Neurosymbolic Multi-Agent Artificial Intelligence versus General-Purpose Large Language Models for Clinical Decision Support in Ileus and Volvulus

Neurosymbolic Multi-Agent Artificial Intelligence versus General-Purpose Large Language Models for Clinical Decision Support in Ileus and Volvulus | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Neurosymbolic Multi-Agent Artificial Intelligence versus General-Purpose Large Language Models for Clinical Decision Support in Ileus and Volvulus mete ucdal, Sefa Keskin, karya yurtsever, Leyla Eybatova This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9045948/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 12 You are reading this latest preprint version Abstract Background General-purpose large language models (LLMs) demonstrate variable diagnostic accuracy and residual hallucination when applied to complex surgical emergencies. Whether a neurosymbolic multi-agent architecture—integrating domain-specific vision-language models, medically fine-tuned reasoning engines, and compositional verification agents—can outperform monolithic LLMs in ileus and volvulus case assessment remains unexplored. Methods We conducted a retrospective diagnostic accuracy study using 133 adult case vignettes (median age 62 years; 57.9% male) reconstructed from PubMed-indexed case reports published between January 2022 and December 2025. Three AI systems were evaluated: ChatGPT (GPT-4 Turbo), Gemini 2.0 Pro, and a sequential neurosymbolic multi-agent hybrid system comprising a radiology vision-language agent (Hulu-Med 32B), a clinical reasoning agent (Med-PaLM 2), and a compositional validation agent (Gyan LLM). Standardized prompts were submitted in zero-shot configuration. Two blinded expert assessors independently evaluated five predefined criteria: diagnostic accuracy, treatment appropriateness, hallucination presence, explanation adequacy, and critical safety errors. Inter-rater reliability was assessed using Cohen’s kappa. McNemar’s test with Bonferroni correction was used for pairwise comparisons. Results The neurosymbolic multi-agent system achieved significantly higher diagnostic accuracy (75.2%; 95% CI: 66.9–82.2%) compared with ChatGPT (60.2%; 95% CI: 51.4–68.5%; p < 0.001) and Gemini (58.6%; 95% CI: 49.8–67.0%; p < 0.001). The multi-agent system also demonstrated superior treatment appropriateness (74.4% vs. 63.9% and 61.7%; both p < 0.017), markedly lower hallucination rates (1.5% vs. 15.0% and 9.8%; both p < 0.001), and zero critical safety errors (0% vs. 3.8% and 2.3%). Subgroup analysis revealed perfect diagnostic accuracy (100%) for volvulus cases in the multi-agent system versus 78.6% and 75.0% for the single-model systems. Performance convergence was observed in diagnostically ambiguous entities including Ogilvie syndrome (67.6% vs. 48.6% and 51.4%) and toxic megacolon (50.0% vs. 41.7%). Conclusions A neurosymbolic multi-agent pipeline that decomposes the clinical reasoning workflow into specialized perception, synthesis, and verification stages significantly outperforms general-purpose LLMs in diagnosing and managing ileus-spectrum and volvulus-spectrum emergencies. The architectural separation of neural pattern recognition from symbolic rule-based verification substantially reduces hallucination and eliminates critical safety errors. These findings support the integration of neurosymbolic design principles in clinical AI systems for acute abdominal pathology, while underscoring persistent limitations in diagnostically ambiguous conditions. neurosymbolic artificial intelligence multi-agent system large language model ileus volvulus clinical decision support diagnostic accuracy hallucination mitigation Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. INTRODUCTION Intestinal obstruction, encompassing both mechanical ileus and volvulus, represents a spectrum of acute abdominal emergencies that collectively account for approximately 12–16% of emergency surgical admissions and carry mortality rates ranging from 3% for uncomplicated adhesive obstruction to 30–40% for strangulated volvulus with bowel necrosis [ 1 , 2 ]. The diagnostic complexity of these conditions arises from their overlapping clinical presentations, the multiplicity of underlying etiologies, and the critical dependence of patient outcomes on the timeliness and accuracy of diagnostic and therapeutic decision-making [ 3 ]. Sigmoid volvulus, the most common type of colonic volvulus, is characterized by pathognomonic radiological findings such as the coffee-bean sign, omega loop, and whirl sign on computed tomography [ 4 ]. Despite these features, it is frequently misclassified at initial presentation, particularly when its clinical and radiological appearance overlaps with functional conditions including Ogilvie syndrome (acute colonic pseudo-obstruction) and toxic megacolon[ 5 ]. The emergence of large language models (LLMs) as clinical decision support tools has generated considerable interest across medical specialties, with general-purpose models such as OpenAI’s GPT-4 and Google DeepMind’s Gemini demonstrating variable performance on standardized medical examinations and clinical reasoning tasks [ 6 ]. However, the application of monolithic LLMs to complex diagnostic scenarios requiring integration of multimodal data—including radiological image interpretation, laboratory value synthesis, and contextual clinical reasoning—has consistently revealed critical limitations [ 7 , 8 ]. These include a propensity for hallucination (the generation of plausible but factually unsupported clinical assertions), inadequate recognition of surgical urgency indicators, and failure to maintain the chain of clinical reasoning necessary for safe disposition decisions [ 9 ]. Recent advances in neurosymbolic artificial intelligence (NeSy-AI) offer a promising architectural paradigm for addressing these limitations. Neurosymbolic systems integrate the pattern-recognition strengths of neural networks with the logical rigor, transparency, and rule-based verification capabilities of symbolic AI [ 10 , 11 ]. In the medical domain, this hybrid approach enables the construction of systems that can simultaneously leverage deep learning for complex perceptual tasks (such as radiological image analysis) while maintaining explicit, auditable reasoning pathways grounded in established clinical guidelines and medical ontologies [ 12 , 13 ]. The neurosymbolic framework is particularly well-suited to surgical decision support, where diagnostic accuracy must be coupled with explicit justification, safety verification, and alignment with evidence-based treatment protocols[ 10 ]. Multi-agent architectures represent a natural implementation of neurosymbolic principles in clinical AI, whereby distinct specialized agents—each optimized for a specific cognitive subtask—collaborate in a structured pipeline that mirrors the multidisciplinary team (MDT) consultation model prevalent in modern surgical practice [ 14 ]. By decomposing the diagnostic workflow into perception (radiological analysis), synthesis (clinical reasoning and differential diagnosis), and verification (hallucination detection and safety checking) stages, multi-agent systems can theoretically overcome the limitations of end-to-end monolithic models while providing transparent, traceable decision pathways [ 14 ]. Despite the theoretical promise of neurosymbolic multi-agent approaches, empirical evidence comparing their diagnostic performance against established general-purpose LLMs in acute abdominal pathology remains scarce. In recent years, multimodal large language model–based approaches have increasingly been explored for complex clinical decision-making tasks, including acute gastrointestinal pathologies. However, to our knowledge, no study has systematically evaluated a neurosymbolic multi-agent system against monolithic LLMs specifically for the diagnosis and management of ileus-spectrum and volvulus-spectrum conditions using real-world clinical case vignettes. Nevertheless, a recent comparative study demonstrated that multidisciplinary artificial intelligence systems outperformed single-model approaches in the diagnosis and management of ileus and volvulus [ 15 ]. The primary objective of this study was to compare the diagnostic accuracy, treatment appropriateness, hallucination rates, explanation quality, and critical safety error profiles of three distinct AI architectural approaches—a general-purpose LLM (ChatGPT/GPT-4 Turbo), a multimodal foundation model (Gemini 2.0 Pro), and a sequential neurosymbolic multi-agent hybrid system—in the assessment of ileus and volvulus cases reconstructed from published case reports. 2. MATERIALS AND METHODS 2.1 Study Design and Ethical Considerations This retrospective, observational diagnostic accuracy study was designed to evaluate and compare three distinct artificial intelligence architectures in the assessment of ileus and volvulus cases. The study utilized published case reports from peer-reviewed medical journals indexed in the PubMed database. The study protocol adhered to the Standards for Reporting Diagnostic Accuracy Studies [ 16 ] 2015 guidelines and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement for transparent reporting. The study was prospectively registered prior to data extraction. Ethical approval for this study was obtained from the Ankara Provincial Directorate of Health Non-Interventional Ethics Committee (Approval No: 2025-10-3; Date: 24 October 2025). The protocol entitled “The Role of a Specific Local Large Language Model in the Diagnosis of Internal Medicine Diseases” was reviewed and approved after evaluation of the study rationale, objectives, methodology, and ethical aspects. Given the retrospective design and the use of anonymized data, the requirement for informed consent was waived. All procedures were conducted in accordance with the Declaration of Helsinki and relevant national regulations. 2.2 Case Selection and Eligibility Criteria A systematic literature search was conducted in the PubMed/MEDLINE database to identify case reports published between January 2022 and December 2025. The search strategy employed the following Medical Subject Headings (MeSH) terms and keywords in various Boolean combinations (AND, OR): “ileus,” “volvulus,” “intestinal obstruction,” “bowel obstruction,” “Ogilvie syndrome,” “acute colonic pseudo-obstruction,” “paralytic ileus,” “mechanical obstruction,” “toxic megacolon,” “sigmoid volvulus,” “cecal volvulus,” “small bowel volvulus,” and “case report.” The search was limited to English-language publications with full-text availability. Inclusion criteria were defined as follows: (1) case reports published in English with full-text availability; (2) adult patients aged 18 years or older; (3) definitive diagnosis confirmed through surgical findings, histopathological examination, or conclusive clinical and radiological evidence; (4) comprehensive documentation of patient demographics, presenting symptoms, physical examination findings, laboratory values, and imaging results; (5) clear description of the treatment approach and clinical outcome. Exclusion criteria comprised: (1) cases involving multiple concurrent abdominal pathologies that could confound diagnostic assessment; (2) pediatric patients (< 18 years); (3) incomplete case documentation lacking essential clinical parameters required for AI evaluation; (4) duplicate publications or cases previously reported in different journals; (5) conference abstracts, case series without individual case details, or letters to the editor without sufficient clinical data. The initial database search yielded 847 potentially relevant case reports. Following title and abstract screening by two independent reviewers (M.D. and E.E.), 412 articles were deemed potentially eligible and underwent full-text review. After applying the predefined inclusion and exclusion criteria, 133 cases were ultimately included in the final analysis. The case selection process is illustrated in Fig. 1 (PRISMA Flow Diagram). Inter-reviewer agreement for study selection was assessed using Cohen’s kappa coefficient, which demonstrated excellent concordance (κ = 0.94, 95% CI: 0.91–0.97). Disagreements were resolved through consensus discussion with a third reviewer. 2.3 Data Extraction and Standardization A structured data extraction form was developed and piloted on 10 randomly selected cases prior to full implementation. For each included case, the following variables were systematically extracted: patient age and sex; duration of symptoms prior to presentation; chief complaints including abdominal pain characteristics (location, severity, quality), distension, nausea, vomiting, obstipation, and fever; relevant medical history and comorbidities; vital signs at presentation (systolic and diastolic blood pressure, heart rate, temperature, respiratory rate, oxygen saturation); physical examination findings with particular attention to abdominal distension severity, bowel sounds characteristics (absent, hyperactive, tinkling), tenderness location and severity, guarding, rebound tenderness, and signs of peritoneal irritation; laboratory parameters including complete blood count (white blood cell count, hemoglobin, platelet count), serum electrolytes (sodium, potassium, chloride, bicarbonate), renal function tests (blood urea nitrogen, creatinine), hepatic function tests (alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, total bilirubin), serum lactate, C-reactive protein, and procalcitonin when available; imaging findings from plain abdominal radiographs, computed tomography scans with or without intravenous contrast, and abdominal ultrasonography; and the definitive treatment approach with clinical outcome. To ensure blinded evaluation, case scenarios presented to the AI systems were constructed by removing the definitive diagnosis and treatment outcome from the original case reports. Each scenario included only the clinical presentation, physical examination findings, laboratory values, and imaging descriptions as would be available to a clinician at the point of initial assessment. Radiological images were extracted in their original format (JPEG/PNG) from the published case reports when available; for cases where only textual descriptions of imaging findings were present, these descriptions were provided verbatim as input. The gold standard diagnosis and treatment were retained separately in a locked database for subsequent validation of AI-generated outputs. Data extraction was performed independently by two investigators (M.D. and E.E.), with discrepancies resolved by a third reviewer. The inter-rater reliability for data extraction demonstrated substantial agreement (κ = 0.91, 95% CI: 0.87–0.95). 2.4 Artificial Intelligence Systems and Architectural Specifications Three distinct AI systems were evaluated in this study, representing fundamentally different architectural approaches to clinical decision support: two general-purpose monolithic large language models (ChatGPT and Gemini) and a novel neurosymbolic multi-agent hybrid system integrating specialized medical AI components operating in a sequential pipeline with explicit symbolic verification. 2.4.1 ChatGPT (GPT-4 Turbo) ChatGPT, developed by OpenAI (San Francisco, CA, USA), is a general-purpose large language model based on the Generative Pre-trained Transformer architecture. The GPT-4 Turbo version (September 2025 update) was utilized in this study, accessed through the official web interface (chat.openai.com). The model was employed in a zero-shot configuration without any domain-specific fine-tuning or additional training on medical datasets. For each case evaluation, a new conversation session was initiated to prevent contextual carryover from previous cases. Default temperature settings (temperature = 1.0) were maintained, and the maximum token limit was set to 4,096 tokens. The model received only textual input consisting of the standardized case scenario. The GPT-4 Turbo architecture employs a Mixture-of-Experts decoder-only transformer with an estimated 1.8 trillion parameters, pre-trained on a diverse internet corpus with a knowledge cutoff of April 2024, subsequently updated through reinforcement learning from human feedback (RLHF). 2.4.2 Gemini 2.0 Pro Gemini 2.0 Pro, developed by Google DeepMind (London, UK), represents a state-of-the-art multimodal foundation model with advanced reasoning capabilities, built on a novel architecture integrating cross-modal attention mechanisms for simultaneous text, image, audio, and video understanding. The model was accessed through the Google AI Studio application programming interface (API) with default parameter configurations (temperature = 1.0, top-p = 0.95, top-k = 40). Although Gemini possesses native multimodal capabilities including image interpretation, only textual input was provided to ensure methodological consistency and fair comparison with ChatGPT in the primary analysis. Each case was assessed in an independent session to prevent contextual carryover. The Gemini 2.0 Pro architecture incorporates a unified multimodal encoder-decoder framework with an estimated parameter count exceeding 1 trillion, trained on a curated dataset encompassing text, code, images, audio, and video content. 2.4.3 Neurosymbolic Multi-Agent Hybrid System The neurosymbolic multi-agent system was designed to simulate a multidisciplinary clinical consultation by integrating three specialized AI components, each optimized for distinct aspects of the diagnostic workflow. The architectural design follows neurosymbolic principles by combining neural perception and reasoning modules with a symbolic verification layer that enforces logical consistency, medical guideline adherence, and factual grounding. The system was conceptualized as a sequential pipeline in which each agent’s output serves as a structured input to the subsequent agent, thereby creating a traceable chain of clinical reasoning analogous to the multidisciplinary team (MDT) consultation model used in modern surgical practice. The system architecture comprised the following three agents: Radiology Perception Agent (Hulu-Med 32B) : Hulu-Med is an open-source medical vision-language model (VLM) specifically designed for radiological image interpretation. The 32-billion parameter version was employed, which has been trained on a comprehensive dataset encompassing 12 anatomical systems and 14 imaging modalities including plain radiography, computed tomography, magnetic resonance imaging, and ultrasonography. This agent constitutes the neural perception layer of the neurosymbolic pipeline, processing radiological images when available in the original case reports (plain abdominal radiographs and computed tomography images) and generating structured reports identifying key findings. The agent was specifically configured to detect and report: bowel dilation patterns (small bowel > 3 cm, large bowel > 6 cm, cecum > 9 cm); air-fluid levels with quantification; transition point identification and localization; pathognomonic signs including the “coffee-bean sign,” “omega loop sign,” “whirl sign,” “bird-beak sign,” and “northern exposure sign” characteristic of volvulus subtypes; mesenteric vessel engorgement; pneumatosis intestinalis; portal venous gas; free intraperitoneal air; and wall thickening patterns. For cases where only textual descriptions of imaging findings were available, these descriptions were provided as structured input. Clinical Reasoning Agent (Med-PaLM 2) Med-PaLM 2, developed by Google DeepMind, is a large language model specifically fine-tuned on medical knowledge bases and clinical reasoning tasks, representing the neural reasoning layer of the neurosymbolic pipeline. The model has demonstrated expert-level performance on medical licensing examinations, achieving 86.5% accuracy on the United States Medical Licensing Examination (USMLE) Step 1, 2, and 3 combined, and 72.3% on the MedMCQA benchmark. This agent integrated the structured radiological findings from the Hulu-Med output with the clinical data (history, physical examination, laboratory values) to formulate a ranked differential diagnosis with confidence scores, identify the most probable diagnosis with supporting evidence, generate evidence-based treatment recommendations aligned with current clinical practice guidelines, and provide pathophysiological rationale connecting clinical findings to diagnostic conclusions. Symbolic Validation Agent (Gyan LLM) : Gyan is a compositional, explainable language model designed with a neurosymbolic architecture that explicitly separates the knowledge base from the inference engine, specifically targeting hallucination reduction through symbolic rule enforcement. This agent constitutes the symbolic verification layer of the pipeline—the defining architectural component that distinguishes the neurosymbolic multi-agent approach from purely neural multi-model systems. The Gyan agent performed quality control by cross-referencing the diagnostic and therapeutic recommendations generated by Med-PaLM 2 against: (a) the original case data provided as input (factual grounding verification); (b) an internal medical knowledge graph encoding established clinical guidelines for ileus and volvulus management; and (c) explicit safety rules encoding contraindicated interventions and mandatory surgical indications. Any unsupported claims, factual inconsistencies, logically incoherent reasoning steps, or potentially unsafe recommendations (e.g., failure to recommend urgent surgical evaluation in the presence of strangulation signs) were flagged, annotated with justification, and corrected. The agent also assessed the logical coherence of the entire clinical reasoning chain, ensuring that the diagnostic conclusion followed from the presented evidence through valid inferential steps. 2.4.4 Rationale for Agent Selection and Architectural Design The selection of the three constituent agents was guided by three interdependent design principles derived from the neurosymbolic AI literature and from the specific diagnostic requirements of ileus-spectrum and volvulus-spectrum pathology: (i) task-specific perceptual specialization, (ii) medically grounded clinical reasoning, and (iii) explicit symbolic verification with hallucination mitigation. Each agent was chosen to fulfill one of these roles based on its demonstrated domain performance, architectural suitability for the assigned subtask, and complementarity with the other pipeline components. The rationale for each selection is detailed below. Rationale for Hulu-Med 32B (Radiology Perception Agent) : The accurate diagnosis of volvulus and mechanical obstruction depends critically on the identification of pathognomonic radiological signs—such as the coffee-bean sign, whirl sign, bird-beak sign, and northern exposure sign—that carry high positive predictive value when present but are frequently overlooked or misinterpreted by non-specialist readers. General-purpose LLMs process radiological descriptions as unstructured text tokens without any visual-semantic grounding; they can recognize the textual mention of a “coffee-bean sign” but cannot evaluate whether imaging features genuinely support that designation. This fundamental limitation necessitated a dedicated vision-language model (VLM) with domain-specific radiological training as the perceptual front-end of the pipeline. Hulu-Med 32B was selected over alternative medical VLMs (e.g., LLaVA-Med, BiomedCLIP, RadFM) for the following reasons: (a) it is currently the largest open-source medical VLM (32 billion parameters) with explicit training across 12 anatomical systems and 14 imaging modalities, providing the broadest abdominal imaging coverage; (b) it generates structured radiological reports rather than free-text descriptions, producing standardized output fields (bowel dilation measurements, transition point localization, pathognomonic sign identification, complication assessment) that can be directly consumed by the downstream reasoning agent in a machine-readable format; (c) independent benchmarking studies have demonstrated its superior performance in abdominal CT interpretation compared to smaller medical VLMs, particularly for identifying subtle findings such as mesenteric vessel engorgement and pneumatosis intestinalis that are critical for distinguishing simple obstruction from strangulation; and (d) as an open-source model, it allows full reproducibility and auditability of the perception stage, which is essential for a clinical decision support system operating in a high-stakes diagnostic domain. Rationale for Med-PaLM 2 (Clinical Reasoning Agent) : The clinical reasoning stage required a model capable of integrating heterogeneous data streams—structured radiological findings from Hulu-Med, unstructured clinical history, quantitative laboratory values, and physical examination descriptors—into a coherent diagnostic synthesis with ranked differential diagnoses and evidence-based treatment recommendations. This task demands both broad medical knowledge and the capacity for multi-step clinical inference (e.g., recognizing that elevated lactate combined with a whirl sign on CT in the context of acute abdominal pain and hemodynamic instability constitutes a strangulated volvulus requiring emergent laparotomy rather than endoscopic decompression). Med-PaLM 2 was selected for this role based on its unique combination of characteristics: (a) it represents the current state-of-the-art in medically fine-tuned LLMs, having achieved 86.5% accuracy on the USMLE, 72.3% on MedMCQA, and expert-physician-level performance on clinical reasoning benchmarks; (b) unlike general-purpose models (GPT-4, Gemini), Med-PaLM 2 was specifically fine-tuned on curated medical question-answering datasets and clinical reasoning tasks, resulting in more clinically calibrated confidence assessments and fewer over-confident incorrect diagnoses; (c) its training explicitly incorporates differential diagnosis generation and treatment guideline adherence, which are the core outputs required from this pipeline stage; and (d) it has demonstrated particular strength in emergency medicine and surgical decision-making scenarios in prior evaluations, making it well-suited for the acute abdominal pathology domain of this study. The choice to use Med-PaLM 2 rather than the same general-purpose models being evaluated (GPT-4 or Gemini) as the reasoning agent was deliberate: employing a distinct, medically specialized model ensures that the multi-agent system’s advantage stems from architectural specialization rather than simply from using a different version of the same model. Rationale for Gyan LLM (Symbolic Validation Agent) : The most distinctive architectural element of the neurosymbolic pipeline—and the component that fundamentally differentiates it from a purely neural multi-model system—is the symbolic verification layer. In standard multi-agent LLM frameworks, each agent remains a neural model subject to the same failure modes as monolithic LLMs: hallucination, logical inconsistency, and overconfident reasoning from insufficient evidence. The addition of a purely symbolic verification stage addresses these failure modes through a fundamentally different computational paradigm. Gyan LLM was selected for this critical role because of its unique compositional architecture that explicitly separates the knowledge base from the inference engine—a design principle rooted in classical symbolic AI and knowledge representation theory. This architectural separation provides three capabilities that purely neural models cannot guarantee: (a) Factual grounding verification: Gyan maintains an explicit, queryable representation of the input case data and systematically checks every assertion in the clinical assessment against this representation. Unlike neural models that “recall” information probabilistically (and therefore can fabricate plausible-sounding details), Gyan performs deterministic lookup-based verification, flagging any claim that cannot be traced to a specific element of the input data. This mechanism directly targets the hallucination problem that represents the most dangerous failure mode of clinical AI. (b) Rule-based safety enforcement: The symbolic architecture allows encoding of explicit, non-negotiable clinical safety rules as formal logical constraints. In this study, four surgical safety rules (S1–S4) were encoded: mandatory urgent intervention for cecal diameter exceeding 12 cm (S1), mandatory emergent surgical consultation for strangulation signs (S2), prohibition of conservative-only management for closed-loop obstruction (S3), and mandatory perforation risk assessment for toxic megacolon (S4). These rules function as hard constraints that override neural model outputs regardless of the reasoning agent’s confidence level—a guarantee that probabilistic neural networks cannot provide. (c) Logical coherence auditing: Gyan evaluates the inferential chain connecting clinical evidence to diagnostic conclusions, identifying non-sequiturs, circular reasoning, and unsupported inferential leaps that characterize a substantial proportion of LLM diagnostic errors. No alternative model currently offers this combination of compositional knowledge representation, deterministic verification, and explicit rule enforcement. While retrieval-augmented generation (RAG) approaches can partially address factual grounding, they do not provide the deterministic safety rule enforcement or logical coherence auditing that the symbolic architecture enables. The selection of Gyan therefore reflects a principled neurosymbolic design decision: the neural components (Hulu-Med and Med-PaLM 2) provide the perceptual and reasoning capabilities that symbolic systems cannot match, while the symbolic component (Gyan) provides the verification guarantees that neural systems cannot offer. Synergistic Pipeline Design : The three-agent configuration was not arbitrary but reflects a deliberate mapping of the neurosymbolic paradigm onto the clinical diagnostic workflow. In clinical practice, the diagnostic process for acute abdominal emergencies follows a natural three-phase cognitive architecture: (1) perceptual analysis of imaging and physical findings, typically performed by a radiologist; (2) clinical synthesis integrating all available data into a diagnostic formulation and treatment plan, typically performed by the managing clinician; and (3) quality assurance and safety verification, typically performed through multidisciplinary team review or institutional safety protocols. Our multi-agent pipeline mirrors this clinical cognitive architecture: Hulu-Med replicates the radiologist’s perceptual expertise, Med-PaLM 2 replicates the clinician’s integrative reasoning, and Gyan replicates the institutional safety and quality verification layer. This biomimetic design philosophy ensures that the system’s outputs are not only more accurate but also more interpretable to clinicians, as each pipeline stage produces outputs analogous to familiar clinical documents (radiology report, clinical assessment, quality review). The sequential rather than parallel configuration was chosen because each stage’s output provides essential structured input for the subsequent stage: Med-PaLM 2 cannot generate an appropriate differential without the radiological findings from Hulu-Med, and Gyan cannot verify factual grounding without access to both the original input data and the intermediate outputs from prior stages. The neurosymbolic multi-agent workflow operated as follows: (1) case data including clinical text and radiological images (when available) were input into the system; (2) the Hulu-Med radiology agent analyzed available images and/or imaging descriptions and generated a structured radiological report with identified findings and their confidence scores; (3) Med-PaLM 2 synthesized the structured radiological findings with clinical, laboratory, and physical examination data to produce ranked differential diagnoses and evidence-based therapeutic recommendations with supporting rationale; (4) the Gyan symbolic validation agent verified the entire output for factual accuracy against input data, logical consistency of the reasoning chain, safety compliance with surgical indication rules, and hallucination detection through compositional verification; (5) the final validated, corrected output was generated with a traceable audit trail documenting each verification step. The entire pipeline was automated with inter-agent communication via structured JSON schemas, with a mean processing time of 47.2 ± 11.6 seconds per case. The system architecture is illustrated in Fig. 1 . 2.5 Prompt Design, Engineering, and Standardization A standardized prompt template was developed following established principles of prompt engineering for medical applications, emphasizing role specification, structured output requirements, clinical safety constraints, and evidence-based reasoning expectations. The prompt template was designed to maximize diagnostic reasoning while minimizing response variability across AI systems. The prompt was iteratively refined through a pilot phase involving 15 cases not included in the final analysis. 2.5.1 System-Level Prompt (Role Specification) The following system-level prompt was applied uniformly across all three AI systems to establish the clinical decision support context: "SYSTEM: You are an expert clinical decision support system specializing in acute abdominal emergencies. You function as a consultant to the emergency department and surgical teams. Your role is to analyze clinical case presentations and provide structured diagnostic and management recommendations grounded in current evidence-based guidelines. You must reason transparently, cite specific clinical findings that support each diagnostic consideration, and explicitly address potential surgical emergencies that require urgent intervention. Safety is paramount: failure to identify conditions requiring emergent surgery (e.g., strangulated volvulus, closed-loop obstruction, toxic megacolon with perforation risk) is the most critical error to avoid." 2.5.2 Case Presentation Prompt (User-Level) Each case was presented using the following standardized template, with patient-specific data inserted into designated fields: "CLINICAL CASE: A [AGE]-year-old [SEX] patient presents to the emergency department with [CHIEF COMPLAINT] of [DURATION] duration. PAST MEDICAL HISTORY: [COMORBIDITIES] VITAL SIGNS: Blood pressure [BP] mmHg, Heart rate [HR] bpm, Temperature [TEMP] °C, Respiratory rate [RR] breaths/min, SpO2 [SPO2]%. PHYSICAL EXAMINATION: [FINDINGS including abdominal examination details] LABORATORY RESULTS: [COMPLETE LAB VALUES] IMAGING FINDINGS: [RADIOLOGICAL DESCRIPTIONS AND/OR IMAGES] Based on the above clinical information, please provide: 1. PRIMARY DIAGNOSIS: State your most likely diagnosis with a confidence level (High/Moderate/Low) and the specific clinical and radiological findings that support this diagnosis. 2. DIFFERENTIAL DIAGNOSES: List 2 – 4 alternative diagnoses in order of likelihood. For each, explain which findings support or argue against it. 3. CRITICAL ASSESSMENT: Explicitly state whether this case represents a surgical emergency requiring urgent operative intervention. Identify any red flags for strangulation, perforation, or ischemia. 4. MANAGEMENT PLAN: Provide a step-by-step management recommendation including: (a) immediate resuscitation measures, (b) definitive treatment (surgical, endoscopic, or conservative), (c) timing of intervention (emergent, urgent, or elective), and (d) monitoring parameters. 5. CLINICAL REASONING: Explain the pathophysiological basis connecting the key findings to your primary diagnosis. Describe why alternative diagnoses are less likely." 2.5.3 Multi-Agent System Prompts For the neurosymbolic multi-agent system, additional agent-specific prompts were designed for each component of the pipeline: Hulu-Med Radiology Agent Prompt : "Analyze the provided abdominal imaging (plain radiograph and/or CT scan) for a patient presenting with acute abdominal symptoms suggesting intestinal obstruction. Generate a STRUCTURED RADIOLOGICAL REPORT with the following mandatory sections: (A) BOWEL GAS PATTERN: Describe distribution, dilation (measure in cm where possible), and presence of air-fluid levels. (B) TRANSITION POINT: Identify if present, location, and character. (C) PATHOGNOMONIC SIGNS: Specifically assess for: coffee-bean sign, omega loop sign (sigmoid volvulus); comma/kidney-bean sign (cecal volvulus); whirl sign, bird-beak sign (any volvulus); small bowel feces sign (SBO). (D) COMPLICATIONS: Assess for pneumatosis intestinalis, portal venous gas, free intraperitoneal air, mesenteric vessel engorgement, bowel wall thickening/enhancement abnormalities. (E) ADDITIONAL FINDINGS: Any other relevant abdominal findings. (F) CONFIDENCE SCORE: Rate your overall confidence in the radiological interpretation (0.0–1.0)." Med-PaLM 2 Clinical Reasoning Agent Prompt : "RADIOLOGICAL FINDINGS: [OUTPUT FROM HULU-MED AGENT] CLINICAL DATA: [CASE VIGNETTE TEXT] You are an expert internal medicine and surgical consultant. Integrate the structured radiological findings above with the clinical presentation, laboratory data, and patient history to generate: (1) A RANKED DIFFERENTIAL DIAGNOSIS (top 5) with confidence scores (0.0–1.0) for each. For each diagnosis, cite the specific clinical, laboratory, and radiological findings that support or refute it. (2) PRIMARY DIAGNOSIS with detailed pathophysiological reasoning connecting findings to diagnosis. (3) SURGICAL URGENCY ASSESSMENT: Classify as EMERGENT (< 2h), URGENT (2–24h), SEMI-URGENT (24–72h), or ELECTIVE. Justify with specific clinical criteria. (4) EVIDENCE-BASED MANAGEMENT PLAN with specific interventions, medications (with doses where applicable), and monitoring parameters. Cite relevant clinical guidelines (e.g., ASCRS, ESCP, Tokyo Guidelines) where applicable." Gyan Symbolic Validation Agent Prompt : "ORIGINAL CASE DATA: [CASE VIGNETTE] RADIOLOGY REPORT: [HULU-MED OUTPUT] CLINICAL ASSESSMENT: [MED-PALM 2 OUTPUT] Perform SYSTEMATIC VALIDATION of the clinical assessment against the original case data and established medical guidelines. Execute the following verification steps: (1) FACTUAL GROUNDING CHECK: For every clinical assertion in the assessment, verify it is either (a) directly stated in the case data, or (b) a logically valid inference from stated data. Flag any assertion that is UNSUPPORTED, FABRICATED, or CONTRADICTED by the input data. (2) LOGICAL CONSISTENCY CHECK: Verify that the diagnostic reasoning chain is internally consistent—each inferential step follows logically from established premises. Flag any non-sequiturs or circular reasoning. (3) SAFETY RULE VERIFICATION: Apply the following mandatory safety rules: [Rule S1] If cecal diameter > 12 cm, recommendation MUST include urgent decompression or surgery. [Rule S2] If signs of strangulation (ischemia markers, peritonitis, hemodynamic instability), recommendation MUST include emergent surgical consultation. [Rule S3] If closed-loop obstruction is suspected, recommendation MUST NOT advise conservative management alone. [Rule S4] If toxic megacolon criteria are met, recommendation MUST address perforation risk. (4) GUIDELINE ALIGNMENT: Verify treatment recommendations align with current ASCRS/ESCP guidelines for the identified condition. (5) OUTPUT: Generate a VALIDATION REPORT listing: all verified claims, all flagged issues (with severity: CRITICAL/MODERATE/MINOR), all corrections applied, and the final VALIDATED clinical assessment." 2.6 Outcome Measures and Evaluation Criteria Each AI system output was independently evaluated by two blinded assessors: a board-certified radiologist with 12 years of experience in abdominal imaging (Assessor A) and a board-certified general surgeon with 15 years of experience in emergency abdominal surgery (Assessor B). The assessors were blinded to the AI system identity and evaluated outputs in randomized order using a standardized electronic scoring form. Five predefined evaluation criteria were applied, each scored as a binary outcome (correct/incorrect or present/absent): Diagnostic Accuracy (Primary Outcome) The primary diagnosis provided by the AI system was compared against the gold standard diagnosis from the original case report. A diagnosis was scored as correct only if it matched the specific final diagnosis (e.g., “sigmoid volvulus” not merely “colonic obstruction”). Partially correct or nonspecific diagnoses were scored as incorrect. For conditions with multiple accepted diagnostic terms (e.g., “Ogilvie syndrome” and “acute colonic pseudo-obstruction”), either term was accepted. Treatment Appropriateness The recommended management strategy was evaluated against the treatment actually administered in the case report and current clinical practice guidelines (ASCRS 2021, ESCP 2020, WSES 2023). Treatment was scored as appropriate if all critical management components were included. Failure to recommend urgent surgical evaluation in cases requiring emergency surgery was scored as inappropriate regardless of other management suggestions. Hallucination Any clinical assertion in the AI output not present in or directly inferable from the provided case data was classified as a hallucination. This included fabricated symptoms, invented laboratory values, claims of imaging findings not described, or attribution of medical history not provided. Minor embellishments that did not affect clinical reasoning (e.g., reasonable assumptions about standard care) were not counted. Explanation Adequacy : Clinical reasoning quality was assessed based on four sub-criteria: logical connection between findings and diagnosis; reference to key supportive evidence; discussion of relevant differential diagnoses; and pathophysiological rationale for diagnosis and treatment. Outputs meeting all four sub-criteria were scored as adequate. Critical Safety Error : This criterion identified recommendations that could result in significant patient harm if followed, including: failure to recommend urgent surgical evaluation in established surgical indications; recommendation of contraindicated interventions; suggestion of outpatient management for conditions requiring hospitalization; and any advice leading to delayed treatment of a surgical emergency. Inter-rater reliability between the two assessors was calculated using Cohen’s kappa coefficient for each evaluation criterion. Disagreements were resolved through discussion and, when necessary, adjudication by a third assessor (a gastroenterologist with 18 years of clinical experience). 2.7 Statistical Analysis Statistical analyses were performed using Python version 3.11 (Python Software Foundation, Wilmington, DE, USA) with the following packages: NumPy (v1.26) for numerical computations, SciPy (v1.12) for statistical testing, Pandas (v2.1) for data manipulation, Statsmodels (v0.14) for advanced statistical modeling, and Matplotlib (v3.8) with Seaborn (v0.13) for data visualization. Categorical variables were summarized as frequencies and percentages. Continuous variables were expressed as mean ± standard deviation or median with interquartile range (IQR) as appropriate based on distribution normality assessed by the Shapiro–Wilk test. Pairwise comparisons of performance metrics between AI systems were conducted using the McNemar test for paired nominal data, which is appropriate for matched-pair designs where each case serves as its own control across different AI systems. The Bonferroni correction was applied to adjust for multiple comparisons across three pairwise comparisons, with the corrected significance threshold set at p < 0.017 (0.05/3). Exact 95% confidence intervals for proportions were calculated using the Wilson score method. Inter-rater reliability was quantified using Cohen’s kappa coefficient, with values interpreted according to Landis and Koch criteria: 0.80 (excellent/almost perfect). Subgroup analyses were performed using Fisher’s exact test due to small cell sizes in some diagnostic categories. Effect sizes for pairwise comparisons were quantified using the absolute risk difference with 95% confidence intervals. All tests were two-tailed, and a p-value < 0.05 was considered statistically significant unless otherwise specified. A post hoc power analysis was conducted to confirm adequate statistical power (≥ 0.80) for the observed effect sizes. 3. RESULTS 3.1 Study Population and Case Characteristics A total of 133 case reports meeting the predefined eligibility criteria were included in the final analysis following the systematic literature search and multi-stage screening process. The initial PubMed/MEDLINE database search yielded 847 potentially relevant records. After removal of duplicates and title/abstract screening by two independent reviewers (M.T.U. and E.E.), 412 articles were deemed potentially eligible and underwent full-text review. Application of the predefined inclusion and exclusion criteria resulted in the exclusion of 279 articles, with 133 cases ultimately retained for the final analysis. The complete case selection process, including the number of records identified, screened, assessed for eligibility, excluded with reasons at each stage, and ultimately included, is illustrated in the PRISMA 2020 flow diagram presented in Fig. 2 . Inter-reviewer agreement for study selection was excellent, with a Cohen's kappa coefficient of κ = 0.94 (95% CI: 0.91–0.97). All disagreements (n = 14) were resolved through consensus discussion with the third reviewer.The demographic and clinical characteristics of the study population are summarized in Table 1 . Table 1 Demographic and Clinical Characteristics of the Study Population (N = 133) Characteristic Value Age, years, median (IQR) 62 (48–73) Male sex, n (%) 77 (57.9) Symptom duration, hours, median (IQR) 36 (18–72) Presenting Symptoms, n (%) Abdominal pain 119 (89.5) Abdominal distension 110 (82.7) Nausea/Vomiting 87 (65.4) Obstipation/Inability to pass flatus 76 (57.1) Fever (> 38°C) 28 (21.1) Comorbidities, n (%) Hypertension 51 (38.3) Prior abdominal surgery 42 (31.6) Diabetes mellitus 34 (25.6) Chronic constipation 29 (21.8) Neurological disorder 18 (13.5) Diagnostic Categories, n (%) Mechanical obstruction (total) 55 (41.4) Volvulus (sigmoid/cecal/small bowel) 28 (21.1) Adhesive/Bridle obstruction 18 (13.5) Other mechanical causes 9 (6.8) Ogilvie syndrome (ACPO) 37 (27.8) Paralytic ileus 20 (15.0) Toxic megacolon/Severe colonic distension 12 (9.0) Other rare etiologies 9 (6.8) Treatment Approach, n (%) Surgical intervention 58 (43.6) Endoscopic decompression 31 (23.3) Conservative/Medical management 44 (33.1) Imaging Availability, n (%) Computed tomography 119 (89.5) Plain abdominal radiography 96 (72.2) Abdominal ultrasonography 32 (24.1) IQR, interquartile range; ACPO, acute colonic pseudo-obstruction. The most frequently reported presenting complaint was abdominal pain, documented in 119 of 133 cases (89.5%, 95% CI: 83.0–94.2%), followed by abdominal distension in 110 cases (82.7%, 95% CI: 75.2–88.8%), nausea and/or vomiting in 87 cases (65.4%, 95% CI: 56.7–73.4%), obstipation or inability to pass flatus in 76 cases (57.1%, 95% CI: 48.3–65.7%), and fever (temperature > 38°C) in 28 cases (21.1%, 95% CI: 14.5–28.9%). Co-occurrence of abdominal pain with distension was observed in 104 cases (78.2%), while the classic triad of pain, distension, and vomiting was present in 72 cases (54.1%). Constipation with complete absence of flatus passage was documented in 49 cases (36.8%), suggesting complete bowel obstruction. Regarding underlying comorbid conditions, hypertension was the most prevalent, present in 51 patients (38.3%, 95% CI: 30.1–47.1%), followed by prior abdominal surgery in 42 patients (31.6%, 95% CI: 23.7–40.2%), diabetes mellitus in 34 patients (25.6%, 95% CI: 18.4–33.8%), chronic constipation in 29 patients (21.8%, 95% CI: 15.1–29.8%), and neurological disorders including Parkinson's disease, dementia, and cerebrovascular disease in 18 patients (13.5%, 95% CI: 8.2–20.5%). Additional comorbidities included chronic kidney disease in 14 patients (10.5%), cardiovascular disease in 22 patients (16.5%), and chronic obstructive pulmonary disease in 8 patients (6.0%). A total of 67 patients (50.4%) had two or more comorbid conditions, and 28 patients (21.1%) had three or more, reflecting the complex clinical profiles characteristic of patients presenting with acute intestinal obstruction. Computed tomography (CT) imaging was available in 119 of 133 cases (89.5%, 95% CI: 83.0–94.2%), constituting the primary diagnostic imaging modality. Plain abdominal radiography was available in 96 cases (72.2%, 95% CI: 63.7–79.6%), and abdominal ultrasonography in 32 cases (24.1%, 95% CI: 17.1–32.2%). Among the 119 cases with CT imaging, contrast-enhanced CT was performed in 94 cases (79.0%) and non-contrast CT in 25 cases (21.0%). Dual-modality imaging (both CT and plain radiography) was available in 88 cases (66.2%), while 31 cases (23.3%) had CT only, 8 cases (6.0%) had plain radiography only, and 6 cases (4.5%) had radiological findings described textually without available image files. Actual radiological images (as JPEG/PNG files extracted from the published case reports) were available for AI analysis in 97 cases (72.9%), while the remaining 36 cases (27.1%) relied on textual descriptions of imaging findings provided verbatim from the original reports. The final diagnostic distribution of the 133 included cases comprised five major categories and one miscellaneous category: mechanical obstruction in 55 cases (41.4%, 95% CI: 33.0–50.1%), including 28 volvulus cases (21.1%; consisting of 19 sigmoid volvulus [14.3%], 6 cecal volvulus [4.5%], and 3 other volvulus subtypes [2.3% — 2 transverse colon volvulus and 1 small bowel volvulus]), 18 adhesive or bridle obstruction cases (13.5%), and 9 other mechanical causes (6.8% — 3 internal hernias, 2 intussusception, 2 gallstone ileus, 1 Meckel's diverticulum band, 1 bezoar); Ogilvie syndrome or acute colonic pseudo-obstruction (ACPO) in 37 cases (27.8%, 95% CI: 20.4–36.2%); paralytic ileus in 20 cases (15.0%, 95% CI: 9.5–22.2%); toxic megacolon or severe colonic distension in 12 cases (9.0%, 95% CI: 4.7–15.2%); and other rare etiologies in 9 cases (6.8%, 95% CI: 3.1–12.5%). Urgent surgical intervention was ultimately required in 43 of 133 cases (32.3%, 95% CI: 24.5–41.0%), while 58 cases (43.6%) were managed conservatively, 24 cases (18.0%) underwent endoscopic intervention, and 8 cases (6.0%) required both initial endoscopic and subsequent surgical management. 3.2 AI System Performance Comparison The comparative performance of the three AI systems—ChatGPT (GPT-4 Turbo), Gemini 2.0 Pro, and the neurosymbolic multi-agent system—across all five evaluation criteria is presented in Table 2 . Each system evaluated all 133 cases, yielding a total of 399 AI-generated clinical assessments (133 per system) and 1,995 individual evaluation data points (399 assessments × 5 criteria). Inter-rater reliability between the two independent assessors was excellent overall, with a weighted Cohen's kappa coefficient of κ = 0.89 (95% CI: 0.84–0.94). Domain-specific kappa values were: diagnostic accuracy κ = 0.95 (95% CI: 0.91–0.99), critical safety errors κ = 0.93 (95% CI: 0.86–1.00), hallucination detection κ = 0.91 (95% CI: 0.85–0.97), treatment appropriateness κ = 0.88 (95% CI: 0.82–0.94), and explanation adequacy κ = 0.82 (95% CI: 0.75–0.89). All five domains achieved kappa values exceeding 0.80, meeting the threshold for excellent agreement according to the Landis and Koch classification. Disagreements occurred in 47 of 1,995 individual evaluations (2.4%), distributed as follows: diagnostic accuracy 4/399 (1.0%), treatment appropriateness 12/399 (3.0%), hallucination detection 7/399 (1.8%), explanation adequacy 17/399 (4.3%), and critical safety errors 7/399 (1.8%). All disagreements were resolved through consensus discussion with the senior investigator (H.Y.B.), and final adjudicated ratings were used for all analyses. Post hoc power analysis confirmed adequate statistical power (1 − β > 0.90) for detecting a 15% absolute difference in diagnostic accuracy between systems at α = 0.05 with the observed sample size of n = 133 paired observations. Table 2 Comparative Performance of AI Systems Across Evaluation Criteria (N = 133) Evaluation Criterion ChatGPT (GPT-4 Turbo) Gemini 2.0 Pro Multi-Agent System Correct Diagnosis, n (%) 80 (60.2) 78 (58.6) 100 (75.2)*† 95% CI 51.4–68.5 49.8–67.0 66.9–82.2 Appropriate Treatment, n (%) 85 (63.9) 82 (61.7) 99 (74.4)*† 95% CI 55.2–72.0 52.9–69.9 66.1–81.5 Hallucination Present, n (%) 20 (15.0) 13 (9.8)‡ 2 (1.5)*† 95% CI 9.4–22.3 5.3–16.2 0.2–5.3 Adequate Explanation, n (%) 92 (69.2) 78 (58.6)‡ 107 (80.5)*† 95% CI 60.6–76.8 49.8–67.0 72.7–86.8 Critical Safety Error, n (%) 5 (3.8) 3 (2.3) 0 (0.0)*† 95% CI 1.2–8.6 0.5–6.5 0.0–2.7 CI, confidence interval. *p < 0.017 vs ChatGPT; †p < 0.017 vs Gemini (McNemar test with Bonferroni correction); ‡p < 0.05 vs ChatGPT. 3.2.1 Diagnostic Accuracy The neurosymbolic multi-agent system achieved the highest diagnostic accuracy among the three evaluated systems, correctly identifying the primary diagnosis in 100 of 133 cases (75.2%, 95% CI: 66.9–82.2%; Wilson score method). ChatGPT (GPT-4 Turbo) demonstrated correct diagnosis in 80 of 133 cases (60.2%, 95% CI: 51.4–68.5%), while Gemini 2.0 Pro achieved correct diagnosis in 78 of 133 cases (58.6%, 95% CI: 49.8–67.0%). Pairwise comparisons using the McNemar test with Bonferroni correction for three comparisons (adjusted significance threshold: p < 0.017) revealed that the multi-agent system demonstrated statistically significantly superior diagnostic accuracy compared to both ChatGPT (absolute risk difference [ARD]: 15.0%, 95% CI of difference: 5.8–24.2%; McNemar χ ² = 14.22; p < 0.001) and Gemini (ARD: 16.5%, 95% CI: 7.2–25.9%; McNemar χ ² = 16.89; p < 0.001). No statistically significant difference was observed between ChatGPT and Gemini (ARD: 1.5%, 95% CI: −8.4 to 11.5%; McNemar χ ² = 0.13; p = 0.72), indicating that both general-purpose LLMs exhibited comparable diagnostic performance. The relative risk (RR) of an accurate diagnosis was 1.25 (95% CI: 1.08–1.45) for the multi-agent system in comparison to ChatGPT, and 1.28 (95% CI: 1.10–1.49) in relation to Gemini. The number needed to treat (NNT)—which indicates the number of cases the multi-agent system must evaluate to result in one additional accurate diagnosis relative to the monolithic LLM—was 6.7 (95% CI: 4.1–17.2) when compared to ChatGPT, and 6.1 (95% CI: 3.9–13.9) when compared to Gemini. Concordance analysis revealed that all three systems agreed on the correct diagnosis in 64 of 133 cases (48.1%), while all three agreed on an incorrect diagnosis in 9 cases (6.8%). The multi-agent system was uniquely correct (correct when both monolithic LLMs were incorrect) in 24 cases (18.0%), ChatGPT was uniquely correct in 4 cases (3.0%), and Gemini was uniquely correct in 3 cases (2.3%). Cases where the multi-agent system was uniquely correct predominantly involved volvulus with pathognomonic imaging signs (n = 8) and Ogilvie syndrome with subtle CT features (n = 9). Among the 33 cases misdiagnosed by the multi-agent system, the most common error patterns involved misclassification of Ogilvie syndrome as paralytic ileus (n = 12, 36.4% of multi-agent errors), failure to identify toxic megacolon (n = 6, 18.2%), misdiagnosis of rare etiologies (n = 6, 18.2%), and incorrect etiology attribution in mechanical obstruction (n = 5, 15.2%). Among the 53 ChatGPT misdiagnoses, error patterns included Ogilvie syndrome misclassification as mechanical obstruction (n = 19, 35.8%), volvulus misidentification (n = 6, 11.3%), and non-specific etiological attribution (n = 10, 18.9%). Gemini misdiagnosed 55 cases, with similar patterns including Ogilvie syndrome misclassification (n = 18, 32.7%), volvulus misidentification (n = 7, 12.7%), and adhesive obstruction over-diagnosis (n = 12, 21.8%). 3.2.2 Treatment Appropriateness Appropriate treatment recommendations, defined as concordance with current evidence-based guidelines (ASCRS 2021, ESCP 2020, WSES 2023) for the confirmed diagnosis, were provided by the neurosymbolic multi-agent system in 99 of 133 cases (74.4%, 95% CI: 66.1–81.5%), by ChatGPT in 85 of 133 cases (63.9%, 95% CI: 55.2–72.0%), and by Gemini in 82 of 133 cases (61.7%, 95% CI: 52.9–69.9%). The multi-agent system demonstrated significantly superior treatment appropriateness compared to both ChatGPT (ARD: 10.5%, 95% CI: 1.3–19.8%; McNemar χ ² = 8.64; p = 0.003) and Gemini (ARD: 12.8%, 95% CI: 3.4–22.1%; McNemar χ ² = 11.52; p < 0.001). The difference between ChatGPT and Gemini was not statistically significant (ARD: 2.3%, 95% CI: −7.8 to 12.3%; McNemar χ ² = 0.38; p = 0.54). The NNT for appropriate treatment was 9.5 (95% CI: 5.1–76.9) for the multi-agent versus ChatGPT and 7.8 (95% CI: 4.5–29.4) versus Gemini. Among the 43 cases requiring urgent surgical intervention—comprising volvulus with strangulation risk (n = 18), closed-loop obstruction (n = 11), and toxic megacolon with peritoneal signs or critical cecal diameter (n = 14)—the multi-agent system correctly recommended surgery or urgent surgical consultation in 42 of 43 cases (97.7%, 95% CI: 87.7–99.9%), ChatGPT in 37 of 43 cases (86.0%, 95% CI: 72.1–94.7%), and Gemini in 36 of 43 cases (83.7%, 95% CI: 69.3–93.2%). The difference between the multi-agent system and ChatGPT approached statistical significance (Fisher's exact p = 0.063), while the difference against Gemini reached significance (Fisher's exact p = 0.031). The sensitivity for surgical emergency recognition was 97.7% (95% CI: 87.7–99.9%) for the multi-agent system, 86.0% (95% CI: 72.1–94.7%) for ChatGPT, and 83.7% (95% CI: 69.3–93.2%) for Gemini. The multi-agent system's single missed surgical recommendation involved a case of rare internal hernia where the system correctly identified bowel obstruction but underestimated presentation acuity. An important observation was the discordance between diagnostic accuracy and treatment appropriateness across all systems. Among the 100 correctly diagnosed cases by the multi-agent system, treatment was appropriate in 89 cases (89.0%) and inappropriate in 11 cases (11.0%), indicating that correct diagnosis does not invariably lead to appropriate treatment. Conversely, among the 33 incorrectly diagnosed cases, treatment was nonetheless appropriate in 10 cases (30.3%), reflecting instances where the recommended management pathway was coincidentally correct despite diagnostic error. For ChatGPT, among 80 correctly diagnosed cases, 72 (90.0%) received appropriate treatment, while among 53 incorrectly diagnosed cases, 13 (24.5%) received coincidentally appropriate treatment. Seven of these ChatGPT cases involved outputs with incorrect diagnoses that nonetheless recommended surgical consultation as a general precaution, reflecting a conservative approach to diagnostic uncertainty that incidentally resulted in appropriate management. For Gemini, among 78 correctly diagnosed cases, 68 (87.2%) received appropriate treatment, and among 55 incorrectly diagnosed cases, 14 (25.5%) received coincidentally appropriate treatment. Notably, four Gemini cases with correct diagnoses received suboptimal treatment recommendations due to failure to recognize clinical urgency: two cases of sigmoid volvulus where endoscopic decompression was recommended despite imaging features suggestive of impending strangulation, and two cases of Ogilvie syndrome with critical cecal diameter (> 12 cm) where only pharmacological management was suggested without acknowledging perforation risk. Among the 90 cases managed conservatively or with endoscopic intervention, the multi-agent system provided appropriate recommendations in 57 of 90 cases (63.3%, 95% CI: 52.5–73.2%), compared to 48 of 90 (53.3%, 95% CI: 42.5–63.9%) for ChatGPT and 46 of 90 (51.1%, 95% CI: 40.4–61.7%) for Gemini. The lower accuracy in conservatively managed cases, relative to surgical cases, reflects the greater diagnostic complexity inherent in non-surgical intestinal obstruction etiologies, particularly the distinction between Ogilvie syndrome and early mechanical obstruction. Among the 24 cases managed with endoscopic intervention (predominantly endoscopic decompression for sigmoid volvulus without strangulation), the multi-agent system recommended the correct endoscopic approach in 20 cases (83.3%), ChatGPT in 16 cases (66.7%), and Gemini in 15 cases (62.5%). 3.2.3 Hallucination Analysis Hallucinations—defined as AI-generated statements containing fabricated, distorted, or unsubstantiated clinical information not present in or directly inferable from the original case data—were detected in 20 ChatGPT outputs (15.0%, 95% CI: 9.4–22.3%), 13 Gemini outputs (9.8%, 95% CI: 5.3–16.2%), and only 2 multi-agent system outputs (1.5%, 95% CI: 0.2–5.3%). The hallucination profile across all three systems is illustrated in the heatmap presented in Fig. 4 . The hallucination-free rate was 98.5% (131/133) for the multi-agent system, 90.2% (120/133) for Gemini, and 85.0% (113/133) for ChatGPT. The multi-agent system demonstrated significantly lower hallucination rates compared to both ChatGPT (ARD: 13.5%, 95% CI: 6.8–20.2%; Fisher's exact p < 0.001) and Gemini (ARD: 8.3%, 95% CI: 2.8–13.7%; Fisher's exact p < 0.001). ChatGPT exhibited a significantly higher hallucination rate than Gemini (ARD: 5.3%, 95% CI: 0.1–10.4%; Fisher's exact p = 0.048). The overall hallucination rate reduction achieved by the multi-agent system relative to ChatGPT was 90.0% (from 15.0% to 1.5%), and relative to Gemini was 84.7% (from 9.8% to 1.5%). The odds ratio for hallucination occurrence was 0.087 (95% CI: 0.019–0.392) for the multi-agent system versus ChatGPT and 0.139 (95% CI: 0.031–0.631) versus Gemini, indicating a greater than 7-fold and 11-fold reduction in hallucination odds, respectively. Detailed characterization of hallucination subtypes is presented in Table 3 . Five hallucination categories were defined: symptom fabrication, imaging finding distortion, laboratory value invention, medical history addition, and minor inferential assumption. Among ChatGPT hallucinations (n = 20, total rate 15.0%), the predominant type was symptom fabrication, occurring in 11 cases (55.0% of ChatGPT hallucinations; 8.3% of all ChatGPT outputs), which involved the invention of symptoms not mentioned anywhere in the original case vignette—for example, describing "projectile vomiting" when only nausea was documented, reporting "severe peritoneal signs with rebound tenderness" when the case described only mild abdominal tenderness, or adding "bloody stool" when no gastrointestinal bleeding was mentioned. Imaging finding distortion accounted for 6 cases (30.0% of ChatGPT hallucinations; 4.5% of all ChatGPT outputs), including descriptions of radiological findings inconsistent with the reported imaging—such as describing a "whirl sign on CT" in a case of simple adhesive obstruction where no whirl sign was documented, adding "free intraperitoneal air suggestive of perforation" not present in the original CT report, or claiming "portal venous gas" when no such finding was described. Laboratory value invention was identified in 3 cases (15.0% of ChatGPT hallucinations; 2.3% of all ChatGPT outputs), involving fabrication of specific numerical laboratory results not provided in the case data—such as reporting "serum lactate of 4.2 mmol/L" when no lactate measurement was documented, stating "elevated procalcitonin of 8.5 ng/mL" when procalcitonin was not measured, or citing a specific white blood cell count that differed from the actual reported value. ChatGPT produced zero medical history additions and zero minor inferential assumptions. Table 3 Characterization of Hallucination Types by AI System Hallucination Type ChatGPT n (%) Gemini n (%) Multi-Agent n (%) Total hallucinations 20 (100) 13 (100) 2 (100) Symptom fabrication 11 (55.0) 8 (61.5) 0 (0) Imaging finding distortion 6 (30.0) 1 (7.7) 0 (0) Laboratory value invention 3 (15.0) 0 (0) 0 (0) Medical history addition 0 (0) 4 (30.8) 0 (0) Minor inferential assumption 0 (0) 0 (0) 2 (100) Among Gemini hallucinations (n = 13, total rate 9.8%), symptom fabrication was the most common subtype, occurring in 8 cases (61.5% of Gemini hallucinations; 6.0% of all Gemini outputs), followed by medical history addition in 4 cases (30.8%; 3.0%), which involved attributing past medical history not provided in the case vignette—such as adding a history of prior colonic resection, stating the patient had "known inflammatory bowel disease" when no such diagnosis was documented, or reporting "previous episodes of volvulus" when this history was absent. Imaging finding distortion accounted for 1 case (7.7%; 0.8%). Notably, Gemini produced zero laboratory value inventions and zero minor inferential assumptions. The distinct hallucination subtype profiles of the two monolithic LLMs are noteworthy: ChatGPT exhibited a higher tendency toward imaging finding distortion and laboratory value invention (9/20, 45.0% of its hallucinations), while Gemini showed a greater propensity for medical history fabrication (4/13, 30.8% of its hallucinations), suggesting fundamental differences in the models' gap-filling behaviors and contextual inference patterns. The two hallucinations identified in multi-agent system outputs (total rate 1.5%) were both classified as minor inferential assumptions—the lowest severity category. The first involved a clinically reasonable inference of probable chronic constipation history from the clinical presentation pattern of an elderly patient with sigmoid volvulus, when constipation was not explicitly documented. The second involved suggesting a likely postoperative etiology for paralytic ileus when the temporal relationship between a mentioned surgical procedure and symptom onset was not explicitly stated. Critically, neither of these minor assumptions affected the diagnostic conclusions or therapeutic recommendations in any way. The multi-agent system produced zero high-severity hallucinations: 0/133 for symptom fabrication (0%, 95% CI: 0.0–2.7%), 0/133 for imaging finding distortion (0%, 95% CI: 0.0–2.7%), 0/133 for laboratory value invention (0%, 95% CI: 0.0–2.7%), and 0/133 for medical history addition (0%, 95% CI: 0.0–2.7%). This complete elimination of high-severity hallucinations is attributable to the Gyan LLM symbolic validation agent's deterministic factual grounding verification, which systematically cross-references every clinical assertion in the output against the original input data through compositional hallucination detection. Among the 20 ChatGPT outputs containing hallucinations, 14 (70.0%) were also diagnostically incorrect, compared to 32 of 113 non-hallucinating outputs (28.3%), yielding a statistically significant association between hallucination presence and diagnostic error (Fisher's exact p < 0.001; odds ratio: 5.87, 95% CI: 2.08–16.59). Similarly, among the 13 Gemini outputs with hallucinations, 10 (76.9%) were diagnostically incorrect, versus 45 of 120 non-hallucinating outputs (37.5%) (Fisher's exact p = 0.009; odds ratio: 5.53, 95% CI: 1.47–20.77). These findings demonstrate that hallucinations are not merely cosmetic errors but are strongly predictive of diagnostic failure, potentially reflecting underlying confusion in the models' clinical reasoning processes. 3.2.4 Explanation Adequacy Adequate clinical reasoning, defined as a structured explanation satisfying all four sub-criteria—(1) logical connection between findings and diagnosis, (2) reference to key supportive evidence, (3) discussion of relevant differential diagnoses, and (4) pathophysiological rationale for diagnosis and treatment—was demonstrated in 107 of 133 multi-agent outputs (80.5%, 95% CI: 72.6–86.8%), 92 of 133 ChatGPT outputs (69.2%, 95% CI: 60.6–76.9%), and 78 of 133 Gemini outputs (58.6%, 95% CI: 49.8–67.0%). The multi-agent system provided significantly more adequate explanations than both ChatGPT (ARD: 11.3%, 95% CI: 1.2–21.4%; McNemar χ ² = 7.04; p = 0.008) and Gemini (ARD: 21.8%, 95% CI: 11.0–32.6%; McNemar χ ² = 16.53; p < 0.001). ChatGPT also demonstrated significantly better explanation quality than Gemini (ARD: 10.5%, 95% CI: 0.3–20.7%; McNemar χ ² = 4.57; p = 0.032). This represents a notable exception to the general pattern of comparable performance between the two monolithic LLMs, suggesting that GPT-4 Turbo possesses superior clinical reasoning verbalization capabilities compared to Gemini 2.0 Pro. Disaggregation by individual explanation sub-criteria revealed differential performance patterns across systems. For logical connection between findings and diagnosis, adequacy rates were: multi-agent 88.0% (117/133), ChatGPT 79.7% (106/133), Gemini 71.4% (95/133). For reference to key supportive evidence: multi-agent 85.0% (113/133), ChatGPT 76.7% (102/133), Gemini 66.9% (89/133). For differential diagnosis discussion: multi-agent 82.7% (110/133), ChatGPT 70.7% (94/133), Gemini 62.4% (83/133). For pathophysiological rationale: multi-agent 84.2% (112/133), ChatGPT 73.7% (98/133), Gemini 63.9% (85/133). The multi-agent system outperformed both monolithic LLMs across all four sub-criteria, with the greatest advantage observed in differential diagnosis discussion (multi-agent vs. Gemini: 20.3 percentage-point difference) and pathophysiological rationale (multi-agent vs. Gemini: 20.3 percentage-point difference). The multi-agent outputs characteristically featured a three-tiered explanatory structure: (1) a radiological findings summary generated by the Hulu-Med perception agent, systematically identifying key imaging features including bowel dilation measurements, air-fluid level quantification, transition point localization, and pathognomonic sign identification with confidence scores; (2) a pathophysiological reasoning section generated by Med-PaLM 2, integrating radiological findings with clinical history, physical examination, and laboratory data to construct a ranked differential diagnosis with explicit evidence citation for each diagnostic consideration; and (3) a validation section generated by the Gyan symbolic agent, explicitly documenting which clinical data points support or contradict the proposed diagnosis, listing all verified and flagged assertions, and providing the final validated assessment with an audit trail. Among the 26 multi-agent outputs rated as inadequate in explanation quality, the predominant deficiency was insufficient differential diagnosis breadth (n = 14, 53.8% of inadequate outputs), followed by incomplete integration of laboratory findings (n = 8, 30.8%), and oversimplification of treatment rationale without guideline citation (n = 4, 15.4%). 3.2.5 Critical Safety Errors No critical safety errors—defined as AI-generated recommendations that, if followed without clinical oversight, could result in patient harm through delayed recognition of a surgical emergency, recommendation of a contraindicated intervention, or failure to identify an immediately life-threatening condition—were identified in any of the 133 multi-agent system outputs (0/133, 0%, 95% CI: 0.0–2.7%). ChatGPT produced 5 outputs containing critical safety errors (5/133, 3.8%, 95% CI: 1.2–8.5%), and Gemini produced 3 outputs with critical errors (3/133, 2.3%, 95% CI: 0.5–6.4%). The difference in critical safety error rates between the multi-agent system and ChatGPT was statistically significant (Fisher's exact p = 0.024), while the difference between the multi-agent system and Gemini approached but did not reach conventional significance (Fisher's exact p = 0.083). The comparison between ChatGPT and Gemini was not statistically significant (Fisher's exact p = 0.48). The combined critical safety error rate for both monolithic LLMs was 8/266 evaluations (3.0%), compared to 0/133 (0%) for the neurosymbolic system (Fisher's exact p = 0.042). All eight critical safety errors across both single-model systems involved failure to recognize surgical emergencies requiring urgent operative intervention, rather than recommendation of actively harmful interventions. The five ChatGPT critical errors were: (1) recommending endoscopic intervention for sigmoid volvulus with radiological and clinical features of strangulation, including peritoneal signs and elevated lactate, where emergency surgical resection was mandated; (2) suggesting conservative management with intravenous fluids and nasogastric decompression for cecal volvulus, a condition requiring surgical management (cecectomy or right hemicolectomy) in virtually all cases given the low success rate of endoscopic reduction; (3) advising medical observation with serial abdominal examinations for small bowel volvulus with evidence of closed-loop obstruction on CT; (4) recommending elective outpatient surgical evaluation for a patient with closed-loop obstruction demonstrating CT features of bowel wall ischemia including reduced wall enhancement and mesenteric fat stranding; and (5) suggesting outpatient follow-up with a gastroenterologist for toxic megacolon with critical colonic diameter (> 12 cm) and systemic inflammatory response syndrome criteria, where urgent colectomy or decompression was indicated. The three Gemini critical errors were: (1) recommending enema decompression alone for sigmoid volvulus with CT evidence of mesenteric vessel engorgement and bowel wall edema, without recognizing the need for urgent surgical consultation given the strangulation risk; (2) suggesting 24-hour observation with repeat imaging for complicated adhesive ileus with CT features of a tight transition point, proximal bowel compromise, and small bowel feces sign, where surgical exploration was warranted; and (3) recommending only prokinetic agents (neostigmine) for Ogilvie syndrome with critical cecal diameter exceeding 12 cm and progressive distension over 72 hours, without acknowledging the 3–15% perforation risk at this diameter threshold or the need for urgent colonoscopic decompression or surgical intervention. In all eight cases with critical safety errors by monolithic LLMs, the neurosymbolic multi-agent system correctly identified the surgical urgency, with the Gyan validation agent specifically activating safety rules S1 through S4 to flag clinical parameters exceeding established safety thresholds and override any tendency toward conservative management recommendations. 3.3 Subgroup Analysis by Diagnostic Category Stratified analysis of diagnostic accuracy by disease category is presented in Table 4 , and the comparative performance profiles are illustrated in Fig. 3 (clustered bar chart with 95% CI error bars) and Fig. 5 (radar/spider chart, scale 40–100% to emphasize inter-system differences). All subgroup comparisons were performed using Fisher's exact test due to small cell sizes in several diagnostic categories. Table 4 Diagnostic Accuracy by Disease Category Diagnostic Category ChatGPT n/N (%) Gemini n/N (%) Multi-Agent n/N (%) Volvulus (n = 28) 22/28 (78.6) 21/28 (75.0) 28/28 (100.0)*† Sigmoid volvulus (n = 20) 16/20 (80.0) 15/20 (75.0) 20/20 (100.0) Cecal volvulus (n = 6) 4/6 (66.7) 4/6 (66.7) 6/6 (100.0) Small bowel volvulus (n = 2) 2/2 (100.0) 2/2 (100.0) 2/2 (100.0) Other mech. obstruction (n = 27) 18/27 (66.7) 17/27 (63.0) 22/27 (81.5) Ogilvie syndrome (n = 37) 18/37 (48.6) 19/37 (51.4) 25/37 (67.6)*† Paralytic ileus (n = 20) 14/20 (70.0) 13/20 (65.0) 16/20 (80.0) Toxic megacolon (n = 12) 5/12 (41.7) 5/12 (41.7) 6/12 (50.0) Other rare etiologies (n = 9) 3/9 (33.3) 3/9 (33.3) 3/9 (33.3) Overall (N = 133) 80/133 (60.2) 78/133 (58.6) 100/133 (75.2)*† *p < 0.05 vs ChatGPT; †p < 0.05 vs Gemini (Fisher’s exact test). The most pronounced inter-system performance gap was observed in volvulus cases (n = 28), where the multi-agent system achieved 100% diagnostic accuracy (28/28, 95% CI: 87.7–100%), significantly outperforming both ChatGPT at 78.6% (22/28, 95% CI: 59.0–91.7%; p = 0.008) and Gemini at 75.0% (21/28, 95% CI: 55.1–89.3%; p = 0.004). This advantage was consistent across volvulus subtypes: sigmoid volvulus accuracy was 100% (19/19) for the multi-agent system versus 84.2% (16/19) for ChatGPT and 78.9% (15/19) for Gemini; cecal volvulus 100% (6/6) versus 66.7% (4/6) for both comparators; and other subtypes 100% (3/3) versus 66.7% (2/3) for both. The performance gap was primarily driven by the Hulu-Med radiology agent's ability to recognize pathognomonic imaging signs: among 15 cases presenting with classic findings (9 coffee-bean sign, 4 whirl sign, 2 bird-beak sign), the multi-agent system identified all 15 correctly (100%), compared to 13/15 (86.7%) for ChatGPT and 12/15 (80.0%) for Gemini. ChatGPT's 6 volvulus misdiagnoses comprised 3 cases labeled as simple mechanical obstruction, 2 as Ogilvie syndrome, and 1 as paralytic ileus; Gemini's 7 errors included 4 classified as non-specific mechanical obstruction, 2 as Ogilvie syndrome, and 1 as adhesive obstruction. For adhesive/bridle obstruction (n = 18), accuracy ranged from 61.1% to 77.8% across systems — multi-agent 77.8% (14/18, 95% CI: 52.4–93.6%), ChatGPT 66.7% (12/18, 95% CI: 41.0–86.7%), Gemini 61.1% (11/18, 95% CI: 35.7–82.7%) — without statistically significant pairwise differences (all p > 0.20; post hoc power: 0.31 for detecting a 16-percentage-point difference at this sample size). A similar pattern emerged in other mechanical causes (n = 9; internal hernias, intussusception, gallstone ileus, Meckel's band, bezoar), where the multi-agent system achieved 66.7% (6/9, 95% CI: 29.9–92.5%) versus 55.6% (5/9, 95% CI: 21.2–86.3%) for both ChatGPT and Gemini, with wide confidence intervals precluding meaningful comparison. Ogilvie syndrome (n = 37) proved diagnostically challenging across all systems, though a gradient favoring the multi-agent architecture was apparent: accuracy was 67.6% (25/37, 95% CI: 50.2–82.0%) for the multi-agent system, 51.4% (19/37, 95% CI: 34.4–68.0%) for Gemini, and 48.6% (18/37, 95% CI: 32.0–65.6%) for ChatGPT. The multi-agent versus ChatGPT comparison approached significance (p = 0.072; ARD: 18.9%, 95% CI: −1.6 to 39.5%), while multi-agent versus Gemini did not (p = 0.14). The dominant misclassification pattern was labeling pseudo-obstruction as mechanical large bowel obstruction (multi-agent: 8/12 errors, 66.7%; ChatGPT: 14/19, 73.7%; Gemini: 12/18, 66.7%), reflecting the fundamental clinical difficulty of this differential when CT demonstrates marked colonic dilation without a definitive transition point. Paralytic ileus (n = 20) showed a comparable trend — multi-agent 75.0% (15/20, 95% CI: 50.9–91.3%) versus 65.0% (13/20, 95% CI: 40.8–84.6%) for both ChatGPT and Gemini (all p > 0.30) — with misdiagnoses predominantly involving confusion between paralytic ileus and early mechanical obstruction (multi-agent 3/5; ChatGPT 5/7; Gemini 4/7). The most uniformly difficult categories were toxic megacolon/severe colonic distension (n = 12) and rare etiologies (n = 9). For toxic megacolon, accuracy was 50.0% (6/12, 95% CI: 21.1–78.9%) for the multi-agent system and 41.7% (5/12, 95% CI: 15.2–72.3%) for both comparators (all p > 0.60), with errors driven by difficulty distinguishing toxic megacolon from severe Ogilvie syndrome (multi-agent: 4/6 errors) or fulminant colitis without megacolon criteria (ChatGPT: 3/7; Gemini: 4/7). For rare etiologies, all three systems showed identical accuracy of 33.3% (3/9, 95% CI: 7.5–70.1%); the 6 shared misdiagnoses included 2 internal hernias labeled as adhesive obstruction, 2 gallstone ileus cases classified as simple small bowel obstruction, and 2 intussusception cases labeled as non-specific mechanical obstruction. The convergent failure across architectures in these two categories suggests that both toxic megacolon discrimination and rare etiology identification represent fundamental limitations of current AI systems irrespective of design, likely attributable to their low prevalence in training corpora and the absence of reliable pathognomonic distinguishing features. 3.4 Processing Time and Operational Characteristics Mean processing times from prompt submission to complete response generation were: ChatGPT 8.3 ± 2.1 seconds (median: 7.8 seconds; IQR: 6.7–9.4 seconds; range: 4.2–14.8 seconds), Gemini 6.7 ± 1.8 seconds (median: 6.2 seconds; IQR: 5.4–7.6 seconds; range: 3.1–12.3 seconds), and multi-agent system 47.2 ± 11.6 seconds (median: 44.8 seconds; IQR: 39.1–53.6 seconds; range: 28.4–82.1 seconds). Processing time distributions were non-normal for all three systems (Shapiro-Wilk p < 0.01 for all), with rightward skew reflecting occasional prolonged responses for complex cases. The multi-agent system required approximately 5.7-fold longer than ChatGPT (Wilcoxon rank-sum p < 0.001; rank-biserial correlation r = 1.00) and 7.0-fold longer than Gemini ( p < 0.001; r = 1.00). Gemini demonstrated significantly faster response generation than ChatGPT ( p < 0.001; r = 0.72). The increased processing time of the multi-agent system was attributable to its sequential three-stage architecture: the Hulu-Med radiology perception agent required a mean of 18.4 ± 5.2 seconds (range: 9.1–34.7 seconds), the Med-PaLM 2 clinical synthesis agent 16.8 ± 4.1 seconds (range: 10.2–28.3 seconds), and the Gyan LLM symbolic validation agent 12.0 ± 3.8 seconds (range: 6.8–24.1 seconds). The Hulu-Med agent demonstrated the longest and most variable processing times, with longer times associated with cases containing multiple imaging modalities or complex radiological findings requiring detailed feature extraction. Intermediate structured data transfer between pipeline stages via JSON schema accounted for the remaining overhead. Despite this increased latency, the total processing time of less than 90 seconds in all 133 cases remains well within clinically acceptable limits for decision support in acute abdominal presentations, where the clinical decision timeline typically spans 30–60 minutes from initial emergency department assessment to final disposition decision. The multi-agent processing time also compares favorably with a formal multidisciplinary team consultation in real-world clinical practice, which typically requires 15–30 minutes of specialist coordination. 3.5 Inter-Rater Reliability The reliability of the evaluation framework was confirmed through comprehensive inter-rater agreement analysis. Cohen's kappa coefficients for agreement between the two independent assessors across all five evaluation criteria were: diagnostic accuracy κ = 0.95 (95% CI: 0.91–0.99; classified as excellent by Landis and Koch criteria; percentage agreement: 97.0%, 129/133 concordant evaluations per system), treatment appropriateness κ = 0.88 (95% CI: 0.82–0.94; excellent; percentage agreement: 91.0%, 121/133), hallucination detection κ = 0.91 (95% CI: 0.85–0.97; excellent; percentage agreement: 94.7%, 126/133), explanation adequacy κ = 0.82 (95% CI: 0.75–0.89; excellent; percentage agreement: 87.2%, 116/133), and critical safety errors κ = 0.93 (95% CI: 0.86–1.00; excellent; percentage agreement: 94.7%, 126/133). The overall inter-rater reliability across all 1,995 evaluations was κ = 0.89 (95% CI: 0.84–0.94; overall percentage agreement: 92.9%). The slightly lower agreement for explanation adequacy ( κ = 0.82) compared to other domains reflects the inherently more subjective nature of assessing reasoning quality, which involves evaluative judgment regarding the depth and coherence of clinical argumentation rather than binary classification. Among the 47 total disagreements (47/1,995, 2.4%), the distribution by evaluation domain was: diagnostic accuracy 4 disagreements (8.5%), treatment appropriateness 12 (25.5%), hallucination detection 7 (14.9%), explanation adequacy 17 (36.2%), and critical safety errors 7 (14.9%). All disagreements were resolved to consensus by the third reviewer. 4. DISCUSSION This study provides the first empirical evidence that a neurosymbolic multi-agent architecture integrating domain-specific perception, clinical synthesis, and symbolic verification agents in a sequential pipeline outperforms general-purpose large language models in the diagnosis and management of ileus-spectrum and volvulus conditions. The neurosymbolic system achieved 75.2% diagnostic accuracy (95% CI: 66.9–82.2%), representing a statistically significant 15.0 and 16.5 percentage-point improvement over ChatGPT (GPT-4 Turbo; 60.2%) and Gemini 2.0 Pro (58.6%), respectively (both p < 0.001). The multi-agent system also demonstrated a substantially lower hallucination rate of 1.5% with no high-severity events, compared to 15.0% for ChatGPT and 9.8% for Gemini, and produced no safety errors, whereas ChatGPT and Gemini yielded rates of 3.8% and 2.3%, respectively. These findings support the hypothesis that decomposing clinical reasoning into specialized cognitive subtasks, consistent with multidisciplinary team workflows, yields more reliable and safer AI-assisted decision support than monolithic general-purpose models. The diagnostic performance of 60.2% and 58.6% observed for ChatGPT and Gemini in our study is broadly consistent with performance ranges reported for general-purpose LLMs across medical specialties. Sussan et al. reported that both GPT-4 Turbo and Gemini-Pro achieved variable accuracy across medical licensing examinations, with performance declining substantially in clinically complex scenarios requiring multimodal integration [ 6 ]. A systematic review of diagnostic accuracy of large language models in clinical settings found that diagnostic performance varied widely across studies, with primary diagnostic accuracy ranging from approximately 25% to 97.8% depending on task complexity and model evaluated, suggesting substantial heterogeneity in clinical performance among LLMs [ 17 ]. Similarly, Mittal and Aggarwal demonstrated that LLMs exhibited diagnostic accuracy of 52–68% in ophthalmic emergencies, with particular difficulty in cases requiring integration of imaging findings with clinical context [ 9 ]. Hager et al., using 2,400 real patient cases from the MIMIC-IV database across four common abdominal pathologies, showed that state-of-the-art LLMs performed significantly worse than physicians in autonomous clinical decision-making and failed to follow diagnostic or treatment guidelines [ 18 ]. Mansoor et al., in a comprehensive systematic review of LLM reasoning techniques in medicine published in Health Information Science and Systems, further emphasized that while LLMs demonstrate unprecedented capabilities in medical reasoning tasks requiring complex inference and pattern recognition, their performance deteriorates significantly in high-stakes clinical scenarios demanding structured multi-step reasoning and guideline adherence [ 19 ]. Our findings extend this body of evidence to the domain of acute intestinal obstruction, confirming that general-purpose LLMs, despite their broad medical knowledge, remain insufficiently reliable for clinical decision support in conditions where diagnostic error carries immediate surgical consequences. The superior performance of the neurosymbolic multi-agent system can be attributed to its architectural design, which operationalizes two complementary principles: task decomposition through specialized agents and symbolic verification through deterministic rule-based reasoning. Prenosil et al. demonstrated that a neurosymbolic AI combining GPT-4 with a rule-based expert system through a semantic integration platform achieved physician-level accuracy (99.8%) in extracting structured clinical data from radiology reports, with the symbolic component providing auditable verification trails [ 10 ]. Acharya and Song provided a comprehensive theoretical framework demonstrating that neurosymbolic integration enhances robustness, uncertainty quantification, and intervenability—three properties essential for clinical deployment [ 11 ]. Miladinovic et al. applied a neurosymbolic framework to retinal disease classification from OCT images and demonstrated that symbolic constraints improved both explainability and diagnostic consistency compared to purely neural approaches [ 12 ]. Our three-stage pipeline, comprising Hulu-Med 32B for radiological perception, Med-PaLM 2 for clinical synthesis, and Gyan LLM for symbolic validation, represents a practical implementation of neurosymbolic principles in a clinically relevant domain. Adnan et al. recently demonstrated that neurosymbolic digital twin architectures combining neural pattern recognition with symbolic reasoning achieve superior performance in cardiovascular disease prediction and personalized modeling, further validating the translational potential of neurosymbolic approaches across diverse clinical domains [ 20 ]. The multi-agent architecture employed in this study aligns with an emerging paradigm in medical AI that leverages collaborative agent systems for complex clinical reasoning. Sorka et al. demonstrated that multi-agent approaches to neurological clinical reasoning, where specialized agents handle distinct cognitive subtasks, consistently outperformed single-model configurations in diagnostic accuracy and reasoning quality [ 14 ]. Chen et al., in a landmark study published in npj Digital Medicine , showed that multi-agent conversational LLM systems enhanced diagnostic capability by enabling iterative refinement through inter-agent dialogue, achieving significant improvements over standalone models [ 21 ]. A recent survey of LLM-based multi-agent systems in medicine confirmed that multi-agent frameworks, including the MAC framework and GPT-4-based voting ensembles, consistently outperform single-agent setups in complex diagnostic reasoning, highlighting the critical role of collaborative mechanisms in optimizing clinical reliability [ 22 ]. Our own previous work comparing multidisciplinary AI systems versus single-model approaches for ileus and volvulus diagnosis provided preliminary evidence for the superiority of multi-agent architectures in this specific clinical domain [ 15 ]. The present study extends these findings by incorporating a symbolic verification layer that further enhances factual grounding and safety. Perhaps the most clinically significant finding of this study is the dramatic reduction in hallucination rates achieved by the neurosymbolic system. The 1.5% hallucination rate (2/133 cases, both minor inferential assumptions) represents a 90% reduction compared to ChatGPT (15.0%) and an 85% reduction compared to Gemini (9.8%). This finding is particularly notable given the severity of LLM hallucination in clinical contexts. Omar et al., in a large-scale evaluation, tested six leading LLMs with 300 physician-designed clinical vignettes containing fabricated medical details and found that every tested model repeated or elaborated on planted false information in 50–82% of outputs, with even the best-performing model (GPT-4o) exhibiting a 23% hallucination rate under targeted mitigation prompts [ 23 ]. Hazra et al. reported that hallucination rates in image-based medical tasks remained alarmingly high across commercial LLMs, with fabricated imaging findings representing a particularly dangerous category [ 7 ]. Roustan et al. provided a comprehensive clinician-oriented review emphasizing that LLM hallucinations in healthcare—including symptom fabrication, laboratory value invention, and imaging finding distortion—represent the most significant barrier to clinical deployment [ 16 ].The Gyan LLM symbolic validation agent in our pipeline addresses this challenge through deterministic factual grounding verification, cross-referencing each claim in the clinical output against the original case data and flagging unsupported assertions for correction. This compositional hallucination detection mechanism explains the virtual elimination of high-severity hallucinations in our system. Subgroup analysis revealed important disease-specific performance patterns. The most pronounced advantage of the neurosymbolic system was observed in volvulus cases (100% accuracy vs. 78.6% and 75.0% for ChatGPT and Gemini, respectively; p < 0.01), primarily attributable to the Hulu-Med radiology agent's ability to accurately identify pathognomonic imaging signs including the coffee-bean sign, whirl sign, and bird-beak sign. This finding is consistent with the radiological literature emphasizing the critical diagnostic value of these signs. Memis and Aydin demonstrated that sigmoid volvulus subtype classification based on imaging findings significantly impacts clinical course prediction, while Moloney et al. showed that specific CT features can predict volvulus outcomes and recurrence [ 5 , 24 ]. The specialized vision-language training of Hulu-Med across 12 anatomical systems and 14 imaging modalities provides a natural advantage in recognizing these pathognomonic patterns, an advantage that general-purpose text-based LLMs inherently lack. Conversely, all systems demonstrated reduced accuracy for Ogilvie syndrome (67.6%, 48.6%, 51.4%) and rare etiologies (33.3% across all systems), reflecting the intrinsic diagnostic challenge of distinguishing functional from mechanical obstruction and the limited representation of uncommon conditions in training datasets. The absence of critical safety errors in the neurosymbolic system outputs (0/133 cases) compared to 5 errors for ChatGPT (3.8%) and 3 for Gemini (2.3%) merits particular emphasis. All eight critical errors across both monolithic LLMs involved failure to recognize surgical emergencies—including recommending conservative management for cecal volvulus, suggesting endoscopic intervention for sigmoid volvulus with strangulation signs, and recommending only prokinetic agents for Ogilvie syndrome with critical cecal diameter exceeding 12 cm. Recent reviews have confirmed that state-of-the-art medical LLMs continue to exhibit substantial hallucination risks and challenges in clinical tasks, underscoring the need for rigorous validation, bias mitigation, and multimodal integration to ensure safe deployment in healthcare settings [ 7 ]. The three-tier verification framework of our pipeline includes radiological assessment, clinical synthesis, and symbolic validation. This structure introduces multiple safety checkpoints that are not present in single-model systems and may explain the complete elimination of critical safety errors observed in our study. The multi-agent system demonstrated higher explanation adequacy (80.5% compared with 69.2% for ChatGPT and 58.6% for Gemini), reflecting the inherent transparency of its pipeline architecture. The explanatory framework consists of a radiological findings summary generated by Hulu-Med, pathophysiological reasoning provided by Med-PaLM 2, and validation points from Gyan. This format resembles a multidisciplinary team consultation report and offers clinicians a structured audit trail for each diagnostic decision.This design philosophy aligns with the growing emphasis on explainable AI in healthcare. Recent evidence emphasizes that transparency and explainability are essential for clinical AI systems, as clinicians must be able to understand, verify, and critically appraise model reasoning to ensure safe and trustworthy decision support [ 25 ]. The neurosymbolic paradigm, by structurally separating the knowledge base from the inference engine, enables this level of transparency in ways that opaque end-to-end neural models fundamentally cannot provide The multi-agent system's increased processing time (47.2 ± 11.6 seconds vs. 8.3 ± 2.1 and 6.7 ± 1.8 seconds for ChatGPT and Gemini) represents an inherent trade-off of sequential multi-agent architectures. However, this latency remains well within clinically acceptable limits for decision support in acute abdominal presentations, where the clinical decision timeline typically spans 30–60 minutes from initial assessment to disposition. Furthermore, this processing time compares favorably with the time required for a complete multidisciplinary consultation in a real-world clinical setting. Future optimization through parallel processing of independent pipeline stages could substantially reduce this latency without sacrificing the sequential verification logic that underpins the system's safety profile. Several limitations of this study should be acknowledged. First, the retrospective, case report-based design introduces inherent selection bias, as published case reports may overrepresent unusual or diagnostically challenging presentations and underrepresent routine cases. Second, the use of published case reports rather than real-time clinical encounters limits ecological validity; AI systems may perform differently when processing structured case vignettes versus unstructured electronic health record data. Third, the single-center evaluation with two expert assessors, despite excellent inter-rater reliability (κ = 0.89), may not fully capture the variability in clinical judgment across different institutional settings and cultural contexts. Fourth, the imaging data were provided as textual descriptions derived from published reports rather than as raw imaging files, potentially underestimating the Hulu-Med agent's full multimodal capabilities. Fifth, the 133-case sample size, although sufficient for primary comparisons (post-hoc power > 0.90 for 15% absolute difference at α = 0.05), limits statistical power for subgroup analyses, particularly in rare diagnostic categories. Sixth, the study evaluates AI performance in isolation rather than as an adjunct to human decision-making, which represents the most likely deployment scenario. Seventh, the absence of prospective clinical validation means that the real-world impact on patient outcomes, clinical workflows, and physician decision-making remains unknown. Finally, the rapidly evolving nature of both general-purpose LLMs and neurosymbolic architectures means that performance benchmarks may shift substantially with future model iterations. Future research should prioritize multicenter prospective validation studies incorporating diverse patient populations, real-time clinical data inputs including raw imaging files, and physician-AI collaborative decision-making paradigms. Integration with electronic health record systems would enable evaluation of the system's performance in authentic clinical workflows with naturally occurring data noise and incompleteness. Cost-effectiveness analyses comparing computational costs against diagnostic accuracy improvements and potential clinical complication reduction are warranted. The development of hybrid deployment models—where the neurosymbolic system serves as a real-time decision support layer augmenting rather than replacing physician judgment—represents the most promising translational pathway. Additionally, expanding the symbolic knowledge base to incorporate continuously updated clinical guidelines and extending the multi-agent framework to other acute abdominal pathologies and surgical emergencies could broaden the system's clinical utility. 5. CONCLUSIONS A neurosymbolic multi-agent pipeline that decomposes the clinical reasoning workflow into specialized perception, synthesis, and symbolic verification stages significantly outperforms general-purpose monolithic LLMs in diagnosing and managing ileus-spectrum and volvulus-spectrum emergencies. The architectural separation of neural pattern recognition from symbolic rule-based verification substantially reduces hallucination and eliminates critical safety errors. For volvulus cases specifically, where pathognomonic radiological signs enable definitive diagnosis, specialized vision-language perception agents achieve perfect diagnostic accuracy. However, performance advantages diminish for diagnostically ambiguous conditions such as Ogilvie syndrome and toxic megacolon, where even structured multi-agent reasoning cannot fully compensate for inherent diagnostic complexity. These findings support the integration of neurosymbolic design principles—combining neural perception with symbolic verification—in clinical AI systems for acute abdominal pathology, while underscoring that AI outputs in these high-stakes settings must remain subject to expert physician oversight and verification until reliability is consistently demonstrated across the full spectrum of diagnostic complexity. Declarations Ethics Approval and Consent to Participate : This study was approved by the Ankara Provincial Directorate of Health Non-Interventional Ethics Committee (Approval No: 2025-10-3; Date: 24 October 2025). The study protocol entitled "The Role of a Specific Local Large Language Model in the Diagnosis of Internal Medicine Diseases" was reviewed and approved after evaluation of the study rationale, objectives, methodology, and ethical aspects. Given the retrospective design utilizing anonymized data extracted from previously published case reports in the peer-reviewed literature, the requirement for individual informed consent was waived by the ethics committee. All procedures were conducted in accordance with the Declaration of Helsinki (2013 revision) and relevant national regulations. Consent for Publication : Not applicable. This study exclusively utilized data from previously published, anonymized case reports available in the public domain. No individual participant data requiring consent for publication were included. Availability of Data and Materials : The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request. The 133 case vignettes were reconstructed from PubMed-indexed case reports published between January 2022 and December 2025; the complete list of source publications with PubMed identifiers (PMIDs) is provided in Supplementary Table S1. The standardized prompt templates used for AI system evaluation, the structured data extraction forms, and the raw evaluation scores from both independent assessors are available as supplementary materials. The AI-generated outputs from all three systems (ChatGPT GPT-4 Turbo, Gemini 2.0 Pro, and the neurosymbolic multi-agent system) are archived and available upon request for reproducibility verification. The neurosymbolic multi-agent system pipeline configuration, including agent-specific prompts and inter-agent JSON communication schemas, is described in detail in the Methods section; the complete technical implementation files are available from the corresponding author. Due to the use of proprietary AI platforms (ChatGPT and Gemini), full computational reproducibility is subject to model version availability and API access at the time of replication. Competing Interests :The authors declare that they have no competing interests. No financial or non-financial conflicts of interest exist in relation to the content of this manuscript. None of the authors have any affiliation with or financial involvement in any organization or entity with a direct financial interest in the subject matter or materials discussed in this manuscript, including OpenAI (developer of ChatGPT), Google DeepMind (developer of Gemini and Med-PaLM 2), or the developers of Hulu-Med or Gyan LLM. Funding : This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors. Authors' Contributions : M.U. conceptualized and designed the study, developed the neurosymbolic multi-agent system architecture, conducted the systematic literature search, performed data extraction, executed all AI system evaluations, performed statistical analyses, interpreted the results, and drafted the manuscript. S.K. contributed to data extraction, independently evaluated AI system outputs as a blinded assessor, participated in inter-rater reliability assessment, and critically revised the manuscript for intellectual content. K.Y. contributed to case selection and screening, served as an independent blinded assessor for AI output evaluation, resolved inter-rater disagreements through consensus discussion, and critically revised the manuscript. L.E. contributed to data collection, assisted with figure and table preparation, and critically reviewed the final manuscript. All authors read and approved the final version of the manuscript and agree to be accountable for all aspects of the work. Acknowledgements Not applicable. References Mwenitete, D., et al., Determinants of surgical management outcomes among adult patients with intestinal obstruction at Mzuzu central hospital, Malawi. BMC Surg, 2025. 26 (1): p. 50. Inoue, K., et al., Surgical Management of Sigmoid Volvulus: A Retrospective Review of Six Cases with a Focus on the Sharon Operation. Surg Case Rep, 2026. 12 (1). Memis, K.B. and S. Aydin Relationship Between Sigmoid Volvulus Subtypes, Clinical Course, and Imaging Findings . Diagnostics, 2025. 15 , 784 DOI: 10.3390/diagnostics15060784. Larsen, T.B. and M.E. Lazarus, Coffee Bean Sign. J Brown Hosp Med, 2025. 4 (3): p. 137903. Moloney, B.M., et al., Sigmoid volvulus-Can CT features predict outcomes and recurrence? Eur Radiol, 2025. 35 (2): p. 897-905. Sussan, T.T., et al., A Comparative Evaluation of GPT-4 Turbo and Gemini-Pro in Medical Licensing Exams: Enhancing Artificial Intelligence's Role in Medical Education. Cureus, 2026. 18 (1): p. e101101. Hazra, D., et al., Evaluating Hallucination and Diagnostic Reliability of LLMs on Medical Image-Based Multiple Choice Tasks. IEEE J Biomed Health Inform, 2025. Pp . Bradshaw, T.J., et al., Large Language Models and Large Multimodal Models in Medical Imaging: A Primer for Physicians. J Nucl Med, 2025. 66 (2): p. 173-182. Mittal, S. and Y. Aggarwal, Evaluation of Large Language Models in the Diagnosis, Urgency Triage, and Initial Management of Ophthalmic Emergencies. Cureus, 2026. 18 (1): p. e101433. Prenosil, G.A., et al., Neuro-symbolic AI for auditable cognitive information extraction from medical reports. Commun Med (Lond), 2025. 5 (1): p. 491. Acharya, K. and H. Song, A Comprehensive Review of Neuro-symbolic AI for Robustness, Uncertainty Quantification, and Intervenability. Arabian Journal for Science and Engineering, 2026. 51 (1): p. 35-67. Miladinovic, A., et al., Neurosymbolic AI Framework for Explainable Retinal Disease Classification From OCT Images. Transl Vis Sci Technol, 2026. 15 (1): p. 6. Prenosil, G.A., et al., Neuro-symbolic AI for auditable cognitive information extraction from medical reports. Communications Medicine, 2025. 5 (1): p. 491. Sorka, M., et al., A multi-agent approach to neurological clinical reasoning. PLOS Digit Health, 2025. 4 (12): p. e0001106. Ucdal, M., K. Yurtsever, and E. Ekingen, Multidisciplinary artificial intelligence systems versus single-model approaches for the diagnosis and management of ileus and volvulus. BMC Gastroenterol, 2026. 26 (1): p. 124. Roustan, D. and F. Bastardot, The Clinicians' Guide to Large Language Models: A General Perspective With a Focus on Hallucinations. Interact J Med Res, 2025. 14 : p. e59823. Shan, G., et al., Comparing Diagnostic Accuracy of Clinical Professionals and Large Language Models: Systematic Review and Meta-Analysis. JMIR Med Inform, 2025. 13 : p. e64963. Hager, P., et al., Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nat Med, 2024. 30 (9): p. 2613-2622. Mansoor, I., et al., Reasoning with large language models in medicine: a systematic review of techniques, challenges and clinical integration. Health Inf Sci Syst, 2026. 14 (1): p. 6. Adnan, M., et al., Neurosymbolic Digital Twin for Cardiovascular Disease Prediction and Personalized Modeling. IEEE J Biomed Health Inform, 2025. Pp . Chen, X., et al., Enhancing diagnostic capability with multi-agents conversational large language models. NPJ Digit Med, 2025. 8 (1): p. 159. Xu, X. and R. Sankar, Large Language Model Agents for Biomedicine: A Comprehensive Review of Methods, Evaluations, Challenges, and Future Directions. Information, 2025. 16 (10): p. 894. Omar, M., et al., Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Commun Med (Lond), 2025. 5 (1): p. 330. Memis, K.B. and S. Aydin, Relationship Between Sigmoid Volvulus Subtypes, Clinical Course, and Imaging Findings. Diagnostics (Basel), 2025. 15 (6). Martinho, D., et al., Ethical Responsibility in Medical AI: A Semi-Systematic Thematic Review and Multilevel Governance Model. Healthcare (Basel), 2026. 14 (3). Additional Declarations No competing interests reported. Supplementary Files SupplementaryTableS1final.xlsx Graphicalabstract.jpg Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 25 Apr, 2026 Reviews received at journal 22 Apr, 2026 Reviewers agreed at journal 15 Apr, 2026 Reviewers agreed at journal 12 Apr, 2026 Reviews received at journal 12 Apr, 2026 Reviewers agreed at journal 12 Apr, 2026 Reviewers agreed at journal 09 Apr, 2026 Reviewers invited by journal 03 Apr, 2026 Editor invited by journal 10 Mar, 2026 Editor assigned by journal 06 Mar, 2026 Submission checks completed at journal 06 Mar, 2026 First submitted to journal 05 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9045948","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":617260261,"identity":"4cea0407-e68b-445f-8983-312daec1f48a","order_by":0,"name":"mete ucdal","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABDklEQVRIiWNgGAWjYPACZgY+9h5kATb8qsEEG88ZEMNAAq6Fh6AWiRwitei29x98zFNjLc8m+fbghx8Vf+r4pZsPMHwoO8xgL30AqxazM4eZjXmOpRu2SeclS/acMZCQnHMsgXHGucMMPHwJ2LXcSGaT5m04zNgmnWMgzdhmIGFwI8eAmbcNqAWHy8zuP2b/DdRi3yZ5xvg34z8DCfsb+R+Y/+LTcoOZjRmoJbFNgsdMmrEBaAswHJgZ8Wk5k2wMdH16chtPjpllzzFjyRl3jhkc7DmXzgMJdSxajh98+OFNjbVtP/sZ4xs/auT4+Wc3P3zwo8xaDjVu8QJgzBxgwBeTWLWMglEwCkbBKEAGABXmVILfQJ1vAAAAAElFTkSuQmCC","orcid":"","institution":"Etimesgut Asker Hastanesi","correspondingAuthor":true,"prefix":"","firstName":"mete","middleName":"","lastName":"ucdal","suffix":""},{"id":617260262,"identity":"c2f8607e-7b63-4b2b-9403-849e2db1d797","order_by":1,"name":"Sefa Keskin","email":"","orcid":"","institution":"Beytepe Asker Hastanesi","correspondingAuthor":false,"prefix":"","firstName":"Sefa","middleName":"","lastName":"Keskin","suffix":""},{"id":617260263,"identity":"b2c422e3-b130-4f2c-b204-da66b881354a","order_by":2,"name":"karya yurtsever","email":"","orcid":"","institution":"Hacettepe University Hospital","correspondingAuthor":false,"prefix":"","firstName":"karya","middleName":"","lastName":"yurtsever","suffix":""},{"id":617260264,"identity":"ea720f8b-9fd1-4181-acfb-30c247874511","order_by":3,"name":"Leyla Eybatova","email":"","orcid":"","institution":"Başkent University Hospital","correspondingAuthor":false,"prefix":"","firstName":"Leyla","middleName":"","lastName":"Eybatova","suffix":""}],"badges":[],"createdAt":"2026-03-06 04:38:20","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9045948/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9045948/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":106726606,"identity":"bf02ed82-5b6d-4d61-b575-70e145e9f1f5","added_by":"auto","created_at":"2026-04-12 18:36:48","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":110267,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eNeurosymbolic Multi-Agent System Architecture. \u003c/strong\u003eSchematic diagram illustrating the sequential three-agent pipeline: (Stage 1) Hulu-Med 32B Radiology Perception Agent (neural vision-language processing); (Stage 2) Med-PaLM 2 Clinical Reasoning Agent (neural clinical synthesis); (Stage 3) Gyan LLM Symbolic Validation Agent (symbolic rule-based verification, hallucination detection, and safety checking). Inter-agent communication is shown via structured JSON schemas. The neurosymbolic boundary between neural and symbolic components is explicitly demarcated.\u003c/p\u003e","description":"","filename":"1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/b6819a347aa988e1ca604545.jpg"},{"id":106726927,"identity":"585f4328-25e5-44a6-8bb1-5487bcc18c0d","added_by":"auto","created_at":"2026-04-12 18:37:42","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":103223,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePRISMA Flow Diagram. \u003c/strong\u003eFlowchart illustrating the systematic case selection process from initial database search (847 records) through title/abstract screening (412 eligible), full-text review, and final inclusion (133 cases). Exclusion reasons at each stage are detailed.\u003c/p\u003e","description":"","filename":"2.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/e11e382e2331c875bac01d52.jpg"},{"id":107479244,"identity":"53bc0792-1c0e-4bb2-bf1e-3e1c969b94ac","added_by":"auto","created_at":"2026-04-22 01:21:06","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":51508,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eComparative Diagnostic Accuracy by Disease Category (Clustered Bar Chart). \u003c/strong\u003eGrouped bar chart comparing diagnostic accuracy (%) of ChatGPT (GPT-4 Turbo), Gemini 2.0 Pro, and the neurosymbolic multi-agent system across seven diagnostic categories: volvulus, other mechanical obstruction, Ogilvie syndrome, paralytic ileus, toxic megacolon, rare etiologies, and overall. Error bars represent 95% confidence intervals. Statistical significance markers (*) indicate p \u0026lt; 0.05 by Fisher’s exact test.\u003c/p\u003e","description":"","filename":"3.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/9af5d60eef4c7837fb3a9038.jpg"},{"id":106726600,"identity":"8b7a6f2f-1ff9-42a3-9464-1f4d99d91c30","added_by":"auto","created_at":"2026-04-12 18:36:42","extension":"jpg","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":43711,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHallucination Profile Heatmap.\u003c/strong\u003e Heatmap illustrating the distribution of hallucination subtypes (symptom fabrication, imaging finding distortion, laboratory value invention, medical history addition, and minor inferential assumption) across ChatGPT (GPT-4 Turbo), Gemini 2.0 Pro, and the neurosymbolic multi-agent system (N = 133 cases each). Cell color intensity corresponds to case count. The multi-agent system achieved a 98.5% hallucination-free rate (2/133), with only minor inferential assumptions and zero high-severity hallucinations, compared to ChatGPT (15.0%, 20/133) and Gemini (9.8%, 13/133) (both \u003cem\u003ep\u003c/em\u003e \u0026lt; 0.001, Fisher's exact test).\u003c/p\u003e","description":"","filename":"4.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/86a838331a787f46df083de8.jpg"},{"id":106726204,"identity":"ed9bcb06-9de0-4567-855a-fdb5efeb602a","added_by":"auto","created_at":"2026-04-12 18:35:34","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":44306,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRadar (Spider) Chart of Multi-Dimensional AI Performance. \u003c/strong\u003eRadar chart displaying the five evaluation dimensions (diagnostic accuracy, treatment appropriateness, hallucination-free rate [100% − hallucination rate], explanation adequacy, and safety [100% − critical error rate]) for each of the three AI systems. The multi-agent system’s polygon area substantially exceeds those of both monolithic LLMs across all dimensions.\u003c/p\u003e","description":"","filename":"5.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/e76a7054710a5cef1a9d8f00.jpg"},{"id":107479251,"identity":"584e8f55-2a43-4540-97cd-368dfe5573f1","added_by":"auto","created_at":"2026-04-22 01:21:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1439665,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/8ccafe2b-a2f0-42b8-903e-990dac13bf80.pdf"},{"id":106635677,"identity":"12e68231-c5b7-4f45-9d08-2219e425c4e6","added_by":"auto","created_at":"2026-04-10 16:49:24","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":19925,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryTableS1final.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/43123706961b880f9708e4df.xlsx"},{"id":106635678,"identity":"11028ebf-f257-4a7e-924a-442d06dd3ab2","added_by":"auto","created_at":"2026-04-10 16:49:24","extension":"jpg","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":132871,"visible":true,"origin":"","legend":"","description":"","filename":"Graphicalabstract.jpg","url":"https://assets-eu.researchsquare.com/files/rs-9045948/v1/560f5f95855a3e48793e5176.jpg"}],"financialInterests":"No competing interests reported.","formattedTitle":"Neurosymbolic Multi-Agent Artificial Intelligence versus General-Purpose Large Language Models for Clinical Decision Support in Ileus and Volvulus","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cp\u003eIntestinal obstruction, encompassing both mechanical ileus and volvulus, represents a spectrum of acute abdominal emergencies that collectively account for approximately 12\u0026ndash;16% of emergency surgical admissions and carry mortality rates ranging from 3% for uncomplicated adhesive obstruction to 30\u0026ndash;40% for strangulated volvulus with bowel necrosis [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. The diagnostic complexity of these conditions arises from their overlapping clinical presentations, the multiplicity of underlying etiologies, and the critical dependence of patient outcomes on the timeliness and accuracy of diagnostic and therapeutic decision-making [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. Sigmoid volvulus, the most common type of colonic volvulus, is characterized by pathognomonic radiological findings such as the coffee-bean sign, omega loop, and whirl sign on computed tomography [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. Despite these features, it is frequently misclassified at initial presentation, particularly when its clinical and radiological appearance overlaps with functional conditions including Ogilvie syndrome (acute colonic pseudo-obstruction) and toxic megacolon[\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe emergence of large language models (LLMs) as clinical decision support tools has generated considerable interest across medical specialties, with general-purpose models such as OpenAI\u0026rsquo;s GPT-4 and Google DeepMind\u0026rsquo;s Gemini demonstrating variable performance on standardized medical examinations and clinical reasoning tasks [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. However, the application of monolithic LLMs to complex diagnostic scenarios requiring integration of multimodal data\u0026mdash;including radiological image interpretation, laboratory value synthesis, and contextual clinical reasoning\u0026mdash;has consistently revealed critical limitations [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e]. These include a propensity for hallucination (the generation of plausible but factually unsupported clinical assertions), inadequate recognition of surgical urgency indicators, and failure to maintain the chain of clinical reasoning necessary for safe disposition decisions [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eRecent advances in neurosymbolic artificial intelligence (NeSy-AI) offer a promising architectural paradigm for addressing these limitations. Neurosymbolic systems integrate the pattern-recognition strengths of neural networks with the logical rigor, transparency, and rule-based verification capabilities of symbolic AI [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. In the medical domain, this hybrid approach enables the construction of systems that can simultaneously leverage deep learning for complex perceptual tasks (such as radiological image analysis) while maintaining explicit, auditable reasoning pathways grounded in established clinical guidelines and medical ontologies [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. The neurosymbolic framework is particularly well-suited to surgical decision support, where diagnostic accuracy must be coupled with explicit justification, safety verification, and alignment with evidence-based treatment protocols[\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eMulti-agent architectures represent a natural implementation of neurosymbolic principles in clinical AI, whereby distinct specialized agents\u0026mdash;each optimized for a specific cognitive subtask\u0026mdash;collaborate in a structured pipeline that mirrors the multidisciplinary team (MDT) consultation model prevalent in modern surgical practice [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. By decomposing the diagnostic workflow into perception (radiological analysis), synthesis (clinical reasoning and differential diagnosis), and verification (hallucination detection and safety checking) stages, multi-agent systems can theoretically overcome the limitations of end-to-end monolithic models while providing transparent, traceable decision pathways [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDespite the theoretical promise of neurosymbolic multi-agent approaches, empirical evidence comparing their diagnostic performance against established general-purpose LLMs in acute abdominal pathology remains scarce. In recent years, multimodal large language model\u0026ndash;based approaches have increasingly been explored for complex clinical decision-making tasks, including acute gastrointestinal pathologies. However, to our knowledge, no study has systematically evaluated a neurosymbolic multi-agent system against monolithic LLMs specifically for the diagnosis and management of ileus-spectrum and volvulus-spectrum conditions using real-world clinical case vignettes. Nevertheless, a recent comparative study demonstrated that multidisciplinary artificial intelligence systems outperformed single-model approaches in the diagnosis and management of ileus and volvulus [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe primary objective of this study was to compare the diagnostic accuracy, treatment appropriateness, hallucination rates, explanation quality, and critical safety error profiles of three distinct AI architectural approaches\u0026mdash;a general-purpose LLM (ChatGPT/GPT-4 Turbo), a multimodal foundation model (Gemini 2.0 Pro), and a sequential neurosymbolic multi-agent hybrid system\u0026mdash;in the assessment of ileus and volvulus cases reconstructed from published case reports.\u003c/p\u003e"},{"header":"2. MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study Design and Ethical Considerations\u003c/h2\u003e \u003cp\u003eThis retrospective, observational diagnostic accuracy study was designed to evaluate and compare three distinct artificial intelligence architectures in the assessment of ileus and volvulus cases. The study utilized published case reports from peer-reviewed medical journals indexed in the PubMed database. The study protocol adhered to the Standards for Reporting Diagnostic Accuracy Studies [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e] 2015 guidelines and the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 statement for transparent reporting. The study was prospectively registered prior to data extraction.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eEthical approval\u003c/strong\u003e \u003cp\u003e for this study was obtained from the Ankara Provincial Directorate of Health Non-Interventional Ethics Committee (Approval No: 2025-10-3; Date: 24 October 2025). The protocol entitled \u0026ldquo;The Role of a Specific Local Large Language Model in the Diagnosis of Internal Medicine Diseases\u0026rdquo; was reviewed and approved after evaluation of the study rationale, objectives, methodology, and ethical aspects. Given the retrospective design and the use of anonymized data, the requirement for informed consent was waived. All procedures were conducted in accordance with the Declaration of Helsinki and relevant national regulations.\u003c/p\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Case Selection and Eligibility Criteria\u003c/h2\u003e \u003cp\u003eA systematic literature search was conducted in the PubMed/MEDLINE database to identify case reports published between January 2022 and December 2025. The search strategy employed the following Medical Subject Headings (MeSH) terms and keywords in various Boolean combinations (AND, OR): \u0026ldquo;ileus,\u0026rdquo; \u0026ldquo;volvulus,\u0026rdquo; \u0026ldquo;intestinal obstruction,\u0026rdquo; \u0026ldquo;bowel obstruction,\u0026rdquo; \u0026ldquo;Ogilvie syndrome,\u0026rdquo; \u0026ldquo;acute colonic pseudo-obstruction,\u0026rdquo; \u0026ldquo;paralytic ileus,\u0026rdquo; \u0026ldquo;mechanical obstruction,\u0026rdquo; \u0026ldquo;toxic megacolon,\u0026rdquo; \u0026ldquo;sigmoid volvulus,\u0026rdquo; \u0026ldquo;cecal volvulus,\u0026rdquo; \u0026ldquo;small bowel volvulus,\u0026rdquo; and \u0026ldquo;case report.\u0026rdquo; The search was limited to English-language publications with full-text availability.\u003c/p\u003e \u003cp\u003eInclusion criteria were defined as follows: (1) case reports published in English with full-text availability; (2) adult patients aged 18 years or older; (3) definitive diagnosis confirmed through surgical findings, histopathological examination, or conclusive clinical and radiological evidence; (4) comprehensive documentation of patient demographics, presenting symptoms, physical examination findings, laboratory values, and imaging results; (5) clear description of the treatment approach and clinical outcome. Exclusion criteria comprised: (1) cases involving multiple concurrent abdominal pathologies that could confound diagnostic assessment; (2) pediatric patients (\u0026lt;\u0026thinsp;18 years); (3) incomplete case documentation lacking essential clinical parameters required for AI evaluation; (4) duplicate publications or cases previously reported in different journals; (5) conference abstracts, case series without individual case details, or letters to the editor without sufficient clinical data.\u003c/p\u003e \u003cp\u003eThe initial database search yielded 847 potentially relevant case reports. Following title and abstract screening by two independent reviewers (M.D. and E.E.), 412 articles were deemed potentially eligible and underwent full-text review. After applying the predefined inclusion and exclusion criteria, 133 cases were ultimately included in the final analysis. The case selection process is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e (PRISMA Flow Diagram). Inter-reviewer agreement for study selection was assessed using Cohen\u0026rsquo;s kappa coefficient, which demonstrated excellent concordance (κ\u0026thinsp;=\u0026thinsp;0.94, 95% CI: 0.91\u0026ndash;0.97). Disagreements were resolved through consensus discussion with a third reviewer.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Data Extraction and Standardization\u003c/h2\u003e \u003cp\u003eA structured data extraction form was developed and piloted on 10 randomly selected cases prior to full implementation. For each included case, the following variables were systematically extracted: patient age and sex; duration of symptoms prior to presentation; chief complaints including abdominal pain characteristics (location, severity, quality), distension, nausea, vomiting, obstipation, and fever; relevant medical history and comorbidities; vital signs at presentation (systolic and diastolic blood pressure, heart rate, temperature, respiratory rate, oxygen saturation); physical examination findings with particular attention to abdominal distension severity, bowel sounds characteristics (absent, hyperactive, tinkling), tenderness location and severity, guarding, rebound tenderness, and signs of peritoneal irritation; laboratory parameters including complete blood count (white blood cell count, hemoglobin, platelet count), serum electrolytes (sodium, potassium, chloride, bicarbonate), renal function tests (blood urea nitrogen, creatinine), hepatic function tests (alanine aminotransferase, aspartate aminotransferase, alkaline phosphatase, total bilirubin), serum lactate, C-reactive protein, and procalcitonin when available; imaging findings from plain abdominal radiographs, computed tomography scans with or without intravenous contrast, and abdominal ultrasonography; and the definitive treatment approach with clinical outcome.\u003c/p\u003e \u003cp\u003eTo ensure blinded evaluation, case scenarios presented to the AI systems were constructed by removing the definitive diagnosis and treatment outcome from the original case reports. Each scenario included only the clinical presentation, physical examination findings, laboratory values, and imaging descriptions as would be available to a clinician at the point of initial assessment. Radiological images were extracted in their original format (JPEG/PNG) from the published case reports when available; for cases where only textual descriptions of imaging findings were present, these descriptions were provided verbatim as input. The gold standard diagnosis and treatment were retained separately in a locked database for subsequent validation of AI-generated outputs. Data extraction was performed independently by two investigators (M.D. and E.E.), with discrepancies resolved by a third reviewer. The inter-rater reliability for data extraction demonstrated substantial agreement (κ\u0026thinsp;=\u0026thinsp;0.91, 95% CI: 0.87\u0026ndash;0.95).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Artificial Intelligence Systems and Architectural Specifications\u003c/h2\u003e \u003cp\u003eThree distinct AI systems were evaluated in this study, representing fundamentally different architectural approaches to clinical decision support: two general-purpose monolithic large language models (ChatGPT and Gemini) and a novel neurosymbolic multi-agent hybrid system integrating specialized medical AI components operating in a sequential pipeline with explicit symbolic verification.\u003c/p\u003e \u003cdiv id=\"Sec7\" class=\"Section3\"\u003e \u003ch2\u003e2.4.1 ChatGPT (GPT-4 Turbo)\u003c/h2\u003e \u003cp\u003eChatGPT, developed by OpenAI (San Francisco, CA, USA), is a general-purpose large language model based on the Generative Pre-trained Transformer architecture. The GPT-4 Turbo version (September 2025 update) was utilized in this study, accessed through the official web interface (chat.openai.com). The model was employed in a zero-shot configuration without any domain-specific fine-tuning or additional training on medical datasets. For each case evaluation, a new conversation session was initiated to prevent contextual carryover from previous cases. Default temperature settings (temperature\u0026thinsp;=\u0026thinsp;1.0) were maintained, and the maximum token limit was set to 4,096 tokens. The model received only textual input consisting of the standardized case scenario. The GPT-4 Turbo architecture employs a Mixture-of-Experts decoder-only transformer with an estimated 1.8 trillion parameters, pre-trained on a diverse internet corpus with a knowledge cutoff of April 2024, subsequently updated through reinforcement learning from human feedback (RLHF).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.4.2 Gemini 2.0 Pro\u003c/h2\u003e \u003cp\u003eGemini 2.0 Pro, developed by Google DeepMind (London, UK), represents a state-of-the-art multimodal foundation model with advanced reasoning capabilities, built on a novel architecture integrating cross-modal attention mechanisms for simultaneous text, image, audio, and video understanding. The model was accessed through the Google AI Studio application programming interface (API) with default parameter configurations (temperature\u0026thinsp;=\u0026thinsp;1.0, top-p\u0026thinsp;=\u0026thinsp;0.95, top-k\u0026thinsp;=\u0026thinsp;40). Although Gemini possesses native multimodal capabilities including image interpretation, only textual input was provided to ensure methodological consistency and fair comparison with ChatGPT in the primary analysis. Each case was assessed in an independent session to prevent contextual carryover. The Gemini 2.0 Pro architecture incorporates a unified multimodal encoder-decoder framework with an estimated parameter count exceeding 1 trillion, trained on a curated dataset encompassing text, code, images, audio, and video content.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.4.3 Neurosymbolic Multi-Agent Hybrid System\u003c/h2\u003e \u003cp\u003eThe neurosymbolic multi-agent system was designed to simulate a multidisciplinary clinical consultation by integrating three specialized AI components, each optimized for distinct aspects of the diagnostic workflow. The architectural design follows neurosymbolic principles by combining neural perception and reasoning modules with a symbolic verification layer that enforces logical consistency, medical guideline adherence, and factual grounding. The system was conceptualized as a sequential pipeline in which each agent\u0026rsquo;s output serves as a structured input to the subsequent agent, thereby creating a traceable chain of clinical reasoning analogous to the multidisciplinary team (MDT) consultation model used in modern surgical practice. The system architecture comprised the following three agents:\u003c/p\u003e \u003cp\u003e \u003cb\u003eRadiology Perception Agent (Hulu-Med 32B)\u003c/b\u003e: Hulu-Med is an open-source medical vision-language model (VLM) specifically designed for radiological image interpretation. The 32-billion parameter version was employed, which has been trained on a comprehensive dataset encompassing 12 anatomical systems and 14 imaging modalities including plain radiography, computed tomography, magnetic resonance imaging, and ultrasonography. This agent constitutes the neural perception layer of the neurosymbolic pipeline, processing radiological images when available in the original case reports (plain abdominal radiographs and computed tomography images) and generating structured reports identifying key findings. The agent was specifically configured to detect and report: bowel dilation patterns (small bowel\u0026thinsp;\u0026gt;\u0026thinsp;3 cm, large bowel\u0026thinsp;\u0026gt;\u0026thinsp;6 cm, cecum\u0026thinsp;\u0026gt;\u0026thinsp;9 cm); air-fluid levels with quantification; transition point identification and localization; pathognomonic signs including the \u0026ldquo;coffee-bean sign,\u0026rdquo; \u0026ldquo;omega loop sign,\u0026rdquo; \u0026ldquo;whirl sign,\u0026rdquo; \u0026ldquo;bird-beak sign,\u0026rdquo; and \u0026ldquo;northern exposure sign\u0026rdquo; characteristic of volvulus subtypes; mesenteric vessel engorgement; pneumatosis intestinalis; portal venous gas; free intraperitoneal air; and wall thickening patterns. For cases where only textual descriptions of imaging findings were available, these descriptions were provided as structured input.\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eClinical Reasoning Agent (Med-PaLM 2)\u003c/strong\u003e \u003cp\u003eMed-PaLM 2, developed by Google DeepMind, is a large language model specifically fine-tuned on medical knowledge bases and clinical reasoning tasks, representing the neural reasoning layer of the neurosymbolic pipeline. The model has demonstrated expert-level performance on medical licensing examinations, achieving 86.5% accuracy on the United States Medical Licensing Examination (USMLE) Step 1, 2, and 3 combined, and 72.3% on the MedMCQA benchmark. This agent integrated the structured radiological findings from the Hulu-Med output with the clinical data (history, physical examination, laboratory values) to formulate a ranked differential diagnosis with confidence scores, identify the most probable diagnosis with supporting evidence, generate evidence-based treatment recommendations aligned with current clinical practice guidelines, and provide pathophysiological rationale connecting clinical findings to diagnostic conclusions.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e\u003cb\u003eSymbolic Validation Agent (Gyan LLM)\u003c/b\u003e: Gyan is a compositional, explainable language model designed with a neurosymbolic architecture that explicitly separates the knowledge base from the inference engine, specifically targeting hallucination reduction through symbolic rule enforcement. This agent constitutes the symbolic verification layer of the pipeline\u0026mdash;the defining architectural component that distinguishes the neurosymbolic multi-agent approach from purely neural multi-model systems. The Gyan agent performed quality control by cross-referencing the diagnostic and therapeutic recommendations generated by Med-PaLM 2 against: (a) the original case data provided as input (factual grounding verification); (b) an internal medical knowledge graph encoding established clinical guidelines for ileus and volvulus management; and (c) explicit safety rules encoding contraindicated interventions and mandatory surgical indications. Any unsupported claims, factual inconsistencies, logically incoherent reasoning steps, or potentially unsafe recommendations (e.g., failure to recommend urgent surgical evaluation in the presence of strangulation signs) were flagged, annotated with justification, and corrected. The agent also assessed the logical coherence of the entire clinical reasoning chain, ensuring that the diagnostic conclusion followed from the presented evidence through valid inferential steps.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section3\"\u003e \u003ch2\u003e2.4.4 Rationale for Agent Selection and Architectural Design\u003c/h2\u003e \u003cp\u003eThe selection of the three constituent agents was guided by three interdependent design principles derived from the neurosymbolic AI literature and from the specific diagnostic requirements of ileus-spectrum and volvulus-spectrum pathology: (i) task-specific perceptual specialization, (ii) medically grounded clinical reasoning, and (iii) explicit symbolic verification with hallucination mitigation. Each agent was chosen to fulfill one of these roles based on its demonstrated domain performance, architectural suitability for the assigned subtask, and complementarity with the other pipeline components. The rationale for each selection is detailed below.\u003c/p\u003e \u003cp\u003e\u003cb\u003eRationale for Hulu-Med 32B (Radiology Perception Agent)\u003c/b\u003e: The accurate diagnosis of volvulus and mechanical obstruction depends critically on the identification of pathognomonic radiological signs\u0026mdash;such as the coffee-bean sign, whirl sign, bird-beak sign, and northern exposure sign\u0026mdash;that carry high positive predictive value when present but are frequently overlooked or misinterpreted by non-specialist readers. General-purpose LLMs process radiological descriptions as unstructured text tokens without any visual-semantic grounding; they can recognize the textual mention of a \u0026ldquo;coffee-bean sign\u0026rdquo; but cannot evaluate whether imaging features genuinely support that designation. This fundamental limitation necessitated a dedicated vision-language model (VLM) with domain-specific radiological training as the perceptual front-end of the pipeline. Hulu-Med 32B was selected over alternative medical VLMs (e.g., LLaVA-Med, BiomedCLIP, RadFM) for the following reasons: (a) it is currently the largest open-source medical VLM (32\u0026nbsp;billion parameters) with explicit training across 12 anatomical systems and 14 imaging modalities, providing the broadest abdominal imaging coverage; (b) it generates structured radiological reports rather than free-text descriptions, producing standardized output fields (bowel dilation measurements, transition point localization, pathognomonic sign identification, complication assessment) that can be directly consumed by the downstream reasoning agent in a machine-readable format; (c) independent benchmarking studies have demonstrated its superior performance in abdominal CT interpretation compared to smaller medical VLMs, particularly for identifying subtle findings such as mesenteric vessel engorgement and pneumatosis intestinalis that are critical for distinguishing simple obstruction from strangulation; and (d) as an open-source model, it allows full reproducibility and auditability of the perception stage, which is essential for a clinical decision support system operating in a high-stakes diagnostic domain.\u003c/p\u003e \u003cp\u003e\u003cb\u003eRationale for Med-PaLM 2 (Clinical Reasoning Agent)\u003c/b\u003e: The clinical reasoning stage required a model capable of integrating heterogeneous data streams\u0026mdash;structured radiological findings from Hulu-Med, unstructured clinical history, quantitative laboratory values, and physical examination descriptors\u0026mdash;into a coherent diagnostic synthesis with ranked differential diagnoses and evidence-based treatment recommendations. This task demands both broad medical knowledge and the capacity for multi-step clinical inference (e.g., recognizing that elevated lactate combined with a whirl sign on CT in the context of acute abdominal pain and hemodynamic instability constitutes a strangulated volvulus requiring emergent laparotomy rather than endoscopic decompression). Med-PaLM 2 was selected for this role based on its unique combination of characteristics: (a) it represents the current state-of-the-art in medically fine-tuned LLMs, having achieved 86.5% accuracy on the USMLE, 72.3% on MedMCQA, and expert-physician-level performance on clinical reasoning benchmarks; (b) unlike general-purpose models (GPT-4, Gemini), Med-PaLM 2 was specifically fine-tuned on curated medical question-answering datasets and clinical reasoning tasks, resulting in more clinically calibrated confidence assessments and fewer over-confident incorrect diagnoses; (c) its training explicitly incorporates differential diagnosis generation and treatment guideline adherence, which are the core outputs required from this pipeline stage; and (d) it has demonstrated particular strength in emergency medicine and surgical decision-making scenarios in prior evaluations, making it well-suited for the acute abdominal pathology domain of this study. The choice to use Med-PaLM 2 rather than the same general-purpose models being evaluated (GPT-4 or Gemini) as the reasoning agent was deliberate: employing a distinct, medically specialized model ensures that the multi-agent system\u0026rsquo;s advantage stems from architectural specialization rather than simply from using a different version of the same model.\u003c/p\u003e \u003cp\u003e \u003cb\u003eRationale for Gyan LLM (Symbolic Validation Agent)\u003c/b\u003e: The most distinctive architectural element of the neurosymbolic pipeline\u0026mdash;and the component that fundamentally differentiates it from a purely neural multi-model system\u0026mdash;is the symbolic verification layer. In standard multi-agent LLM frameworks, each agent remains a neural model subject to the same failure modes as monolithic LLMs: hallucination, logical inconsistency, and overconfident reasoning from insufficient evidence. The addition of a purely symbolic verification stage addresses these failure modes through a fundamentally different computational paradigm. Gyan LLM was selected for this critical role because of its unique compositional architecture that explicitly separates the knowledge base from the inference engine\u0026mdash;a design principle rooted in classical symbolic AI and knowledge representation theory. This architectural separation provides three capabilities that purely neural models cannot guarantee: (a) Factual grounding verification: Gyan maintains an explicit, queryable representation of the input case data and systematically checks every assertion in the clinical assessment against this representation. Unlike neural models that \u0026ldquo;recall\u0026rdquo; information probabilistically (and therefore can fabricate plausible-sounding details), Gyan performs deterministic lookup-based verification, flagging any claim that cannot be traced to a specific element of the input data. This mechanism directly targets the hallucination problem that represents the most dangerous failure mode of clinical AI. (b) Rule-based safety enforcement: The symbolic architecture allows encoding of explicit, non-negotiable clinical safety rules as formal logical constraints. In this study, four surgical safety rules (S1\u0026ndash;S4) were encoded: mandatory urgent intervention for cecal diameter exceeding 12 cm (S1), mandatory emergent surgical consultation for strangulation signs (S2), prohibition of conservative-only management for closed-loop obstruction (S3), and mandatory perforation risk assessment for toxic megacolon (S4). These rules function as hard constraints that override neural model outputs regardless of the reasoning agent\u0026rsquo;s confidence level\u0026mdash;a guarantee that probabilistic neural networks cannot provide. (c) Logical coherence auditing: Gyan evaluates the inferential chain connecting clinical evidence to diagnostic conclusions, identifying non-sequiturs, circular reasoning, and unsupported inferential leaps that characterize a substantial proportion of LLM diagnostic errors. No alternative model currently offers this combination of compositional knowledge representation, deterministic verification, and explicit rule enforcement. While retrieval-augmented generation (RAG) approaches can partially address factual grounding, they do not provide the deterministic safety rule enforcement or logical coherence auditing that the symbolic architecture enables. The selection of Gyan therefore reflects a principled neurosymbolic design decision: the neural components (Hulu-Med and Med-PaLM 2) provide the perceptual and reasoning capabilities that symbolic systems cannot match, while the symbolic component (Gyan) provides the verification guarantees that neural systems cannot offer.\u003c/p\u003e \u003cp\u003e\u003cb\u003eSynergistic Pipeline Design\u003c/b\u003e: The three-agent configuration was not arbitrary but reflects a deliberate mapping of the neurosymbolic paradigm onto the clinical diagnostic workflow. In clinical practice, the diagnostic process for acute abdominal emergencies follows a natural three-phase cognitive architecture: (1) perceptual analysis of imaging and physical findings, typically performed by a radiologist; (2) clinical synthesis integrating all available data into a diagnostic formulation and treatment plan, typically performed by the managing clinician; and (3) quality assurance and safety verification, typically performed through multidisciplinary team review or institutional safety protocols. Our multi-agent pipeline mirrors this clinical cognitive architecture: Hulu-Med replicates the radiologist\u0026rsquo;s perceptual expertise, Med-PaLM 2 replicates the clinician\u0026rsquo;s integrative reasoning, and Gyan replicates the institutional safety and quality verification layer. This biomimetic design philosophy ensures that the system\u0026rsquo;s outputs are not only more accurate but also more interpretable to clinicians, as each pipeline stage produces outputs analogous to familiar clinical documents (radiology report, clinical assessment, quality review). The sequential rather than parallel configuration was chosen because each stage\u0026rsquo;s output provides essential structured input for the subsequent stage: Med-PaLM 2 cannot generate an appropriate differential without the radiological findings from Hulu-Med, and Gyan cannot verify factual grounding without access to both the original input data and the intermediate outputs from prior stages.\u003c/p\u003e \u003cp\u003e The neurosymbolic multi-agent workflow operated as follows: (1) case data including clinical text and radiological images (when available) were input into the system; (2) the Hulu-Med radiology agent analyzed available images and/or imaging descriptions and generated a structured radiological report with identified findings and their confidence scores; (3) Med-PaLM 2 synthesized the structured radiological findings with clinical, laboratory, and physical examination data to produce ranked differential diagnoses and evidence-based therapeutic recommendations with supporting rationale; (4) the Gyan symbolic validation agent verified the entire output for factual accuracy against input data, logical consistency of the reasoning chain, safety compliance with surgical indication rules, and hallucination detection through compositional verification; (5) the final validated, corrected output was generated with a traceable audit trail documenting each verification step. The entire pipeline was automated with inter-agent communication via structured JSON schemas, with a mean processing time of 47.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.6 seconds per case. The system architecture is illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Prompt Design, Engineering, and Standardization\u003c/h2\u003e \u003cp\u003eA standardized prompt template was developed following established principles of prompt engineering for medical applications, emphasizing role specification, structured output requirements, clinical safety constraints, and evidence-based reasoning expectations. The prompt template was designed to maximize diagnostic reasoning while minimizing response variability across AI systems. The prompt was iteratively refined through a pilot phase involving 15 cases not included in the final analysis.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003e2.5.1 System-Level Prompt (Role Specification)\u003c/h2\u003e \u003cp\u003eThe following system-level prompt was applied uniformly across all three AI systems to establish the clinical decision support context:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e\u003cem\u003e\"SYSTEM: You are an expert clinical decision support system specializing in acute abdominal emergencies. You function as a consultant to the emergency department and surgical teams. Your role is to analyze clinical case presentations and provide structured diagnostic and management recommendations grounded in current evidence-based guidelines. You must reason transparently, cite specific clinical findings that support each diagnostic consideration, and explicitly address potential surgical emergencies that require urgent intervention. Safety is paramount: failure to identify conditions requiring emergent surgery (e.g., strangulated volvulus, closed-loop obstruction, toxic megacolon with perforation risk) is the most critical error to avoid.\"\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section3\"\u003e \u003ch2\u003e2.5.2 Case Presentation Prompt (User-Level)\u003c/h2\u003e \u003cp\u003eEach case was presented using the following standardized template, with patient-specific data inserted into designated fields:\u003cdiv class=\"BlockQuote\"\u003e\u003cp\u003e \u003cem\u003e\"CLINICAL CASE: A [AGE]-year-old [SEX] patient presents to the emergency department with [CHIEF COMPLAINT] of [DURATION] duration. PAST MEDICAL HISTORY: [COMORBIDITIES] VITAL SIGNS: Blood pressure [BP] mmHg, Heart rate [HR] bpm, Temperature [TEMP] \u0026deg;C, Respiratory rate [RR] breaths/min, SpO2 [SPO2]%. PHYSICAL EXAMINATION: [FINDINGS including abdominal examination details] LABORATORY RESULTS: [COMPLETE LAB VALUES] IMAGING FINDINGS: [RADIOLOGICAL DESCRIPTIONS AND/OR IMAGES] Based on the above clinical information, please provide: 1. PRIMARY DIAGNOSIS: State your most likely diagnosis with a confidence level (High/Moderate/Low) and the specific clinical and radiological findings that support this diagnosis. 2. DIFFERENTIAL DIAGNOSES: List 2 \u0026ndash; 4 alternative diagnoses in order of likelihood. For each, explain which findings support or argue against it. 3. CRITICAL ASSESSMENT: Explicitly state whether this case represents a surgical emergency requiring urgent operative intervention. Identify any red flags for strangulation, perforation, or ischemia. 4. MANAGEMENT PLAN: Provide a step-by-step management recommendation including: (a) immediate resuscitation measures, (b) definitive treatment (surgical, endoscopic, or conservative), (c) timing of intervention (emergent, urgent, or elective), and (d) monitoring parameters. 5. CLINICAL REASONING: Explain the pathophysiological basis connecting the key findings to your primary diagnosis. Describe why alternative diagnoses are less likely.\"\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section3\"\u003e \u003ch2\u003e2.5.3 Multi-Agent System Prompts\u003c/h2\u003e \u003cp\u003eFor the neurosymbolic multi-agent system, additional agent-specific prompts were designed for each component of the pipeline:\u003c/p\u003e \u003cp\u003e \u003cb\u003eHulu-Med Radiology Agent Prompt\u003c/b\u003e: \"Analyze the provided abdominal imaging (plain radiograph and/or CT scan) for a patient presenting with acute abdominal symptoms suggesting intestinal obstruction. Generate a STRUCTURED RADIOLOGICAL REPORT with the following mandatory sections: (A) BOWEL GAS PATTERN: Describe distribution, dilation (measure in cm where possible), and presence of air-fluid levels. (B) TRANSITION POINT: Identify if present, location, and character. (C) PATHOGNOMONIC SIGNS: Specifically assess for: coffee-bean sign, omega loop sign (sigmoid volvulus); comma/kidney-bean sign (cecal volvulus); whirl sign, bird-beak sign (any volvulus); small bowel feces sign (SBO). (D) COMPLICATIONS: Assess for pneumatosis intestinalis, portal venous gas, free intraperitoneal air, mesenteric vessel engorgement, bowel wall thickening/enhancement abnormalities. (E) ADDITIONAL FINDINGS: Any other relevant abdominal findings. (F) CONFIDENCE SCORE: Rate your overall confidence in the radiological interpretation (0.0\u0026ndash;1.0).\"\u003c/p\u003e \u003cp\u003e\u003cb\u003eMed-PaLM 2 Clinical Reasoning Agent Prompt\u003c/b\u003e: \"RADIOLOGICAL FINDINGS: [OUTPUT FROM HULU-MED AGENT] CLINICAL DATA: [CASE VIGNETTE TEXT] You are an expert internal medicine and surgical consultant. Integrate the structured radiological findings above with the clinical presentation, laboratory data, and patient history to generate: (1) A RANKED DIFFERENTIAL DIAGNOSIS (top 5) with confidence scores (0.0\u0026ndash;1.0) for each. For each diagnosis, cite the specific clinical, laboratory, and radiological findings that support or refute it. (2) PRIMARY DIAGNOSIS with detailed pathophysiological reasoning connecting findings to diagnosis. (3) SURGICAL URGENCY ASSESSMENT: Classify as EMERGENT (\u0026lt;\u0026thinsp;2h), URGENT (2\u0026ndash;24h), SEMI-URGENT (24\u0026ndash;72h), or ELECTIVE. Justify with specific clinical criteria. (4) EVIDENCE-BASED MANAGEMENT PLAN with specific interventions, medications (with doses where applicable), and monitoring parameters. Cite relevant clinical guidelines (e.g., ASCRS, ESCP, Tokyo Guidelines) where applicable.\"\u003c/p\u003e \u003cp\u003e\u003cb\u003eGyan Symbolic Validation Agent Prompt\u003c/b\u003e: \"ORIGINAL CASE DATA: [CASE VIGNETTE] RADIOLOGY REPORT: [HULU-MED OUTPUT] CLINICAL ASSESSMENT: [MED-PALM 2 OUTPUT] Perform SYSTEMATIC VALIDATION of the clinical assessment against the original case data and established medical guidelines. Execute the following verification steps: (1) FACTUAL GROUNDING CHECK: For every clinical assertion in the assessment, verify it is either (a) directly stated in the case data, or (b) a logically valid inference from stated data. Flag any assertion that is UNSUPPORTED, FABRICATED, or CONTRADICTED by the input data. (2) LOGICAL CONSISTENCY CHECK: Verify that the diagnostic reasoning chain is internally consistent\u0026mdash;each inferential step follows logically from established premises. Flag any non-sequiturs or circular reasoning. (3) SAFETY RULE VERIFICATION: Apply the following mandatory safety rules: [Rule S1] If cecal diameter\u0026thinsp;\u0026gt;\u0026thinsp;12 cm, recommendation MUST include urgent decompression or surgery. [Rule S2] If signs of strangulation (ischemia markers, peritonitis, hemodynamic instability), recommendation MUST include emergent surgical consultation. [Rule S3] If closed-loop obstruction is suspected, recommendation MUST NOT advise conservative management alone. [Rule S4] If toxic megacolon criteria are met, recommendation MUST address perforation risk. (4) GUIDELINE ALIGNMENT: Verify treatment recommendations align with current ASCRS/ESCP guidelines for the identified condition. (5) OUTPUT: Generate a VALIDATION REPORT listing: all verified claims, all flagged issues (with severity: CRITICAL/MODERATE/MINOR), all corrections applied, and the final VALIDATED clinical assessment.\"\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Outcome Measures and Evaluation Criteria\u003c/h2\u003e \u003cp\u003eEach AI system output was independently evaluated by two blinded assessors: a board-certified radiologist with 12 years of experience in abdominal imaging (Assessor A) and a board-certified general surgeon with 15 years of experience in emergency abdominal surgery (Assessor B). The assessors were blinded to the AI system identity and evaluated outputs in randomized order using a standardized electronic scoring form. Five predefined evaluation criteria were applied, each scored as a binary outcome (correct/incorrect or present/absent):\u003c/p\u003e \u003cp\u003e \u003cstrong\u003eDiagnostic Accuracy (Primary Outcome)\u003c/strong\u003e \u003cp\u003eThe primary diagnosis provided by the AI system was compared against the gold standard diagnosis from the original case report. A diagnosis was scored as correct only if it matched the specific final diagnosis (e.g., \u0026ldquo;sigmoid volvulus\u0026rdquo; not merely \u0026ldquo;colonic obstruction\u0026rdquo;). Partially correct or nonspecific diagnoses were scored as incorrect. For conditions with multiple accepted diagnostic terms (e.g., \u0026ldquo;Ogilvie syndrome\u0026rdquo; and \u0026ldquo;acute colonic pseudo-obstruction\u0026rdquo;), either term was accepted.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eTreatment Appropriateness\u003c/strong\u003e \u003cp\u003e The recommended management strategy was evaluated against the treatment actually administered in the case report and current clinical practice guidelines (ASCRS 2021, ESCP 2020, WSES 2023). Treatment was scored as appropriate if all critical management components were included. Failure to recommend urgent surgical evaluation in cases requiring emergency surgery was scored as inappropriate regardless of other management suggestions.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cstrong\u003eHallucination\u003c/strong\u003e \u003cp\u003eAny clinical assertion in the AI output not present in or directly inferable from the provided case data was classified as a hallucination. This included fabricated symptoms, invented laboratory values, claims of imaging findings not described, or attribution of medical history not provided. Minor embellishments that did not affect clinical reasoning (e.g., reasonable assumptions about standard care) were not counted.\u003c/p\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eExplanation Adequacy\u003c/b\u003e: Clinical reasoning quality was assessed based on four sub-criteria: logical connection between findings and diagnosis; reference to key supportive evidence; discussion of relevant differential diagnoses; and pathophysiological rationale for diagnosis and treatment. Outputs meeting all four sub-criteria were scored as adequate.\u003c/p\u003e \u003cp\u003e \u003cb\u003eCritical Safety Error\u003c/b\u003e: This criterion identified recommendations that could result in significant patient harm if followed, including: failure to recommend urgent surgical evaluation in established surgical indications; recommendation of contraindicated interventions; suggestion of outpatient management for conditions requiring hospitalization; and any advice leading to delayed treatment of a surgical emergency.\u003c/p\u003e \u003cp\u003eInter-rater reliability between the two assessors was calculated using Cohen\u0026rsquo;s kappa coefficient for each evaluation criterion. Disagreements were resolved through discussion and, when necessary, adjudication by a third assessor (a gastroenterologist with 18 years of clinical experience).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Statistical Analysis\u003c/h2\u003e \u003cp\u003eStatistical analyses were performed using Python version 3.11 (Python Software Foundation, Wilmington, DE, USA) with the following packages: NumPy (v1.26) for numerical computations, SciPy (v1.12) for statistical testing, Pandas (v2.1) for data manipulation, Statsmodels (v0.14) for advanced statistical modeling, and Matplotlib (v3.8) with Seaborn (v0.13) for data visualization. Categorical variables were summarized as frequencies and percentages. Continuous variables were expressed as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation or median with interquartile range (IQR) as appropriate based on distribution normality assessed by the Shapiro\u0026ndash;Wilk test.\u003c/p\u003e \u003cp\u003ePairwise comparisons of performance metrics between AI systems were conducted using the McNemar test for paired nominal data, which is appropriate for matched-pair designs where each case serves as its own control across different AI systems. The Bonferroni correction was applied to adjust for multiple comparisons across three pairwise comparisons, with the corrected significance threshold set at p\u0026thinsp;\u0026lt;\u0026thinsp;0.017 (0.05/3). Exact 95% confidence intervals for proportions were calculated using the Wilson score method. Inter-rater reliability was quantified using Cohen\u0026rsquo;s kappa coefficient, with values interpreted according to Landis and Koch criteria: \u0026lt; 0.20, 0.21\u0026ndash;0.40 (fair), 0.41\u0026ndash;0.60 (moderate), 0.61\u0026ndash;0.80 (substantial), and \u0026gt;\u0026thinsp;0.80 (excellent/almost perfect). Subgroup analyses were performed using Fisher\u0026rsquo;s exact test due to small cell sizes in some diagnostic categories. Effect sizes for pairwise comparisons were quantified using the absolute risk difference with 95% confidence intervals. All tests were two-tailed, and a p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant unless otherwise specified. A post hoc power analysis was conducted to confirm adequate statistical power (\u0026ge;\u0026thinsp;0.80) for the observed effect sizes.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. RESULTS","content":"\u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Study Population and Case Characteristics\u003c/h2\u003e \u003cp\u003eA total of 133 case reports meeting the predefined eligibility criteria were included in the final analysis following the systematic literature search and multi-stage screening process. The initial PubMed/MEDLINE database search yielded 847 potentially relevant records. After removal of duplicates and title/abstract screening by two independent reviewers (M.T.U. and E.E.), 412 articles were deemed potentially eligible and underwent full-text review. Application of the predefined inclusion and exclusion criteria resulted in the exclusion of 279 articles, with 133 cases ultimately retained for the final analysis. The complete case selection process, including the number of records identified, screened, assessed for eligibility, excluded with reasons at each stage, and ultimately included, is illustrated in the PRISMA 2020 flow diagram presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Inter-reviewer agreement for study selection was excellent, with a Cohen's kappa coefficient of \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.94 (95% CI: 0.91\u0026ndash;0.97). All disagreements (n\u0026thinsp;=\u0026thinsp;14) were resolved through consensus discussion with the third reviewer.The demographic and clinical characteristics of the study population are summarized in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDemographic and Clinical Characteristics of the Study Population (N\u0026thinsp;=\u0026thinsp;133)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003e Characteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eValue\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAge, years, median (IQR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e62 (48\u0026ndash;73)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMale sex, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e77 (57.9)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSymptom duration, hours, median (IQR)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e36 (18\u0026ndash;72)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003ePresenting Symptoms, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbdominal pain\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e119 (89.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbdominal distension\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e110 (82.7)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNausea/Vomiting\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e87 (65.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eObstipation/Inability to pass flatus\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e76 (57.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFever (\u0026gt;\u0026thinsp;38\u0026deg;C)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28 (21.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eComorbidities, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHypertension\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e51 (38.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePrior abdominal surgery\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e42 (31.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiabetes mellitus\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e34 (25.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChronic constipation\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e29 (21.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNeurological disorder\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18 (13.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDiagnostic Categories, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMechanical obstruction (total)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e55 (41.4)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVolvulus (sigmoid/cecal/small bowel)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e28 (21.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAdhesive/Bridle obstruction\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e18 (13.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther mechanical causes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9 (6.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOgilvie syndrome (ACPO)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e37 (27.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eParalytic ileus\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e20 (15.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eToxic megacolon/Severe colonic distension\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12 (9.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther rare etiologies\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e9 (6.8)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTreatment Approach, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSurgical intervention\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e58 (43.6)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEndoscopic decompression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e31 (23.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConservative/Medical management\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e44 (33.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eImaging Availability, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eComputed tomography\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e119 (89.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePlain abdominal radiography\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e96 (72.2)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAbdominal ultrasonography\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e32 (24.1)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"2\"\u003e\u003cem\u003eIQR, interquartile range; ACPO, acute colonic pseudo-obstruction.\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe most frequently reported presenting complaint was abdominal pain, documented in 119 of 133 cases (89.5%, 95% CI: 83.0\u0026ndash;94.2%), followed by abdominal distension in 110 cases (82.7%, 95% CI: 75.2\u0026ndash;88.8%), nausea and/or vomiting in 87 cases (65.4%, 95% CI: 56.7\u0026ndash;73.4%), obstipation or inability to pass flatus in 76 cases (57.1%, 95% CI: 48.3\u0026ndash;65.7%), and fever (temperature\u0026thinsp;\u0026gt;\u0026thinsp;38\u0026deg;C) in 28 cases (21.1%, 95% CI: 14.5\u0026ndash;28.9%). Co-occurrence of abdominal pain with distension was observed in 104 cases (78.2%), while the classic triad of pain, distension, and vomiting was present in 72 cases (54.1%). Constipation with complete absence of flatus passage was documented in 49 cases (36.8%), suggesting complete bowel obstruction.\u003c/p\u003e \u003cp\u003eRegarding underlying comorbid conditions, hypertension was the most prevalent, present in 51 patients (38.3%, 95% CI: 30.1\u0026ndash;47.1%), followed by prior abdominal surgery in 42 patients (31.6%, 95% CI: 23.7\u0026ndash;40.2%), diabetes mellitus in 34 patients (25.6%, 95% CI: 18.4\u0026ndash;33.8%), chronic constipation in 29 patients (21.8%, 95% CI: 15.1\u0026ndash;29.8%), and neurological disorders including Parkinson's disease, dementia, and cerebrovascular disease in 18 patients (13.5%, 95% CI: 8.2\u0026ndash;20.5%). Additional comorbidities included chronic kidney disease in 14 patients (10.5%), cardiovascular disease in 22 patients (16.5%), and chronic obstructive pulmonary disease in 8 patients (6.0%). A total of 67 patients (50.4%) had two or more comorbid conditions, and 28 patients (21.1%) had three or more, reflecting the complex clinical profiles characteristic of patients presenting with acute intestinal obstruction.\u003c/p\u003e \u003cp\u003eComputed tomography (CT) imaging was available in 119 of 133 cases (89.5%, 95% CI: 83.0\u0026ndash;94.2%), constituting the primary diagnostic imaging modality. Plain abdominal radiography was available in 96 cases (72.2%, 95% CI: 63.7\u0026ndash;79.6%), and abdominal ultrasonography in 32 cases (24.1%, 95% CI: 17.1\u0026ndash;32.2%). Among the 119 cases with CT imaging, contrast-enhanced CT was performed in 94 cases (79.0%) and non-contrast CT in 25 cases (21.0%). Dual-modality imaging (both CT and plain radiography) was available in 88 cases (66.2%), while 31 cases (23.3%) had CT only, 8 cases (6.0%) had plain radiography only, and 6 cases (4.5%) had radiological findings described textually without available image files. Actual radiological images (as JPEG/PNG files extracted from the published case reports) were available for AI analysis in 97 cases (72.9%), while the remaining 36 cases (27.1%) relied on textual descriptions of imaging findings provided verbatim from the original reports.\u003c/p\u003e \u003cp\u003eThe final diagnostic distribution of the 133 included cases comprised five major categories and one miscellaneous category: mechanical obstruction in 55 cases (41.4%, 95% CI: 33.0\u0026ndash;50.1%), including 28 volvulus cases (21.1%; consisting of 19 sigmoid volvulus [14.3%], 6 cecal volvulus [4.5%], and 3 other volvulus subtypes [2.3% \u0026mdash; 2 transverse colon volvulus and 1 small bowel volvulus]), 18 adhesive or bridle obstruction cases (13.5%), and 9 other mechanical causes (6.8% \u0026mdash; 3 internal hernias, 2 intussusception, 2 gallstone ileus, 1 Meckel's diverticulum band, 1 bezoar); Ogilvie syndrome or acute colonic pseudo-obstruction (ACPO) in 37 cases (27.8%, 95% CI: 20.4\u0026ndash;36.2%); paralytic ileus in 20 cases (15.0%, 95% CI: 9.5\u0026ndash;22.2%); toxic megacolon or severe colonic distension in 12 cases (9.0%, 95% CI: 4.7\u0026ndash;15.2%); and other rare etiologies in 9 cases (6.8%, 95% CI: 3.1\u0026ndash;12.5%). Urgent surgical intervention was ultimately required in 43 of 133 cases (32.3%, 95% CI: 24.5\u0026ndash;41.0%), while 58 cases (43.6%) were managed conservatively, 24 cases (18.0%) underwent endoscopic intervention, and 8 cases (6.0%) required both initial endoscopic and subsequent surgical management.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e3.2 AI System Performance Comparison\u003c/h2\u003e \u003cp\u003eThe comparative performance of the three AI systems\u0026mdash;ChatGPT (GPT-4 Turbo), Gemini 2.0 Pro, and the neurosymbolic multi-agent system\u0026mdash;across all five evaluation criteria is presented in Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e. Each system evaluated all 133 cases, yielding a total of 399 AI-generated clinical assessments (133 per system) and 1,995 individual evaluation data points (399 assessments \u0026times; 5 criteria). Inter-rater reliability between the two independent assessors was excellent overall, with a weighted Cohen's kappa coefficient of \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.89 (95% CI: 0.84\u0026ndash;0.94). Domain-specific kappa values were: diagnostic accuracy \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.95 (95% CI: 0.91\u0026ndash;0.99), critical safety errors \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.93 (95% CI: 0.86\u0026ndash;1.00), hallucination detection \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.91 (95% CI: 0.85\u0026ndash;0.97), treatment appropriateness \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.88 (95% CI: 0.82\u0026ndash;0.94), and explanation adequacy \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.82 (95% CI: 0.75\u0026ndash;0.89). All five domains achieved kappa values exceeding 0.80, meeting the threshold for excellent agreement according to the Landis and Koch classification. Disagreements occurred in 47 of 1,995 individual evaluations (2.4%), distributed as follows: diagnostic accuracy 4/399 (1.0%), treatment appropriateness 12/399 (3.0%), hallucination detection 7/399 (1.8%), explanation adequacy 17/399 (4.3%), and critical safety errors 7/399 (1.8%). All disagreements were resolved through consensus discussion with the senior investigator (H.Y.B.), and final adjudicated ratings were used for all analyses. Post hoc power analysis confirmed adequate statistical power (1\u0026thinsp;\u0026minus;\u0026thinsp;β\u0026thinsp;\u0026gt;\u0026thinsp;0.90) for detecting a 15% absolute difference in diagnostic accuracy between systems at \u003cem\u003eα\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.05 with the observed sample size of n\u0026thinsp;=\u0026thinsp;133 paired observations.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eComparative Performance of AI Systems Across Evaluation Criteria (N\u0026thinsp;=\u0026thinsp;133)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEvaluation Criterion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eChatGPT (GPT-4 Turbo)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGemini 2.0 Pro\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMulti-Agent System\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCorrect Diagnosis, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e80 (60.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e78 (58.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e100 (75.2)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e51.4\u0026ndash;68.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e49.8\u0026ndash;67.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e66.9\u0026ndash;82.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAppropriate Treatment, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e85 (63.9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e82 (61.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e99 (74.4)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e55.2\u0026ndash;72.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e52.9\u0026ndash;69.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e66.1\u0026ndash;81.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eHallucination Present, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e20 (15.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e13 (9.8)\u0026Dagger;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2 (1.5)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9.4\u0026ndash;22.3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5.3\u0026ndash;16.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.2\u0026ndash;5.3\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAdequate Explanation, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e92 (69.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e78 (58.6)\u0026Dagger;\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e107 (80.5)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e60.6\u0026ndash;76.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e49.8\u0026ndash;67.0\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e72.7\u0026ndash;86.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eCritical Safety Error, n (%)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5 (3.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3 (2.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0 (0.0)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1.2\u0026ndash;8.6\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.5\u0026ndash;6.5\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.0\u0026ndash;2.7\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003e\u003cem\u003eCI, confidence interval. *p\u0026thinsp;\u0026lt;\u0026thinsp;0.017 vs ChatGPT; \u0026dagger;p\u0026thinsp;\u0026lt;\u0026thinsp;0.017 vs Gemini (McNemar test with Bonferroni correction); \u0026Dagger;p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 vs ChatGPT.\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cdiv id=\"Sec20\" class=\"Section3\"\u003e \u003ch2\u003e3.2.1 Diagnostic Accuracy\u003c/h2\u003e \u003cp\u003eThe neurosymbolic multi-agent system achieved the highest diagnostic accuracy among the three evaluated systems, correctly identifying the primary diagnosis in 100 of 133 cases (75.2%, 95% CI: 66.9\u0026ndash;82.2%; Wilson score method). ChatGPT (GPT-4 Turbo) demonstrated correct diagnosis in 80 of 133 cases (60.2%, 95% CI: 51.4\u0026ndash;68.5%), while Gemini 2.0 Pro achieved correct diagnosis in 78 of 133 cases (58.6%, 95% CI: 49.8\u0026ndash;67.0%).\u003c/p\u003e \u003cp\u003ePairwise comparisons using the McNemar test with Bonferroni correction for three comparisons (adjusted significance threshold: \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.017) revealed that the multi-agent system demonstrated statistically significantly superior diagnostic accuracy compared to both ChatGPT (absolute risk difference [ARD]: 15.0%, 95% CI of difference: 5.8\u0026ndash;24.2%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 14.22; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and Gemini (ARD: 16.5%, 95% CI: 7.2\u0026ndash;25.9%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 16.89; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). No statistically significant difference was observed between ChatGPT and Gemini (ARD: 1.5%, 95% CI: \u0026minus;8.4 to 11.5%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 0.13; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.72), indicating that both general-purpose LLMs exhibited comparable diagnostic performance.\u003c/p\u003e \u003cp\u003eThe relative risk (RR) of an accurate diagnosis was 1.25 (95% CI: 1.08\u0026ndash;1.45) for the multi-agent system in comparison to ChatGPT, and 1.28 (95% CI: 1.10\u0026ndash;1.49) in relation to Gemini. The number needed to treat (NNT)\u0026mdash;which indicates the number of cases the multi-agent system must evaluate to result in one additional accurate diagnosis relative to the monolithic LLM\u0026mdash;was 6.7 (95% CI: 4.1\u0026ndash;17.2) when compared to ChatGPT, and 6.1 (95% CI: 3.9\u0026ndash;13.9) when compared to Gemini.\u003c/p\u003e \u003cp\u003eConcordance analysis revealed that all three systems agreed on the correct diagnosis in 64 of 133 cases (48.1%), while all three agreed on an incorrect diagnosis in 9 cases (6.8%). The multi-agent system was uniquely correct (correct when both monolithic LLMs were incorrect) in 24 cases (18.0%), ChatGPT was uniquely correct in 4 cases (3.0%), and Gemini was uniquely correct in 3 cases (2.3%). Cases where the multi-agent system was uniquely correct predominantly involved volvulus with pathognomonic imaging signs (n\u0026thinsp;=\u0026thinsp;8) and Ogilvie syndrome with subtle CT features (n\u0026thinsp;=\u0026thinsp;9). Among the 33 cases misdiagnosed by the multi-agent system, the most common error patterns involved misclassification of Ogilvie syndrome as paralytic ileus (n\u0026thinsp;=\u0026thinsp;12, 36.4% of multi-agent errors), failure to identify toxic megacolon (n\u0026thinsp;=\u0026thinsp;6, 18.2%), misdiagnosis of rare etiologies (n\u0026thinsp;=\u0026thinsp;6, 18.2%), and incorrect etiology attribution in mechanical obstruction (n\u0026thinsp;=\u0026thinsp;5, 15.2%). Among the 53 ChatGPT misdiagnoses, error patterns included Ogilvie syndrome misclassification as mechanical obstruction (n\u0026thinsp;=\u0026thinsp;19, 35.8%), volvulus misidentification (n\u0026thinsp;=\u0026thinsp;6, 11.3%), and non-specific etiological attribution (n\u0026thinsp;=\u0026thinsp;10, 18.9%). Gemini misdiagnosed 55 cases, with similar patterns including Ogilvie syndrome misclassification (n\u0026thinsp;=\u0026thinsp;18, 32.7%), volvulus misidentification (n\u0026thinsp;=\u0026thinsp;7, 12.7%), and adhesive obstruction over-diagnosis (n\u0026thinsp;=\u0026thinsp;12, 21.8%).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section3\"\u003e \u003ch2\u003e3.2.2 Treatment Appropriateness\u003c/h2\u003e \u003cp\u003e Appropriate treatment recommendations, defined as concordance with current evidence-based guidelines (ASCRS 2021, ESCP 2020, WSES 2023) for the confirmed diagnosis, were provided by the neurosymbolic multi-agent system in 99 of 133 cases (74.4%, 95% CI: 66.1\u0026ndash;81.5%), by ChatGPT in 85 of 133 cases (63.9%, 95% CI: 55.2\u0026ndash;72.0%), and by Gemini in 82 of 133 cases (61.7%, 95% CI: 52.9\u0026ndash;69.9%). The multi-agent system demonstrated significantly superior treatment appropriateness compared to both ChatGPT (ARD: 10.5%, 95% CI: 1.3\u0026ndash;19.8%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 8.64; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.003) and Gemini (ARD: 12.8%, 95% CI: 3.4\u0026ndash;22.1%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 11.52; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). The difference between ChatGPT and Gemini was not statistically significant (ARD: 2.3%, 95% CI: \u0026minus;7.8 to 12.3%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 0.38; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.54). The NNT for appropriate treatment was 9.5 (95% CI: 5.1\u0026ndash;76.9) for the multi-agent versus ChatGPT and 7.8 (95% CI: 4.5\u0026ndash;29.4) versus Gemini.\u003c/p\u003e \u003cp\u003eAmong the 43 cases requiring urgent surgical intervention\u0026mdash;comprising volvulus with strangulation risk (n\u0026thinsp;=\u0026thinsp;18), closed-loop obstruction (n\u0026thinsp;=\u0026thinsp;11), and toxic megacolon with peritoneal signs or critical cecal diameter (n\u0026thinsp;=\u0026thinsp;14)\u0026mdash;the multi-agent system correctly recommended surgery or urgent surgical consultation in 42 of 43 cases (97.7%, 95% CI: 87.7\u0026ndash;99.9%), ChatGPT in 37 of 43 cases (86.0%, 95% CI: 72.1\u0026ndash;94.7%), and Gemini in 36 of 43 cases (83.7%, 95% CI: 69.3\u0026ndash;93.2%). The difference between the multi-agent system and ChatGPT approached statistical significance (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.063), while the difference against Gemini reached significance (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.031). The sensitivity for surgical emergency recognition was 97.7% (95% CI: 87.7\u0026ndash;99.9%) for the multi-agent system, 86.0% (95% CI: 72.1\u0026ndash;94.7%) for ChatGPT, and 83.7% (95% CI: 69.3\u0026ndash;93.2%) for Gemini. The multi-agent system's single missed surgical recommendation involved a case of rare internal hernia where the system correctly identified bowel obstruction but underestimated presentation acuity.\u003c/p\u003e \u003cp\u003eAn important observation was the discordance between diagnostic accuracy and treatment appropriateness across all systems. Among the 100 correctly diagnosed cases by the multi-agent system, treatment was appropriate in 89 cases (89.0%) and inappropriate in 11 cases (11.0%), indicating that correct diagnosis does not invariably lead to appropriate treatment. Conversely, among the 33 incorrectly diagnosed cases, treatment was nonetheless appropriate in 10 cases (30.3%), reflecting instances where the recommended management pathway was coincidentally correct despite diagnostic error. For ChatGPT, among 80 correctly diagnosed cases, 72 (90.0%) received appropriate treatment, while among 53 incorrectly diagnosed cases, 13 (24.5%) received coincidentally appropriate treatment. Seven of these ChatGPT cases involved outputs with incorrect diagnoses that nonetheless recommended surgical consultation as a general precaution, reflecting a conservative approach to diagnostic uncertainty that incidentally resulted in appropriate management. For Gemini, among 78 correctly diagnosed cases, 68 (87.2%) received appropriate treatment, and among 55 incorrectly diagnosed cases, 14 (25.5%) received coincidentally appropriate treatment. Notably, four Gemini cases with correct diagnoses received suboptimal treatment recommendations due to failure to recognize clinical urgency: two cases of sigmoid volvulus where endoscopic decompression was recommended despite imaging features suggestive of impending strangulation, and two cases of Ogilvie syndrome with critical cecal diameter (\u0026gt;\u0026thinsp;12 cm) where only pharmacological management was suggested without acknowledging perforation risk.\u003c/p\u003e \u003cp\u003eAmong the 90 cases managed conservatively or with endoscopic intervention, the multi-agent system provided appropriate recommendations in 57 of 90 cases (63.3%, 95% CI: 52.5\u0026ndash;73.2%), compared to 48 of 90 (53.3%, 95% CI: 42.5\u0026ndash;63.9%) for ChatGPT and 46 of 90 (51.1%, 95% CI: 40.4\u0026ndash;61.7%) for Gemini. The lower accuracy in conservatively managed cases, relative to surgical cases, reflects the greater diagnostic complexity inherent in non-surgical intestinal obstruction etiologies, particularly the distinction between Ogilvie syndrome and early mechanical obstruction. Among the 24 cases managed with endoscopic intervention (predominantly endoscopic decompression for sigmoid volvulus without strangulation), the multi-agent system recommended the correct endoscopic approach in 20 cases (83.3%), ChatGPT in 16 cases (66.7%), and Gemini in 15 cases (62.5%).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section3\"\u003e \u003ch2\u003e3.2.3 Hallucination Analysis\u003c/h2\u003e \u003cp\u003eHallucinations\u0026mdash;defined as AI-generated statements containing fabricated, distorted, or unsubstantiated clinical information not present in or directly inferable from the original case data\u0026mdash;were detected in 20 ChatGPT outputs (15.0%, 95% CI: 9.4\u0026ndash;22.3%), 13 Gemini outputs (9.8%, 95% CI: 5.3\u0026ndash;16.2%), and only 2 multi-agent system outputs (1.5%, 95% CI: 0.2\u0026ndash;5.3%). The hallucination profile across all three systems is illustrated in the heatmap presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The hallucination-free rate was 98.5% (131/133) for the multi-agent system, 90.2% (120/133) for Gemini, and 85.0% (113/133) for ChatGPT. The multi-agent system demonstrated significantly lower hallucination rates compared to both ChatGPT (ARD: 13.5%, 95% CI: 6.8\u0026ndash;20.2%; Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and Gemini (ARD: 8.3%, 95% CI: 2.8\u0026ndash;13.7%; Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). ChatGPT exhibited a significantly higher hallucination rate than Gemini (ARD: 5.3%, 95% CI: 0.1\u0026ndash;10.4%; Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.048). The overall hallucination rate reduction achieved by the multi-agent system relative to ChatGPT was 90.0% (from 15.0% to 1.5%), and relative to Gemini was 84.7% (from 9.8% to 1.5%). The odds ratio for hallucination occurrence was 0.087 (95% CI: 0.019\u0026ndash;0.392) for the multi-agent system versus ChatGPT and 0.139 (95% CI: 0.031\u0026ndash;0.631) versus Gemini, indicating a greater than 7-fold and 11-fold reduction in hallucination odds, respectively.\u003c/p\u003e \u003cp\u003eDetailed characterization of hallucination subtypes is presented in Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e. Five hallucination categories were defined: symptom fabrication, imaging finding distortion, laboratory value invention, medical history addition, and minor inferential assumption. Among ChatGPT hallucinations (n\u0026thinsp;=\u0026thinsp;20, total rate 15.0%), the predominant type was symptom fabrication, occurring in 11 cases (55.0% of ChatGPT hallucinations; 8.3% of all ChatGPT outputs), which involved the invention of symptoms not mentioned anywhere in the original case vignette\u0026mdash;for example, describing \"projectile vomiting\" when only nausea was documented, reporting \"severe peritoneal signs with rebound tenderness\" when the case described only mild abdominal tenderness, or adding \"bloody stool\" when no gastrointestinal bleeding was mentioned. Imaging finding distortion accounted for 6 cases (30.0% of ChatGPT hallucinations; 4.5% of all ChatGPT outputs), including descriptions of radiological findings inconsistent with the reported imaging\u0026mdash;such as describing a \"whirl sign on CT\" in a case of simple adhesive obstruction where no whirl sign was documented, adding \"free intraperitoneal air suggestive of perforation\" not present in the original CT report, or claiming \"portal venous gas\" when no such finding was described. Laboratory value invention was identified in 3 cases (15.0% of ChatGPT hallucinations; 2.3% of all ChatGPT outputs), involving fabrication of specific numerical laboratory results not provided in the case data\u0026mdash;such as reporting \"serum lactate of 4.2 mmol/L\" when no lactate measurement was documented, stating \"elevated procalcitonin of 8.5 ng/mL\" when procalcitonin was not measured, or citing a specific white blood cell count that differed from the actual reported value. ChatGPT produced zero medical history additions and zero minor inferential assumptions.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eCharacterization of Hallucination Types by AI System\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHallucination Type\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eChatGPT n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGemini n (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMulti-Agent n (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eTotal hallucinations\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e20 (100)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e13 (100)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (100)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eSymptom fabrication\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e11 (55.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e8 (61.5)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eImaging finding distortion\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e6 (30.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1 (7.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLaboratory value invention\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e3 (15.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMedical history addition\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e4 (30.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eMinor inferential assumption\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0 (0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e2 (100)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eAmong Gemini hallucinations (n\u0026thinsp;=\u0026thinsp;13, total rate 9.8%), symptom fabrication was the most common subtype, occurring in 8 cases (61.5% of Gemini hallucinations; 6.0% of all Gemini outputs), followed by medical history addition in 4 cases (30.8%; 3.0%), which involved attributing past medical history not provided in the case vignette\u0026mdash;such as adding a history of prior colonic resection, stating the patient had \"known inflammatory bowel disease\" when no such diagnosis was documented, or reporting \"previous episodes of volvulus\" when this history was absent. Imaging finding distortion accounted for 1 case (7.7%; 0.8%). Notably, Gemini produced zero laboratory value inventions and zero minor inferential assumptions. The distinct hallucination subtype profiles of the two monolithic LLMs are noteworthy: ChatGPT exhibited a higher tendency toward imaging finding distortion and laboratory value invention (9/20, 45.0% of its hallucinations), while Gemini showed a greater propensity for medical history fabrication (4/13, 30.8% of its hallucinations), suggesting fundamental differences in the models' gap-filling behaviors and contextual inference patterns.\u003c/p\u003e \u003cp\u003eThe two hallucinations identified in multi-agent system outputs (total rate 1.5%) were both classified as minor inferential assumptions\u0026mdash;the lowest severity category. The first involved a clinically reasonable inference of probable chronic constipation history from the clinical presentation pattern of an elderly patient with sigmoid volvulus, when constipation was not explicitly documented. The second involved suggesting a likely postoperative etiology for paralytic ileus when the temporal relationship between a mentioned surgical procedure and symptom onset was not explicitly stated. Critically, neither of these minor assumptions affected the diagnostic conclusions or therapeutic recommendations in any way. The multi-agent system produced zero high-severity hallucinations: 0/133 for symptom fabrication (0%, 95% CI: 0.0\u0026ndash;2.7%), 0/133 for imaging finding distortion (0%, 95% CI: 0.0\u0026ndash;2.7%), 0/133 for laboratory value invention (0%, 95% CI: 0.0\u0026ndash;2.7%), and 0/133 for medical history addition (0%, 95% CI: 0.0\u0026ndash;2.7%). This complete elimination of high-severity hallucinations is attributable to the Gyan LLM symbolic validation agent's deterministic factual grounding verification, which systematically cross-references every clinical assertion in the output against the original input data through compositional hallucination detection.\u003c/p\u003e \u003cp\u003eAmong the 20 ChatGPT outputs containing hallucinations, 14 (70.0%) were also diagnostically incorrect, compared to 32 of 113 non-hallucinating outputs (28.3%), yielding a statistically significant association between hallucination presence and diagnostic error (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001; odds ratio: 5.87, 95% CI: 2.08\u0026ndash;16.59). Similarly, among the 13 Gemini outputs with hallucinations, 10 (76.9%) were diagnostically incorrect, versus 45 of 120 non-hallucinating outputs (37.5%) (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.009; odds ratio: 5.53, 95% CI: 1.47\u0026ndash;20.77). These findings demonstrate that hallucinations are not merely cosmetic errors but are strongly predictive of diagnostic failure, potentially reflecting underlying confusion in the models' clinical reasoning processes.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003e3.2.4 Explanation Adequacy\u003c/h2\u003e \u003cp\u003eAdequate clinical reasoning, defined as a structured explanation satisfying all four sub-criteria\u0026mdash;(1) logical connection between findings and diagnosis, (2) reference to key supportive evidence, (3) discussion of relevant differential diagnoses, and (4) pathophysiological rationale for diagnosis and treatment\u0026mdash;was demonstrated in 107 of 133 multi-agent outputs (80.5%, 95% CI: 72.6\u0026ndash;86.8%), 92 of 133 ChatGPT outputs (69.2%, 95% CI: 60.6\u0026ndash;76.9%), and 78 of 133 Gemini outputs (58.6%, 95% CI: 49.8\u0026ndash;67.0%).\u003c/p\u003e \u003cp\u003eThe multi-agent system provided significantly more adequate explanations than both ChatGPT (ARD: 11.3%, 95% CI: 1.2\u0026ndash;21.4%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 7.04; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.008) and Gemini (ARD: 21.8%, 95% CI: 11.0\u0026ndash;32.6%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 16.53; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). ChatGPT also demonstrated significantly better explanation quality than Gemini (ARD: 10.5%, 95% CI: 0.3\u0026ndash;20.7%; McNemar \u003cem\u003eχ\u003c/em\u003e\u0026sup2; = 4.57; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.032). This represents a notable exception to the general pattern of comparable performance between the two monolithic LLMs, suggesting that GPT-4 Turbo possesses superior clinical reasoning verbalization capabilities compared to Gemini 2.0 Pro.\u003c/p\u003e \u003cp\u003eDisaggregation by individual explanation sub-criteria revealed differential performance patterns across systems. For logical connection between findings and diagnosis, adequacy rates were: multi-agent 88.0% (117/133), ChatGPT 79.7% (106/133), Gemini 71.4% (95/133). For reference to key supportive evidence: multi-agent 85.0% (113/133), ChatGPT 76.7% (102/133), Gemini 66.9% (89/133). For differential diagnosis discussion: multi-agent 82.7% (110/133), ChatGPT 70.7% (94/133), Gemini 62.4% (83/133). For pathophysiological rationale: multi-agent 84.2% (112/133), ChatGPT 73.7% (98/133), Gemini 63.9% (85/133). The multi-agent system outperformed both monolithic LLMs across all four sub-criteria, with the greatest advantage observed in differential diagnosis discussion (multi-agent vs. Gemini: 20.3 percentage-point difference) and pathophysiological rationale (multi-agent vs. Gemini: 20.3 percentage-point difference).\u003c/p\u003e \u003cp\u003eThe multi-agent outputs characteristically featured a three-tiered explanatory structure: (1) a radiological findings summary generated by the Hulu-Med perception agent, systematically identifying key imaging features including bowel dilation measurements, air-fluid level quantification, transition point localization, and pathognomonic sign identification with confidence scores; (2) a pathophysiological reasoning section generated by Med-PaLM 2, integrating radiological findings with clinical history, physical examination, and laboratory data to construct a ranked differential diagnosis with explicit evidence citation for each diagnostic consideration; and (3) a validation section generated by the Gyan symbolic agent, explicitly documenting which clinical data points support or contradict the proposed diagnosis, listing all verified and flagged assertions, and providing the final validated assessment with an audit trail. Among the 26 multi-agent outputs rated as inadequate in explanation quality, the predominant deficiency was insufficient differential diagnosis breadth (n\u0026thinsp;=\u0026thinsp;14, 53.8% of inadequate outputs), followed by incomplete integration of laboratory findings (n\u0026thinsp;=\u0026thinsp;8, 30.8%), and oversimplification of treatment rationale without guideline citation (n\u0026thinsp;=\u0026thinsp;4, 15.4%).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section3\"\u003e \u003ch2\u003e3.2.5 Critical Safety Errors\u003c/h2\u003e \u003cp\u003eNo critical safety errors\u0026mdash;defined as AI-generated recommendations that, if followed without clinical oversight, could result in patient harm through delayed recognition of a surgical emergency, recommendation of a contraindicated intervention, or failure to identify an immediately life-threatening condition\u0026mdash;were identified in any of the 133 multi-agent system outputs (0/133, 0%, 95% CI: 0.0\u0026ndash;2.7%). ChatGPT produced 5 outputs containing critical safety errors (5/133, 3.8%, 95% CI: 1.2\u0026ndash;8.5%), and Gemini produced 3 outputs with critical errors (3/133, 2.3%, 95% CI: 0.5\u0026ndash;6.4%). The difference in critical safety error rates between the multi-agent system and ChatGPT was statistically significant (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.024), while the difference between the multi-agent system and Gemini approached but did not reach conventional significance (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.083). The comparison between ChatGPT and Gemini was not statistically significant (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.48). The combined critical safety error rate for both monolithic LLMs was 8/266 evaluations (3.0%), compared to 0/133 (0%) for the neurosymbolic system (Fisher's exact \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.042).\u003c/p\u003e \u003cp\u003eAll eight critical safety errors across both single-model systems involved failure to recognize surgical emergencies requiring urgent operative intervention, rather than recommendation of actively harmful interventions. The five ChatGPT critical errors were: (1) recommending endoscopic intervention for sigmoid volvulus with radiological and clinical features of strangulation, including peritoneal signs and elevated lactate, where emergency surgical resection was mandated; (2) suggesting conservative management with intravenous fluids and nasogastric decompression for cecal volvulus, a condition requiring surgical management (cecectomy or right hemicolectomy) in virtually all cases given the low success rate of endoscopic reduction; (3) advising medical observation with serial abdominal examinations for small bowel volvulus with evidence of closed-loop obstruction on CT; (4) recommending elective outpatient surgical evaluation for a patient with closed-loop obstruction demonstrating CT features of bowel wall ischemia including reduced wall enhancement and mesenteric fat stranding; and (5) suggesting outpatient follow-up with a gastroenterologist for toxic megacolon with critical colonic diameter (\u0026gt;\u0026thinsp;12 cm) and systemic inflammatory response syndrome criteria, where urgent colectomy or decompression was indicated.\u003c/p\u003e \u003cp\u003eThe three Gemini critical errors were: (1) recommending enema decompression alone for sigmoid volvulus with CT evidence of mesenteric vessel engorgement and bowel wall edema, without recognizing the need for urgent surgical consultation given the strangulation risk; (2) suggesting 24-hour observation with repeat imaging for complicated adhesive ileus with CT features of a tight transition point, proximal bowel compromise, and small bowel feces sign, where surgical exploration was warranted; and (3) recommending only prokinetic agents (neostigmine) for Ogilvie syndrome with critical cecal diameter exceeding 12 cm and progressive distension over 72 hours, without acknowledging the 3\u0026ndash;15% perforation risk at this diameter threshold or the need for urgent colonoscopic decompression or surgical intervention. In all eight cases with critical safety errors by monolithic LLMs, the neurosymbolic multi-agent system correctly identified the surgical urgency, with the Gyan validation agent specifically activating safety rules S1 through S4 to flag clinical parameters exceeding established safety thresholds and override any tendency toward conservative management recommendations.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec25\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Subgroup Analysis by Diagnostic Category\u003c/h2\u003e \u003cp\u003eStratified analysis of diagnostic accuracy by disease category is presented in Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e, and the comparative performance profiles are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e (clustered bar chart with 95% CI error bars) and Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e (radar/spider chart, scale 40\u0026ndash;100% to emphasize inter-system differences). All subgroup comparisons were performed using Fisher's exact test due to small cell sizes in several diagnostic categories.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDiagnostic Accuracy by Disease Category\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic Category\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eChatGPT n/N (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eGemini n/N (%)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eMulti-Agent n/N (%)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eVolvulus (n\u0026thinsp;=\u0026thinsp;28)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e22/28 (78.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e21/28 (75.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e28/28 (100.0)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSigmoid volvulus (n\u0026thinsp;=\u0026thinsp;20)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e16/20 (80.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e15/20 (75.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e20/20 (100.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCecal volvulus (n\u0026thinsp;=\u0026thinsp;6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e4/6 (66.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e4/6 (66.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6/6 (100.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSmall bowel volvulus (n\u0026thinsp;=\u0026thinsp;2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2/2 (100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e2/2 (100.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e2/2 (100.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther mech. obstruction (n\u0026thinsp;=\u0026thinsp;27)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e18/27 (66.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e17/27 (63.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e22/27 (81.5)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eOgilvie syndrome (n\u0026thinsp;=\u0026thinsp;37)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e18/37 (48.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19/37 (51.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25/37 (67.6)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eParalytic ileus (n\u0026thinsp;=\u0026thinsp;20)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14/20 (70.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e13/20 (65.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e16/20 (80.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eToxic megacolon (n\u0026thinsp;=\u0026thinsp;12)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e5/12 (41.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e5/12 (41.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e6/12 (50.0)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOther rare etiologies (n\u0026thinsp;=\u0026thinsp;9)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e3/9 (33.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e3/9 (33.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e3/9 (33.3)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eOverall (N\u0026thinsp;=\u0026thinsp;133)\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e80/133 (60.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e78/133 (58.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e100/133 (75.2)*\u0026dagger;\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"4\"\u003e\u003cem\u003e*p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 vs ChatGPT; \u0026dagger;p\u0026thinsp;\u0026lt;\u0026thinsp;0.05 vs Gemini (Fisher\u0026rsquo;s exact test).\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe most pronounced inter-system performance gap was observed in volvulus cases (n\u0026thinsp;=\u0026thinsp;28), where the multi-agent system achieved 100% diagnostic accuracy (28/28, 95% CI: 87.7\u0026ndash;100%), significantly outperforming both ChatGPT at 78.6% (22/28, 95% CI: 59.0\u0026ndash;91.7%; p\u0026thinsp;=\u0026thinsp;0.008) and Gemini at 75.0% (21/28, 95% CI: 55.1\u0026ndash;89.3%; p\u0026thinsp;=\u0026thinsp;0.004). This advantage was consistent across volvulus subtypes: sigmoid volvulus accuracy was 100% (19/19) for the multi-agent system versus 84.2% (16/19) for ChatGPT and 78.9% (15/19) for Gemini; cecal volvulus 100% (6/6) versus 66.7% (4/6) for both comparators; and other subtypes 100% (3/3) versus 66.7% (2/3) for both. The performance gap was primarily driven by the Hulu-Med radiology agent's ability to recognize pathognomonic imaging signs: among 15 cases presenting with classic findings (9 coffee-bean sign, 4 whirl sign, 2 bird-beak sign), the multi-agent system identified all 15 correctly (100%), compared to 13/15 (86.7%) for ChatGPT and 12/15 (80.0%) for Gemini. ChatGPT's 6 volvulus misdiagnoses comprised 3 cases labeled as simple mechanical obstruction, 2 as Ogilvie syndrome, and 1 as paralytic ileus; Gemini's 7 errors included 4 classified as non-specific mechanical obstruction, 2 as Ogilvie syndrome, and 1 as adhesive obstruction.\u003c/p\u003e \u003cp\u003eFor adhesive/bridle obstruction (n\u0026thinsp;=\u0026thinsp;18), accuracy ranged from 61.1% to 77.8% across systems \u0026mdash; multi-agent 77.8% (14/18, 95% CI: 52.4\u0026ndash;93.6%), ChatGPT 66.7% (12/18, 95% CI: 41.0\u0026ndash;86.7%), Gemini 61.1% (11/18, 95% CI: 35.7\u0026ndash;82.7%) \u0026mdash; without statistically significant pairwise differences (all p\u0026thinsp;\u0026gt;\u0026thinsp;0.20; post hoc power: 0.31 for detecting a 16-percentage-point difference at this sample size). A similar pattern emerged in other mechanical causes (n\u0026thinsp;=\u0026thinsp;9; internal hernias, intussusception, gallstone ileus, Meckel's band, bezoar), where the multi-agent system achieved 66.7% (6/9, 95% CI: 29.9\u0026ndash;92.5%) versus 55.6% (5/9, 95% CI: 21.2\u0026ndash;86.3%) for both ChatGPT and Gemini, with wide confidence intervals precluding meaningful comparison.\u003c/p\u003e \u003cp\u003eOgilvie syndrome (n\u0026thinsp;=\u0026thinsp;37) proved diagnostically challenging across all systems, though a gradient favoring the multi-agent architecture was apparent: accuracy was 67.6% (25/37, 95% CI: 50.2\u0026ndash;82.0%) for the multi-agent system, 51.4% (19/37, 95% CI: 34.4\u0026ndash;68.0%) for Gemini, and 48.6% (18/37, 95% CI: 32.0\u0026ndash;65.6%) for ChatGPT. The multi-agent versus ChatGPT comparison approached significance (p\u0026thinsp;=\u0026thinsp;0.072; ARD: 18.9%, 95% CI: \u0026minus;1.6 to 39.5%), while multi-agent versus Gemini did not (p\u0026thinsp;=\u0026thinsp;0.14). The dominant misclassification pattern was labeling pseudo-obstruction as mechanical large bowel obstruction (multi-agent: 8/12 errors, 66.7%; ChatGPT: 14/19, 73.7%; Gemini: 12/18, 66.7%), reflecting the fundamental clinical difficulty of this differential when CT demonstrates marked colonic dilation without a definitive transition point. Paralytic ileus (n\u0026thinsp;=\u0026thinsp;20) showed a comparable trend \u0026mdash; multi-agent 75.0% (15/20, 95% CI: 50.9\u0026ndash;91.3%) versus 65.0% (13/20, 95% CI: 40.8\u0026ndash;84.6%) for both ChatGPT and Gemini (all p\u0026thinsp;\u0026gt;\u0026thinsp;0.30) \u0026mdash; with misdiagnoses predominantly involving confusion between paralytic ileus and early mechanical obstruction (multi-agent 3/5; ChatGPT 5/7; Gemini 4/7).\u003c/p\u003e \u003cp\u003eThe most uniformly difficult categories were toxic megacolon/severe colonic distension (n\u0026thinsp;=\u0026thinsp;12) and rare etiologies (n\u0026thinsp;=\u0026thinsp;9). For toxic megacolon, accuracy was 50.0% (6/12, 95% CI: 21.1\u0026ndash;78.9%) for the multi-agent system and 41.7% (5/12, 95% CI: 15.2\u0026ndash;72.3%) for both comparators (all p\u0026thinsp;\u0026gt;\u0026thinsp;0.60), with errors driven by difficulty distinguishing toxic megacolon from severe Ogilvie syndrome (multi-agent: 4/6 errors) or fulminant colitis without megacolon criteria (ChatGPT: 3/7; Gemini: 4/7). For rare etiologies, all three systems showed identical accuracy of 33.3% (3/9, 95% CI: 7.5\u0026ndash;70.1%); the 6 shared misdiagnoses included 2 internal hernias labeled as adhesive obstruction, 2 gallstone ileus cases classified as simple small bowel obstruction, and 2 intussusception cases labeled as non-specific mechanical obstruction. The convergent failure across architectures in these two categories suggests that both toxic megacolon discrimination and rare etiology identification represent fundamental limitations of current AI systems irrespective of design, likely attributable to their low prevalence in training corpora and the absence of reliable pathognomonic distinguishing features.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section2\"\u003e \u003ch2\u003e3.4 Processing Time and Operational Characteristics\u003c/h2\u003e \u003cp\u003eMean processing times from prompt submission to complete response generation were: ChatGPT 8.3\u0026thinsp;\u0026plusmn;\u0026thinsp;2.1 seconds (median: 7.8 seconds; IQR: 6.7\u0026ndash;9.4 seconds; range: 4.2\u0026ndash;14.8 seconds), Gemini 6.7\u0026thinsp;\u0026plusmn;\u0026thinsp;1.8 seconds (median: 6.2 seconds; IQR: 5.4\u0026ndash;7.6 seconds; range: 3.1\u0026ndash;12.3 seconds), and multi-agent system 47.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.6 seconds (median: 44.8 seconds; IQR: 39.1\u0026ndash;53.6 seconds; range: 28.4\u0026ndash;82.1 seconds). Processing time distributions were non-normal for all three systems (Shapiro-Wilk \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01 for all), with rightward skew reflecting occasional prolonged responses for complex cases. The multi-agent system required approximately 5.7-fold longer than ChatGPT (Wilcoxon rank-sum \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001; rank-biserial correlation r\u0026thinsp;=\u0026thinsp;1.00) and 7.0-fold longer than Gemini (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001; r\u0026thinsp;=\u0026thinsp;1.00). Gemini demonstrated significantly faster response generation than ChatGPT (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001; r\u0026thinsp;=\u0026thinsp;0.72).\u003c/p\u003e \u003cp\u003eThe increased processing time of the multi-agent system was attributable to its sequential three-stage architecture: the Hulu-Med radiology perception agent required a mean of 18.4\u0026thinsp;\u0026plusmn;\u0026thinsp;5.2 seconds (range: 9.1\u0026ndash;34.7 seconds), the Med-PaLM 2 clinical synthesis agent 16.8\u0026thinsp;\u0026plusmn;\u0026thinsp;4.1 seconds (range: 10.2\u0026ndash;28.3 seconds), and the Gyan LLM symbolic validation agent 12.0\u0026thinsp;\u0026plusmn;\u0026thinsp;3.8 seconds (range: 6.8\u0026ndash;24.1 seconds). The Hulu-Med agent demonstrated the longest and most variable processing times, with longer times associated with cases containing multiple imaging modalities or complex radiological findings requiring detailed feature extraction. Intermediate structured data transfer between pipeline stages via JSON schema accounted for the remaining overhead. Despite this increased latency, the total processing time of less than 90 seconds in all 133 cases remains well within clinically acceptable limits for decision support in acute abdominal presentations, where the clinical decision timeline typically spans 30\u0026ndash;60 minutes from initial emergency department assessment to final disposition decision. The multi-agent processing time also compares favorably with a formal multidisciplinary team consultation in real-world clinical practice, which typically requires 15\u0026ndash;30 minutes of specialist coordination.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section2\"\u003e \u003ch2\u003e3.5 Inter-Rater Reliability\u003c/h2\u003e \u003cp\u003eThe reliability of the evaluation framework was confirmed through comprehensive inter-rater agreement analysis. Cohen's kappa coefficients for agreement between the two independent assessors across all five evaluation criteria were: diagnostic accuracy \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.95 (95% CI: 0.91\u0026ndash;0.99; classified as excellent by Landis and Koch criteria; percentage agreement: 97.0%, 129/133 concordant evaluations per system), treatment appropriateness \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.88 (95% CI: 0.82\u0026ndash;0.94; excellent; percentage agreement: 91.0%, 121/133), hallucination detection \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.91 (95% CI: 0.85\u0026ndash;0.97; excellent; percentage agreement: 94.7%, 126/133), explanation adequacy \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.82 (95% CI: 0.75\u0026ndash;0.89; excellent; percentage agreement: 87.2%, 116/133), and critical safety errors \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.93 (95% CI: 0.86\u0026ndash;1.00; excellent; percentage agreement: 94.7%, 126/133). The overall inter-rater reliability across all 1,995 evaluations was \u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.89 (95% CI: 0.84\u0026ndash;0.94; overall percentage agreement: 92.9%). The slightly lower agreement for explanation adequacy (\u003cem\u003eκ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.82) compared to other domains reflects the inherently more subjective nature of assessing reasoning quality, which involves evaluative judgment regarding the depth and coherence of clinical argumentation rather than binary classification. Among the 47 total disagreements (47/1,995, 2.4%), the distribution by evaluation domain was: diagnostic accuracy 4 disagreements (8.5%), treatment appropriateness 12 (25.5%), hallucination detection 7 (14.9%), explanation adequacy 17 (36.2%), and critical safety errors 7 (14.9%). All disagreements were resolved to consensus by the third reviewer.\u003c/p\u003e \u003c/div\u003e"},{"header":"4. DISCUSSION","content":"\u003cp\u003eThis study provides the first empirical evidence that a neurosymbolic multi-agent architecture integrating domain-specific perception, clinical synthesis, and symbolic verification agents in a sequential pipeline outperforms general-purpose large language models in the diagnosis and management of ileus-spectrum and volvulus conditions. The neurosymbolic system achieved 75.2% diagnostic accuracy (95% CI: 66.9\u0026ndash;82.2%), representing a statistically significant 15.0 and 16.5 percentage-point improvement over ChatGPT (GPT-4 Turbo; 60.2%) and Gemini 2.0 Pro (58.6%), respectively (both p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). The multi-agent system also demonstrated a substantially lower hallucination rate of 1.5% with no high-severity events, compared to 15.0% for ChatGPT and 9.8% for Gemini, and produced no safety errors, whereas ChatGPT and Gemini yielded rates of 3.8% and 2.3%, respectively. These findings support the hypothesis that decomposing clinical reasoning into specialized cognitive subtasks, consistent with multidisciplinary team workflows, yields more reliable and safer AI-assisted decision support than monolithic general-purpose models.\u003c/p\u003e \u003cp\u003eThe diagnostic performance of 60.2% and 58.6% observed for ChatGPT and Gemini in our study is broadly consistent with performance ranges reported for general-purpose LLMs across medical specialties. Sussan et al. reported that both GPT-4 Turbo and Gemini-Pro achieved variable accuracy across medical licensing examinations, with performance declining substantially in clinically complex scenarios requiring multimodal integration [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. A systematic review of diagnostic accuracy of large language models in clinical settings found that diagnostic performance varied widely across studies, with primary diagnostic accuracy ranging from approximately 25% to 97.8% depending on task complexity and model evaluated, suggesting substantial heterogeneity in clinical performance among LLMs [\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Similarly, Mittal and Aggarwal demonstrated that LLMs exhibited diagnostic accuracy of 52\u0026ndash;68% in ophthalmic emergencies, with particular difficulty in cases requiring integration of imaging findings with clinical context [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Hager et al., using 2,400 real patient cases from the MIMIC-IV database across four common abdominal pathologies, showed that state-of-the-art LLMs performed significantly worse than physicians in autonomous clinical decision-making and failed to follow diagnostic or treatment guidelines [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. Mansoor et al., in a comprehensive systematic review of LLM reasoning techniques in medicine published in Health Information Science and Systems, further emphasized that while LLMs demonstrate unprecedented capabilities in medical reasoning tasks requiring complex inference and pattern recognition, their performance deteriorates significantly in high-stakes clinical scenarios demanding structured multi-step reasoning and guideline adherence [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e]. Our findings extend this body of evidence to the domain of acute intestinal obstruction, confirming that general-purpose LLMs, despite their broad medical knowledge, remain insufficiently reliable for clinical decision support in conditions where diagnostic error carries immediate surgical consequences.\u003c/p\u003e \u003cp\u003eThe superior performance of the neurosymbolic multi-agent system can be attributed to its architectural design, which operationalizes two complementary principles: task decomposition through specialized agents and symbolic verification through deterministic rule-based reasoning. Prenosil et al. demonstrated that a neurosymbolic AI combining GPT-4 with a rule-based expert system through a semantic integration platform achieved physician-level accuracy (99.8%) in extracting structured clinical data from radiology reports, with the symbolic component providing auditable verification trails [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Acharya and Song provided a comprehensive theoretical framework demonstrating that neurosymbolic integration enhances robustness, uncertainty quantification, and intervenability\u0026mdash;three properties essential for clinical deployment [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e]. Miladinovic et al. applied a neurosymbolic framework to retinal disease classification from OCT images and demonstrated that symbolic constraints improved both explainability and diagnostic consistency compared to purely neural approaches [\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Our three-stage pipeline, comprising Hulu-Med 32B for radiological perception, Med-PaLM 2 for clinical synthesis, and Gyan LLM for symbolic validation, represents a practical implementation of neurosymbolic principles in a clinically relevant domain. Adnan et al. recently demonstrated that neurosymbolic digital twin architectures combining neural pattern recognition with symbolic reasoning achieve superior performance in cardiovascular disease prediction and personalized modeling, further validating the translational potential of neurosymbolic approaches across diverse clinical domains [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe multi-agent architecture employed in this study aligns with an emerging paradigm in medical AI that leverages collaborative agent systems for complex clinical reasoning. Sorka et al. demonstrated that multi-agent approaches to neurological clinical reasoning, where specialized agents handle distinct cognitive subtasks, consistently outperformed single-model configurations in diagnostic accuracy and reasoning quality [\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e]. Chen et al., in a landmark study published in \u003cem\u003enpj Digital Medicine\u003c/em\u003e, showed that multi-agent conversational LLM systems enhanced diagnostic capability by enabling iterative refinement through inter-agent dialogue, achieving significant improvements over standalone models [\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e]. A recent survey of LLM-based multi-agent systems in medicine confirmed that multi-agent frameworks, including the MAC framework and GPT-4-based voting ensembles, consistently outperform single-agent setups in complex diagnostic reasoning, highlighting the critical role of collaborative mechanisms in optimizing clinical reliability [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e]. Our own previous work comparing multidisciplinary AI systems versus single-model approaches for ileus and volvulus diagnosis provided preliminary evidence for the superiority of multi-agent architectures in this specific clinical domain [\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e]. The present study extends these findings by incorporating a symbolic verification layer that further enhances factual grounding and safety.\u003c/p\u003e \u003cp\u003ePerhaps the most clinically significant finding of this study is the dramatic reduction in hallucination rates achieved by the neurosymbolic system. The 1.5% hallucination rate (2/133 cases, both minor inferential assumptions) represents a 90% reduction compared to ChatGPT (15.0%) and an 85% reduction compared to Gemini (9.8%). This finding is particularly notable given the severity of LLM hallucination in clinical contexts. Omar et al., in a large-scale evaluation, tested six leading LLMs with 300 physician-designed clinical vignettes containing fabricated medical details and found that every tested model repeated or elaborated on planted false information in 50\u0026ndash;82% of outputs, with even the best-performing model (GPT-4o) exhibiting a 23% hallucination rate under targeted mitigation prompts [\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e]. Hazra et al. reported that hallucination rates in image-based medical tasks remained alarmingly high across commercial LLMs, with fabricated imaging findings representing a particularly dangerous category [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. Roustan et al. provided a comprehensive clinician-oriented review emphasizing that LLM hallucinations in healthcare\u0026mdash;including symptom fabrication, laboratory value invention, and imaging finding distortion\u0026mdash;represent the most significant barrier to clinical deployment [\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e].The Gyan LLM symbolic validation agent in our pipeline addresses this challenge through deterministic factual grounding verification, cross-referencing each claim in the clinical output against the original case data and flagging unsupported assertions for correction. This compositional hallucination detection mechanism explains the virtual elimination of high-severity hallucinations in our system.\u003c/p\u003e \u003cp\u003eSubgroup analysis revealed important disease-specific performance patterns. The most pronounced advantage of the neurosymbolic system was observed in volvulus cases (100% accuracy vs. 78.6% and 75.0% for ChatGPT and Gemini, respectively; \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01), primarily attributable to the Hulu-Med radiology agent's ability to accurately identify pathognomonic imaging signs including the coffee-bean sign, whirl sign, and bird-beak sign. This finding is consistent with the radiological literature emphasizing the critical diagnostic value of these signs. Memis and Aydin demonstrated that sigmoid volvulus subtype classification based on imaging findings significantly impacts clinical course prediction, while Moloney et al. showed that specific CT features can predict volvulus outcomes and recurrence [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e, \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e]. The specialized vision-language training of Hulu-Med across 12 anatomical systems and 14 imaging modalities provides a natural advantage in recognizing these pathognomonic patterns, an advantage that general-purpose text-based LLMs inherently lack. Conversely, all systems demonstrated reduced accuracy for Ogilvie syndrome (67.6%, 48.6%, 51.4%) and rare etiologies (33.3% across all systems), reflecting the intrinsic diagnostic challenge of distinguishing functional from mechanical obstruction and the limited representation of uncommon conditions in training datasets.\u003c/p\u003e \u003cp\u003eThe absence of critical safety errors in the neurosymbolic system outputs (0/133 cases) compared to 5 errors for ChatGPT (3.8%) and 3 for Gemini (2.3%) merits particular emphasis. All eight critical errors across both monolithic LLMs involved failure to recognize surgical emergencies\u0026mdash;including recommending conservative management for cecal volvulus, suggesting endoscopic intervention for sigmoid volvulus with strangulation signs, and recommending only prokinetic agents for Ogilvie syndrome with critical cecal diameter exceeding 12 cm. Recent reviews have confirmed that state-of-the-art medical LLMs continue to exhibit substantial hallucination risks and challenges in clinical tasks, underscoring the need for rigorous validation, bias mitigation, and multimodal integration to ensure safe deployment in healthcare settings [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe three-tier verification framework of our pipeline includes radiological assessment, clinical synthesis, and symbolic validation. This structure introduces multiple safety checkpoints that are not present in single-model systems and may explain the complete elimination of critical safety errors observed in our study.\u003c/p\u003e \u003cp\u003eThe multi-agent system demonstrated higher explanation adequacy (80.5% compared with 69.2% for ChatGPT and 58.6% for Gemini), reflecting the inherent transparency of its pipeline architecture. The explanatory framework consists of a radiological findings summary generated by Hulu-Med, pathophysiological reasoning provided by Med-PaLM 2, and validation points from Gyan. This format resembles a multidisciplinary team consultation report and offers clinicians a structured audit trail for each diagnostic decision.This design philosophy aligns with the growing emphasis on explainable AI in healthcare. Recent evidence emphasizes that transparency and explainability are essential for clinical AI systems, as clinicians must be able to understand, verify, and critically appraise model reasoning to ensure safe and trustworthy decision support [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e]. The neurosymbolic paradigm, by structurally separating the knowledge base from the inference engine, enables this level of transparency in ways that opaque end-to-end neural models fundamentally cannot provide\u003c/p\u003e \u003cp\u003eThe multi-agent system's increased processing time (47.2\u0026thinsp;\u0026plusmn;\u0026thinsp;11.6 seconds vs. 8.3\u0026thinsp;\u0026plusmn;\u0026thinsp;2.1 and 6.7\u0026thinsp;\u0026plusmn;\u0026thinsp;1.8 seconds for ChatGPT and Gemini) represents an inherent trade-off of sequential multi-agent architectures. However, this latency remains well within clinically acceptable limits for decision support in acute abdominal presentations, where the clinical decision timeline typically spans 30\u0026ndash;60 minutes from initial assessment to disposition. Furthermore, this processing time compares favorably with the time required for a complete multidisciplinary consultation in a real-world clinical setting. Future optimization through parallel processing of independent pipeline stages could substantially reduce this latency without sacrificing the sequential verification logic that underpins the system's safety profile.\u003c/p\u003e \u003cp\u003eSeveral limitations of this study should be acknowledged. First, the retrospective, case report-based design introduces inherent selection bias, as published case reports may overrepresent unusual or diagnostically challenging presentations and underrepresent routine cases. Second, the use of published case reports rather than real-time clinical encounters limits ecological validity; AI systems may perform differently when processing structured case vignettes versus unstructured electronic health record data. Third, the single-center evaluation with two expert assessors, despite excellent inter-rater reliability (κ\u0026thinsp;=\u0026thinsp;0.89), may not fully capture the variability in clinical judgment across different institutional settings and cultural contexts. Fourth, the imaging data were provided as textual descriptions derived from published reports rather than as raw imaging files, potentially underestimating the Hulu-Med agent's full multimodal capabilities. Fifth, the 133-case sample size, although sufficient for primary comparisons (post-hoc power\u0026thinsp;\u0026gt;\u0026thinsp;0.90 for 15% absolute difference at α\u0026thinsp;=\u0026thinsp;0.05), limits statistical power for subgroup analyses, particularly in rare diagnostic categories. Sixth, the study evaluates AI performance in isolation rather than as an adjunct to human decision-making, which represents the most likely deployment scenario. Seventh, the absence of prospective clinical validation means that the real-world impact on patient outcomes, clinical workflows, and physician decision-making remains unknown. Finally, the rapidly evolving nature of both general-purpose LLMs and neurosymbolic architectures means that performance benchmarks may shift substantially with future model iterations.\u003c/p\u003e \u003cp\u003eFuture research should prioritize multicenter prospective validation studies incorporating diverse patient populations, real-time clinical data inputs including raw imaging files, and physician-AI collaborative decision-making paradigms. Integration with electronic health record systems would enable evaluation of the system's performance in authentic clinical workflows with naturally occurring data noise and incompleteness. Cost-effectiveness analyses comparing computational costs against diagnostic accuracy improvements and potential clinical complication reduction are warranted. The development of hybrid deployment models\u0026mdash;where the neurosymbolic system serves as a real-time decision support layer augmenting rather than replacing physician judgment\u0026mdash;represents the most promising translational pathway. Additionally, expanding the symbolic knowledge base to incorporate continuously updated clinical guidelines and extending the multi-agent framework to other acute abdominal pathologies and surgical emergencies could broaden the system's clinical utility.\u003c/p\u003e"},{"header":"5. CONCLUSIONS","content":"\u003cp\u003eA neurosymbolic multi-agent pipeline that decomposes the clinical reasoning workflow into specialized perception, synthesis, and symbolic verification stages significantly outperforms general-purpose monolithic LLMs in diagnosing and managing ileus-spectrum and volvulus-spectrum emergencies. The architectural separation of neural pattern recognition from symbolic rule-based verification substantially reduces hallucination and eliminates critical safety errors. For volvulus cases specifically, where pathognomonic radiological signs enable definitive diagnosis, specialized vision-language perception agents achieve perfect diagnostic accuracy. However, performance advantages diminish for diagnostically ambiguous conditions such as Ogilvie syndrome and toxic megacolon, where even structured multi-agent reasoning cannot fully compensate for inherent diagnostic complexity. These findings support the integration of neurosymbolic design principles\u0026mdash;combining neural perception with symbolic verification\u0026mdash;in clinical AI systems for acute abdominal pathology, while underscoring that AI outputs in these high-stakes settings must remain subject to expert physician oversight and verification until reliability is consistently demonstrated across the full spectrum of diagnostic complexity.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics Approval and Consent to Participate\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e This study was approved by the Ankara Provincial Directorate of Health Non-Interventional Ethics Committee (Approval No: 2025-10-3; Date: 24 October 2025). The study protocol entitled \u0026quot;The Role of a Specific Local Large Language Model in the Diagnosis of Internal Medicine Diseases\u0026quot; was reviewed and approved after evaluation of the study rationale, objectives, methodology, and ethical aspects. Given the retrospective design utilizing anonymized data extracted from previously published case reports in the peer-reviewed literature, the requirement for individual informed consent was waived by the ethics committee. All procedures were conducted in accordance with the Declaration of Helsinki (2013 revision) and relevant national regulations.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for Publication\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e Not applicable. This study exclusively utilized data from previously published, anonymized case reports available in the public domain. No individual participant data requiring consent for publication were included.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of Data and Materials\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request. The 133 case vignettes were reconstructed from PubMed-indexed case reports published between January 2022 and December 2025; the complete list of source publications with PubMed identifiers (PMIDs) is provided in Supplementary Table S1. The standardized prompt templates used for AI system evaluation, the structured data extraction forms, and the raw evaluation scores from both independent assessors are available as supplementary materials. The AI-generated outputs from all three systems (ChatGPT GPT-4 Turbo, Gemini 2.0 Pro, and the neurosymbolic multi-agent system) are archived and available upon request for reproducibility verification. The neurosymbolic multi-agent system pipeline configuration, including agent-specific prompts and inter-agent JSON communication schemas, is described in detail in the Methods section; the complete technical implementation files are available from the corresponding author. Due to the use of proprietary AI platforms (ChatGPT and Gemini), full computational reproducibility is subject to model version availability and API access at the time of replication.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e:The authors declare that they have no competing interests. No financial or non-financial conflicts of interest exist in relation to the content of this manuscript. None of the authors have any affiliation with or financial involvement in any organization or entity with a direct financial interest in the subject matter or materials discussed in this manuscript, including OpenAI (developer of ChatGPT), Google DeepMind (developer of Gemini and Med-PaLM 2), or the developers of Hulu-Med or Gyan LLM.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; Contributions\u003c/strong\u003e\u003cstrong\u003e:\u003c/strong\u003e M.U. conceptualized and designed the study, developed the neurosymbolic multi-agent system architecture, conducted the systematic literature search, performed data extraction, executed all AI system evaluations, performed statistical analyses, interpreted the results, and drafted the manuscript. S.K. contributed to data extraction, independently evaluated AI system outputs as a blinded assessor, participated in inter-rater reliability assessment, and critically revised the manuscript for intellectual content. K.Y. contributed to case selection and screening, served as an independent blinded assessor for AI output evaluation, resolved inter-rater disagreements through consensus discussion, and critically revised the manuscript. L.E. contributed to data collection, assisted with figure and table preparation, and critically reviewed the final manuscript. All authors read and approved the final version of the manuscript and agree to be accountable for all aspects of the work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e Not applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMwenitete, D., et al., \u003cem\u003eDeterminants of surgical management outcomes among adult patients with intestinal obstruction at Mzuzu central hospital, Malawi.\u003c/em\u003e BMC Surg, 2025. \u003cstrong\u003e26\u003c/strong\u003e(1): p. 50.\u003c/li\u003e\n\u003cli\u003eInoue, K., et al., \u003cem\u003eSurgical Management of Sigmoid Volvulus: A Retrospective Review of Six Cases with a Focus on the Sharon Operation.\u003c/em\u003e Surg Case Rep, 2026. \u003cstrong\u003e12\u003c/strong\u003e(1).\u003c/li\u003e\n\u003cli\u003eMemis, K.B. and S. Aydin \u003cem\u003eRelationship Between Sigmoid Volvulus Subtypes, Clinical Course, and Imaging Findings\u003c/em\u003e. Diagnostics, 2025. \u003cstrong\u003e15\u003c/strong\u003e, 784 DOI: 10.3390/diagnostics15060784.\u003c/li\u003e\n\u003cli\u003eLarsen, T.B. and M.E. Lazarus, \u003cem\u003eCoffee Bean Sign.\u003c/em\u003e J Brown Hosp Med, 2025. \u003cstrong\u003e4\u003c/strong\u003e(3): p. 137903.\u003c/li\u003e\n\u003cli\u003eMoloney, B.M., et al., \u003cem\u003eSigmoid volvulus-Can CT features predict outcomes and recurrence?\u003c/em\u003e Eur Radiol, 2025. \u003cstrong\u003e35\u003c/strong\u003e(2): p. 897-905.\u003c/li\u003e\n\u003cli\u003eSussan, T.T., et al., \u003cem\u003eA Comparative Evaluation of GPT-4 Turbo and Gemini-Pro in Medical Licensing Exams: Enhancing Artificial Intelligence\u0026apos;s Role in Medical Education.\u003c/em\u003e Cureus, 2026. \u003cstrong\u003e18\u003c/strong\u003e(1): p. e101101.\u003c/li\u003e\n\u003cli\u003eHazra, D., et al., \u003cem\u003eEvaluating Hallucination and Diagnostic Reliability of LLMs on Medical Image-Based Multiple Choice Tasks.\u003c/em\u003e IEEE J Biomed Health Inform, 2025. \u003cstrong\u003ePp\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eBradshaw, T.J., et al., \u003cem\u003eLarge Language Models and Large Multimodal Models in Medical Imaging: A Primer for Physicians.\u003c/em\u003e J Nucl Med, 2025. \u003cstrong\u003e66\u003c/strong\u003e(2): p. 173-182.\u003c/li\u003e\n\u003cli\u003eMittal, S. and Y. Aggarwal, \u003cem\u003eEvaluation of Large Language Models in the Diagnosis, Urgency Triage, and Initial Management of Ophthalmic Emergencies.\u003c/em\u003e Cureus, 2026. \u003cstrong\u003e18\u003c/strong\u003e(1): p. e101433.\u003c/li\u003e\n\u003cli\u003ePrenosil, G.A., et al., \u003cem\u003eNeuro-symbolic AI for auditable cognitive information extraction from medical reports.\u003c/em\u003e Commun Med (Lond), 2025. \u003cstrong\u003e5\u003c/strong\u003e(1): p. 491.\u003c/li\u003e\n\u003cli\u003eAcharya, K. and H. Song, \u003cem\u003eA Comprehensive Review of Neuro-symbolic AI for Robustness, Uncertainty Quantification, and Intervenability.\u003c/em\u003e Arabian Journal for Science and Engineering, 2026. \u003cstrong\u003e51\u003c/strong\u003e(1): p. 35-67.\u003c/li\u003e\n\u003cli\u003eMiladinovic, A., et al., \u003cem\u003eNeurosymbolic AI Framework for Explainable Retinal Disease Classification From OCT Images.\u003c/em\u003e Transl Vis Sci Technol, 2026. \u003cstrong\u003e15\u003c/strong\u003e(1): p. 6.\u003c/li\u003e\n\u003cli\u003ePrenosil, G.A., et al., \u003cem\u003eNeuro-symbolic AI for auditable cognitive information extraction from medical reports.\u003c/em\u003e Communications Medicine, 2025. \u003cstrong\u003e5\u003c/strong\u003e(1): p. 491.\u003c/li\u003e\n\u003cli\u003eSorka, M., et al., \u003cem\u003eA multi-agent approach to neurological clinical reasoning.\u003c/em\u003e PLOS Digit Health, 2025. \u003cstrong\u003e4\u003c/strong\u003e(12): p. e0001106.\u003c/li\u003e\n\u003cli\u003eUcdal, M., K. Yurtsever, and E. Ekingen, \u003cem\u003eMultidisciplinary artificial intelligence systems versus single-model approaches for the diagnosis and management of ileus and volvulus.\u003c/em\u003e BMC Gastroenterol, 2026. \u003cstrong\u003e26\u003c/strong\u003e(1): p. 124.\u003c/li\u003e\n\u003cli\u003eRoustan, D. and F. Bastardot, \u003cem\u003eThe Clinicians\u0026apos; Guide to Large Language Models: A General Perspective With a Focus on Hallucinations.\u003c/em\u003e Interact J Med Res, 2025. \u003cstrong\u003e14\u003c/strong\u003e: p. e59823.\u003c/li\u003e\n\u003cli\u003eShan, G., et al., \u003cem\u003eComparing Diagnostic Accuracy of Clinical Professionals and Large Language Models: Systematic Review and Meta-Analysis.\u003c/em\u003e JMIR Med Inform, 2025. \u003cstrong\u003e13\u003c/strong\u003e: p. e64963.\u003c/li\u003e\n\u003cli\u003eHager, P., et al., \u003cem\u003eEvaluation and mitigation of the limitations of large language models in clinical decision-making.\u003c/em\u003e Nat Med, 2024. \u003cstrong\u003e30\u003c/strong\u003e(9): p. 2613-2622.\u003c/li\u003e\n\u003cli\u003eMansoor, I., et al., \u003cem\u003eReasoning with large language models in medicine: a systematic review of techniques, challenges and clinical integration.\u003c/em\u003e Health Inf Sci Syst, 2026. \u003cstrong\u003e14\u003c/strong\u003e(1): p. 6.\u003c/li\u003e\n\u003cli\u003eAdnan, M., et al., \u003cem\u003eNeurosymbolic Digital Twin for Cardiovascular Disease Prediction and Personalized Modeling.\u003c/em\u003e IEEE J Biomed Health Inform, 2025. \u003cstrong\u003ePp\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eChen, X., et al., \u003cem\u003eEnhancing diagnostic capability with multi-agents conversational large language models.\u003c/em\u003e NPJ Digit Med, 2025. \u003cstrong\u003e8\u003c/strong\u003e(1): p. 159.\u003c/li\u003e\n\u003cli\u003eXu, X. and R. Sankar, \u003cem\u003eLarge Language Model Agents for Biomedicine: A Comprehensive Review of Methods, Evaluations, Challenges, and Future Directions.\u003c/em\u003e Information, 2025. \u003cstrong\u003e16\u003c/strong\u003e(10): p. 894.\u003c/li\u003e\n\u003cli\u003eOmar, M., et al., \u003cem\u003eMulti-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support.\u003c/em\u003e Commun Med (Lond), 2025. \u003cstrong\u003e5\u003c/strong\u003e(1): p. 330.\u003c/li\u003e\n\u003cli\u003eMemis, K.B. and S. Aydin, \u003cem\u003eRelationship Between Sigmoid Volvulus Subtypes, Clinical Course, and Imaging Findings.\u003c/em\u003e Diagnostics (Basel), 2025. \u003cstrong\u003e15\u003c/strong\u003e(6).\u003c/li\u003e\n\u003cli\u003eMartinho, D., et al., \u003cem\u003eEthical Responsibility in Medical AI: A Semi-Systematic Thematic Review and Multilevel Governance Model.\u003c/em\u003e Healthcare (Basel), 2026. \u003cstrong\u003e14\u003c/strong\u003e(3).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"neurosymbolic artificial intelligence, multi-agent system, large language model, ileus, volvulus, clinical decision support, diagnostic accuracy, hallucination mitigation","lastPublishedDoi":"10.21203/rs.3.rs-9045948/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9045948/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003e\u003cstrong\u003eBackground\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eGeneral-purpose large language models (LLMs) demonstrate variable diagnostic accuracy and residual hallucination when applied to complex surgical emergencies. Whether a neurosymbolic multi-agent architecture—integrating domain-specific vision-language models, medically fine-tuned reasoning engines, and compositional verification agents—can outperform monolithic LLMs in ileus and volvulus case assessment remains unexplored.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe conducted a retrospective diagnostic accuracy study using 133 adult case vignettes (median age 62 years; 57.9% male) reconstructed from PubMed-indexed case reports published between January 2022 and December 2025. Three AI systems were evaluated: ChatGPT (GPT-4 Turbo), Gemini 2.0 Pro, and a sequential neurosymbolic multi-agent hybrid system comprising a radiology vision-language agent (Hulu-Med 32B), a clinical reasoning agent (Med-PaLM 2), and a compositional validation agent (Gyan LLM). Standardized prompts were submitted in zero-shot configuration. Two blinded expert assessors independently evaluated five predefined criteria: diagnostic accuracy, treatment appropriateness, hallucination presence, explanation adequacy, and critical safety errors. Inter-rater reliability was assessed using Cohen’s kappa. McNemar’s test with Bonferroni correction was used for pairwise comparisons.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe neurosymbolic multi-agent system achieved significantly higher diagnostic accuracy (75.2%; 95% CI: 66.9–82.2%) compared with ChatGPT (60.2%; 95% CI: 51.4–68.5%; p \u0026lt; 0.001) and Gemini (58.6%; 95% CI: 49.8–67.0%; p \u0026lt; 0.001). The multi-agent system also demonstrated superior treatment appropriateness (74.4% vs. 63.9% and 61.7%; both p \u0026lt; 0.017), markedly lower hallucination rates (1.5% vs. 15.0% and 9.8%; both p \u0026lt; 0.001), and zero critical safety errors (0% vs. 3.8% and 2.3%). Subgroup analysis revealed perfect diagnostic accuracy (100%) for volvulus cases in the multi-agent system versus 78.6% and 75.0% for the single-model systems. Performance convergence was observed in diagnostically ambiguous entities including Ogilvie syndrome (67.6% vs. 48.6% and 51.4%) and toxic megacolon (50.0% vs. 41.7%).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA neurosymbolic multi-agent pipeline that decomposes the clinical reasoning workflow into specialized perception, synthesis, and verification stages significantly outperforms general-purpose LLMs in diagnosing and managing ileus-spectrum and volvulus-spectrum emergencies. The architectural separation of neural pattern recognition from symbolic rule-based verification substantially reduces hallucination and eliminates critical safety errors. These findings support the integration of neurosymbolic design principles in clinical AI systems for acute abdominal pathology, while underscoring persistent limitations in diagnostically ambiguous conditions.\u003c/p\u003e","manuscriptTitle":"Neurosymbolic Multi-Agent Artificial Intelligence versus General-Purpose Large Language Models for Clinical Decision Support in Ileus and Volvulus","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-10 16:49:19","doi":"10.21203/rs.3.rs-9045948/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-04-25T20:41:38+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-22T23:42:55+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"218620668793299110752184731322231508506","date":"2026-04-15T06:53:15+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"146735918892524047675235864865486790914","date":"2026-04-12T15:23:36+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-04-12T07:30:29+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"279374749437171339447367387220989127630","date":"2026-04-12T07:15:05+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"56017329322423843905019601924080624974","date":"2026-04-09T18:37:13+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-03T14:56:40+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-03-10T09:29:17+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-06T12:28:45+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-06T12:23:56+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2026-03-06T04:26:23+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-medical-informatics-and-decision-making","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"midm","sideBox":"Learn more about [BMC Medical Informatics and Decision Making](http://bmcmedinformdecismak.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/midm/default.aspx","title":"BMC Medical Informatics and Decision Making","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8cb25e62-1527-468d-a9f9-2738563b945e","owner":[],"postedDate":"April 10th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-04-10T16:49:20+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-10 16:49:19","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9045948","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9045948","identity":"rs-9045948","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}