Performance Comparison of a Neuro-Symbolic Large Language Model System Versus Conventional AI Models and Human Experts in Cholangitis Management

Performance Comparison of a Neuro-Symbolic Large Language Model System Versus Conventional AI Models and Human Experts in Cholangitis Management | 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 Performance Comparison of a Neuro-Symbolic Large Language Model System Versus Conventional AI Models and Human Experts in Cholangitis Management Mete Ucdal, Evren Ekingen This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8714103/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Background Large language models (LLMs) have shown promising results in medical decision support; Background: Large language models (LLMs) have demonstrated promising outcomes in medical decision support; however, their efficacy in managing complex hepatobiliary conditions remains insufficiently examined. We have developed a genetic neuro-symbolic LLM system that integrates multiple AI agents with neural-symbolic reasoning for the management of cholangitis, and we have compared its performance to that of conventional LLMs and human experts.genetic neuro-symbolic LLM system integrating multiple AI agents with neural-symbolic reasoning for cholangitis management and compared its performance against conventional LLMs and human experts. Methods This multi-center cross-sectional study included 30 case-based questions from American Board of Internal Medicine (ABIM) gastroenterology subspecialty examinations covering acute cholangitis. Questions were categorized into diagnosis (n = 10), treatment (n = 10), and complications/prognosis (n = 10). Performance of a genetic neuro-symbolic LLM system orchestrated via LangGraph was compared against Claude 4.5 Sonnet, ChatGPT 5.2, Gemini 3 Pro, 10 gastroenterology specialists, and 4 emergency medicine physicians from four tertiary centers in Turkey. Results The genetic neuro-symbolic system achieved the highest overall accuracy (100%, 30/30), significantly outperforming Claude 4.5 Sonnet (90.0%), ChatGPT 5.2 (53.3%), Gemini 3 Pro (56.7%), gastroenterology experts (mean 96.3% ± 2.1%), and emergency medicine physicians (mean 89.2% ± 4.8%). The neuro-symbolic system demonstrated superior performance across all categories and cholangitis subtypes. Among human participants, gastroenterologists outperformed emergency physicians in treatment decisions (p = 0.012) and showed non-inferior performance to Gemini 3 Pro overall (p = 0.034). ROC analysis revealed excellent discrimination for the neuro-symbolic system (AUC = 1.000) compared to Claude (AUC = 0.924), ChatGPT (AUC = 0.687), and Gemini (AUC = 0.712). Conclusions The genetic neuro-symbolic LLM system demonstrated superior accuracy in cholangitis management compared to all conventional AI models and human experts. This multi-agent architecture with neural-symbolic reasoning represents a significant advancement in AI-assisted clinical decision support for complex hepatobiliary conditions. Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction The integration of artificial intelligence (AI) into clinical decision support systems exemplifies one of the most revolutionary advancements in contemporary medicine. Large Language Models (LLMs), in particular, have exhibited exceptional capacities in analyzing unstructured clinical data, responding to medical inquiries, and assisting in diagnostic reasoning ( 1 ). Their application encompasses various medical specialties, providing potential tools for education, documentation, and initial analysis. Nevertheless, considerable challenges remain in implementing these models within complex, high-stakes clinical settings, particularly in specialized domains such as hepatology. Conventional large language models frequently produce fluent and contextually appropriate responses; however, these outputs are not consistently underpinned by explicit clinical reasoning, systematic application of established medical guidelines, or dependable execution of multi-step diagnostic decision-making processes ( 2 , 3 ). This limitation manifests in inconsistencies, factual hallucinations, and an inability to reliably apply structured clinical algorithms to nuanced patient presentations ( 4 ). These limitations are particularly apparent in the management of acute cholangitis. Acute cholangitis is a potentially life-threatening biliary tract infection that requires prompt diagnosis, severity assessment, and timely intervention ( 5 ). The Tokyo Guidelines 2018 (TG18) provide a standardized framework for the diagnosis and severity grading of acute cholangitis, incorporating clinical signs, laboratory findings, and imaging features ( 6 ). Optimal management requires precise integration of these parameters within the TG18 framework to determine appropriate treatment strategies, including antibiotic therapy, biliary drainage timing, and intervention modality selection ( 6 ).Current LLMs, when tasked with such challenges, frequently exhibit guideline misapplication, difficulty in synthesizing multimodal data, and failure in complex differential diagnosis, leading to potentially unsafe recommendations ( 7 ). Consequently, their utility as standalone clinical tools remains limited without mechanisms for verification, reasoning traceability, and adherence to medical knowledge structures. To address these limitations, neuro-symbolic AI has emerged as a promising paradigm. This approach combines the pattern recognition capabilities of neural networks with the logical reasoning and explicit knowledge representation of symbolic AI ( 8 ). Neural networks excel at processing unstructured data, while symbolic systems operate on defined rules and enable deductive reasoning. A neuro-symbolic system can parse clinical vignettes using its neural components, extract relevant features, and map them onto a symbolic knowledge graph of medical guidelines ( 9 ). This process enables explicit and auditable clinical reasoning. Recent studies have demonstrated the feasibility of such architectures for specific medical tasks, with improvements in accuracy, reliability, and explainability compared to conventional LLMs ( 10 , 11 ). However, the application of neuro-symbolic AI to the comprehensive management of complex disease groups such as cholangitis, which involves multiple subtypes and management phases, remains largely unexplored ( 10 ). Building on this foundation, we implemented an NS-LLM system for cholangitis management that uses a multi-agent framework to orchestrate multiple base LLMs and integrates a structured symbolic knowledge base of current clinical guidelines and scoring systems; key innovations include a genetic algorithm for dynamic prompt optimization and a neural-symbolic integration layer that converts natural language input into structured queries against the knowledge graph to support clinical reasoning. Therefore, the primary objective of this multi-center, cross-sectional study was to conduct a head-to-head performance comparison of this novel Genetic Neuro-Symbolic LLM system against leading conventional LLMs (Claude 4.5 Sonnet, ChatGPT 5.2, Gemini 3 Pro) and human expert physicians (gastroenterologists and emergency medicine specialists) across a validated set of case-based questions covering the diagnosis, treatment, and prognosis of various cholangitis subtypes. We hypothesized that the neuro-symbolic system would achieve superior diagnostic accuracy and clinical reasoning fidelity by mitigating the core limitations of purely neural approaches, thereby establishing a new benchmark for AI-assisted decision support in complex hepatobiliary disease. Methods Study Design and Setting This multi-center cross-sectional diagnostic accuracy study was conducted between October 2025 and January 2026 at four tertiary healthcare centers in Turkey: Ankara Bilkent City Hospital, Ankara Etlik City Hospital, Elazığ Fethi Sekin City Hospital (EAH), and Etimesgut Şehit Sait Ertürk State Hospital Emergency Department. The study was designed in accordance with the Standards for Reporting Diagnostic Accuracy Studies (STARD) guidelines and the STROBE checklist for cross-sectional studies. Institutional ethics committee approval was obtained from Hacettepe University Faculty of Medicine Ethics Committee (Protocol No: 2025-GOA-0847, Date: October 15, 2025). Participants Human Expert Selection We recruited 14 physician participants from four tertiary care centers in Turkey. The gastroenterology group consisted of 10 specialists from the following institutions: Ankara Bilkent City Hospital (n = 3; two professors and one associate professor), Ankara Etlik City Hospital (n = 3; one professor and two associate professors), Elazığ Fethi Sekin City Hospital (n = 2; one professor and one associate professor), and Etimesgut Şehit Sait Ertürk State Hospital (n = 2; two specialists). The emergency medicine group consisted of 4 specialists: Ankara Bilkent City Hospital (n = 2), Ankara Etlik City Hospital (n = 1), and Elazığ Fethi Sekin City Hospital (n = 1). Inclusion criteria for gastroenterology specialists : Board certification in gastroenterology or hepatology Minimum of 10 years of clinical experience following residency Academic appointment at associate professor level or higher (for university-affiliated centers) Active clinical practice involving hepatobiliary disorders Current affiliation with a tertiary care center performing endoscopic retrograde cholangiopancreatography (ERCP) Inclusion criteria for emergency medicine specialists : Board certification in emergency medicine Minimum of 10 years of clinical experience Regular management of acute biliary emergencies Employment at a tertiary care center with 24-hour ERCP capability Exclusion criteria for all participants : Prior involvement in AI-related clinical research Previous exposure to the study questions AI Model Selection Four AI systems were evaluated: ( 1 ) Claude 4.5 Sonnet (Anthropic, 2025 version), ( 2 ) ChatGPT 5.2 (OpenAI, GPT-4 based, 2025 version), ( 3 ) Gemini 3 Pro (Google, 2025 version), and ( 4 ) our proprietary Genetic Neuro-Symbolic LLM System. All conventional LLMs were accessed via their respective APIs without additional fine-tuning, using standardized prompting protocols. Each model completed three independent runs of the 30-question TG18-based assessment. The complete study flow diagram is presented in Fig. 1 . Test Instrument The evaluation instrument consisted of 30 case-based multiple-choice questions adapted from ABIM gastroenterology subspecialty board examination question banks. Each question presented a detailed clinical vignette including patient demographics, presenting symptoms, physical examination findings, laboratory results, and imaging reports (ultrasonography, MRCP, CT, or ERCP). All questions focused exclusively on acute cholangitis.Questions were stratified by clinical domain: diagnosis (n = 10), treatment (n = 10), and complications/prognosis (n = 10). Cases represented the full spectrum of acute cholangitis severity (Grade I, II, and III per TG18) and etiology including choledocholithiasis (n = 12), malignant biliary obstruction (n = 6), post-procedural/iatrogenic causes (n = 6), benign strictures (n = 4), and parasitic cholangitis (n = 2).Gold standard answers were established through consensus by a panel of three hepatology professors with more than 20 years of experience each, using the Tokyo Guidelines 2018 (TG18) for diagnostic criteria, severity grading, and management recommendations ( 6 ). Genetic Neuro-Symbolic LLM System Architecture The genetic neuro-symbolic large language model (LLM) system was developed as a multi-agent orchestration framework based on the LangGraph architecture (LangChain, 2024), an open-source library designed for constructing stateful, multi-agent applications utilizing large language models. The LangGraph framework facilitates the creation of cyclical computational graphs whereby multiple AI agents can interact, exchange information, and iteratively enhance their outputs through structured workflows. Our implementation employs this framework to coordinate parallel reasoning processes while ensuring explicit state management throughout the decision-making pipeline. Multi-Agent Architecture The system employs a parallel agent deployment strategy utilizing two state-of-the-art large language models: Gemini 2.0 Flash (Google DeepMind, 2025) and GPT-5.2 (OpenAI, 2025). Each agent operates as an independent reasoning entity that receives identical clinical vignettes as input. The agents process the clinical information through their respective neural architectures and generate candidate answers accompanied by calibrated confidence scores ranging from 0 to 1. This dual-agent approach serves two purposes: first, it provides redundancy that reduces single-point-of-failure errors inherent in individual LLM outputs; second, it enables cross-validation of reasoning pathways, as agreement between architecturally distinct models increases confidence in the generated response. The agents communicate through a shared state object maintained by LangGraph, which tracks intermediate reasoning steps, extracted clinical features, and provisional diagnoses throughout the inference pipeline ( 12 ). Symbolic Knowledge Base The symbolic reasoning component incorporates a structured knowledge graph implemented using a graph database architecture with Neo4j as the underlying storage engine. This knowledge base encodes the Tokyo Guidelines 2018 (TG18) clinical criteria as interconnected nodes (representing clinical findings, laboratory thresholds, severity grades, and management decisions) and directed edges (representing diagnostic implications, severity escalation pathways, and treatment algorithms). The graph structure enables efficient traversal of diagnostic and management pathways and supports complex queries that mirror clinical reasoning patterns for acute cholangitis. Reference Guideline The knowledge base is built exclusively upon the Tokyo Guidelines 2018, the international consensus guideline for diagnosis, severity grading, and management of acute cholangitis published in the Journal of Hepato-Biliary-Pancreatic Sciences. TG18 provides evidence-based, algorithmically structured recommendations that are ideally suited for symbolic encoding and rule-based inference. TG18 Diagnostic Criteria Module The diagnostic module encodes the TG18 three-domain requirement system for acute cholangitis diagnosis. Each domain is represented as a parent node with child nodes representing individual criteria: Domain A - Systemic Inflammation : - A-1: Fever > 38°C (> 100.4°F) - A-2: Laboratory evidence of inflammatory response: - White blood cell count 10,000/µL - C-reactive protein ≥ 1 mg/dL Domain B - Cholestasis : - B-1: Jaundice (total bilirubin ≥ 2 mg/dL) - B-2: Abnormal liver function tests: - Alkaline phosphatase > 1.5× upper limit of normal (ULN) - Gamma-glutamyl transferase > 1.5× ULN - Aspartate aminotransferase > 1.5× ULN - Alanine aminotransferase > 1.5× ULN Domain C - Biliary Imaging : - C-1: Biliary dilatation on imaging (ultrasound, CT, MRCP, or EUS) - C-2: Evidence of etiology on imaging: - Choledocholithiasis (stone visualization) - Biliary stricture (benign or malignant) - Biliary stent (with or without occlusion) - Other obstruction (parasitic, extrinsic compression) Diagnostic Classification Rules: The knowledge base encodes the following production rules for diagnosis: RULE_DX_DEFINITE: IF (Domain_A ≥ 1 criterion) AND (Domain_B ≥ 1 criterion) AND (Domain_C ≥ 1 criterion) THEN Diagnosis = "Definite Acute Cholangitis" [Confidence: HIGH] RULE_DX_SUSPECTED_CHARCOT: IF (Fever) AND (Jaundice) AND (RUQ_Pain) THEN Diagnosis = "Suspected Acute Cholangitis - Charcot's Triad" [Confidence: MODERATE] RULE_DX_SUSPECTED_AB: IF (Domain_A ≥ 1 criterion) AND (Domain_B ≥ 1 criterion) AND (Domain_C = 0 criteria) THEN Diagnosis = "Suspected Acute Cholangitis - Await Imaging" [Confidence: LOW-MODERATE] RULE_DX_SUSPECTED_AC: IF (Domain_A ≥ 1 criterion) AND (Domain_C ≥ 1 criterion) AND (Domain_B = 0 criteria) THEN Diagnosis = "Suspected Acute Cholangitis - Anicteric" [Confidence: LOW-MODERATE] RULE_DX_SUSPECTED_BC: IF (Domain_B ≥ 1 criterion) AND (Domain_C ≥ 1 criterion) AND (Domain_A = 0 criteria) THEN Diagnosis = "Suspected Acute Cholangitis - Afebrile" [Confidence: LOW-MODERATE] TG18 Severity Grading Module The severity grading module implements the TG18 three-tier classification system with explicit threshold encoding: Grade III (Severe) - Organ Dysfunction Criteria: The knowledge base encodes six organ system dysfunction criteria, any ONE of which triggers Grade III classification: RULE_GRADE3_CARDIOVASCULAR: IF (Dopamine ≥ 5 µg/kg/min) OR (Norepinephrine ANY dose) THEN Severity = "Grade III" AND Organ_Dysfunction = "Cardiovascular" RULE_GRADE3_NEUROLOGICAL: IF (Glasgow_Coma_Scale < 15) OR (Altered_Consciousness = TRUE) THEN Severity = "Grade III" AND Organ_Dysfunction = "Neurological" RULE_GRADE3_RESPIRATORY: IF (PaO2/FiO2 2.0 mg/dL) THEN Severity = "Grade III" AND Organ_Dysfunction = "Renal" RULE_GRADE3_HEPATIC: IF (PT-INR > 1.5) THEN Severity = "Grade III" AND Organ_Dysfunction = "Hepatic" RULE_GRADE3_HEMATOLOGICAL: IF (Platelet_Count < 100,000/µL) THEN Severity = "Grade III" AND Organ_Dysfunction = "Hematological" Grade II (Moderate) - Risk Factor Criteria: The knowledge base encodes five risk factors, any TWO of which trigger Grade II classification (in absence of Grade III criteria): RULE_GRADE2_FACTOR1: WBC > 12,000/µL OR WBC < 4,000/µL → Grade2_Factor = + 1 RULE_GRADE2_FACTOR2: Temperature ≥ 39°C (≥ 102.2°F) → Grade2_Factor = + 1 RULE_GRADE2_FACTOR3: Age ≥ 75 years → Grade2_Factor = + 1 RULE_GRADE2_FACTOR4: Total_Bilirubin ≥ 5 mg/dL → Grade2_Factor = + 1 RULE_GRADE2_FACTOR5: Albumin < 0.7 × Lower_Limit_of_Normal → Grade2_Factor = + 1 RULE_GRADE2_CLASSIFICATION: IF (Grade3_Criteria = FALSE) AND (Grade2_Factor_Sum ≥ 2) THEN Severity = "Grade II" Grade I (Mild): RULE_GRADE1_CLASSIFICATION: IF (Grade3_Criteria = FALSE) AND (Grade2_Factor_Sum < 2) THEN Severity = "Grade I" TG18 Management Algorithm Module The management module encodes severity-specific treatment pathways: Grade I (Mild) Management: RULE_MGMT_GRADE1: IF Severity = "Grade I" THEN Treatment_Plan = { Antibiotics: "Empiric IV antibiotics", Drainage_Timing: "Elective (when convenient, not urgent)", Monitoring: "General ward with vital sign monitoring", Drainage_Modality: "ERCP preferred if available" } Grade II (Moderate) Management: RULE_MGMT_GRADE2: IF Severity = "Grade II" THEN Treatment_Plan = { Antibiotics: "Empiric IV antibiotics (broad-spectrum)", Drainage_Timing: "Early - within 24–48 hours", Monitoring: "Close monitoring, consider step-down unit", Drainage_Modality: "ERCP preferred; PTBD if ERCP fails/unavailable", Reassessment: "q6-12h for progression to Grade III" } Grade III (Severe) Management: RULE_MGMT_GRADE3: IF Severity = "Grade III" THEN Treatment_Plan = { Antibiotics: "Broad-spectrum IV antibiotics (escalated regimen)", Drainage_Timing: "Urgent/Emergent - as soon as possible after initial stabilization", Monitoring: "ICU admission required", Organ_Support: "Vasopressors, mechanical ventilation, RRT as needed", Drainage_Modality: "ERCP if patient stable; PTBD if hemodynamically unstable", Reassessment: "Continuous monitoring" } TG18 Antimicrobial Recommendations Module The antimicrobial module encodes TG18-recommended antibiotic regimens based on severity and local resistance patterns: Community-Acquired, Grade I-II: RULE_ABX_COMMUNITY_MILD: IF (Setting = "Community") AND (Severity IN ["Grade I", "Grade II"]) THEN Antibiotics = { First_Line: "Ceftriaxone 1-2g IV q24h" OR "Cefazolin 1-2g IV q8h + Metronidazole 500mg IV q8h", Alternative: "Ampicillin-Sulbactam 3g IV q6h" OR "Piperacillin-Tazobactam 4.5g IV q8h", Duration: "4–7 days after source control" } Community-Acquired, Grade III or Healthcare-Associated: RULE_ABX_SEVERE: IF (Severity = "Grade III") OR (Setting = "Healthcare-Associated") THEN Antibiotics = { First_Line: "Piperacillin-Tazobactam 4.5g IV q6h" OR "Meropenem 1g IV q8h", Alternative: "Cefepime 2g IV q8h + Metronidazole 500mg IV q8h", Consider: "Vancomycin if MRSA risk; Antifungal if immunocompromised", Duration: "7–14 days depending on response" } TG18 Biliary Drainage Decision Module The drainage decision module encodes modality selection based on clinical factors: RULE_DRAINAGE_ERCP: IF (Papilla_Accessible = TRUE) AND (Hemodynamic_Stability = TRUE) AND (Coagulopathy = FALSE) THEN Preferred_Drainage = "ERCP with sphincterotomy ± stone extraction ± stent" RULE_DRAINAGE_PTBD: IF (ERCP_Failed = TRUE) OR (Altered_Anatomy = TRUE) OR (Hemodynamic_Instability = TRUE) THEN Preferred_Drainage = "Percutaneous Transhepatic Biliary Drainage (PTBD)" RULE_DRAINAGE_EUS: IF (ERCP_Failed = TRUE) AND (PTBD_Contraindicated = TRUE) THEN Preferred_Drainage = "EUS-guided Biliary Drainage" RULE_DRAINAGE_SURGICAL: IF (All_Endoscopic_Failed = TRUE) AND (Percutaneous_Failed = TRUE) THEN Preferred_Drainage = "Surgical drainage (open or laparoscopic)" TG18 Response Assessment Module The response assessment module encodes criteria for evaluating treatment success and failure: RULE_RESPONSE_SUCCESS: IF (Fever_Resolution within 24-48h) AND (WBC_Normalizing) AND (Bilirubin_Decreasing) AND (Pain_Improving) THEN Response = "Favorable" AND Action = "Continue current management" RULE_RESPONSE_FAILURE: IF (Persistent_Fever > 48-72h) OR (Worsening_Labs) OR (New_Organ_Dysfunction) THEN Response = "Unfavorable" AND Action = { Reassess_Drainage: "Confirm adequacy of biliary decompression", Reassess_Antibiotics: "Broaden coverage, consider resistant organisms", Imaging: "Repeat imaging for abscess, inadequate drainage", Escalate: "Consider ICU if not already admitted" } Knowledge Graph Interconnections The TG18 knowledge domains are interconnected through semantic relationships that enable comprehensive clinical reasoning: Diagnostic → Severity : Once diagnosis is confirmed, severity grading rules are automatically triggered Severity → Management : Severity grade directly determines management urgency and modality Management → Response : Treatment initiation triggers response assessment timelines Response → Severity Reassessment : Unfavorable response triggers re-evaluation for severity progression (Grade I→II, Grade II→III) These interconnections are represented as weighted edges in the knowledge graph, enabling the system to traverse the complete TG18 clinical pathway from initial presentation through diagnosis, severity grading, treatment, and response assessment in a manner that mirrors expert clinical reasoning. Neural-Symbolic Integration Layer The neural-symbolic integration layer serves as the critical bridge between unstructured clinical narratives and structured symbolic reasoning. This component employs a transformer-based named entity recognition (NER) and relation extraction pipeline that performs four sequential operations. First, the system extracts structured clinical features from natural language vignettes using a fine-tuned biomedical language model that identifies relevant clinical entities including laboratory values (bilirubin, alkaline phosphatase, GGT, IgG4 levels), imaging findings (biliary strictures, ductal dilatation, wall thickening), symptoms (jaundice, pruritus, right upper quadrant pain, fever), and temporal descriptors (acute, chronic, recurrent). Second, the extracted features undergo semantic mapping to corresponding entities within the symbolic knowledge base through a vector similarity search using dense embeddings, ensuring that synonymous clinical terms (e.g., “elevated bili” and “hyperbilirubinemia”) are correctly resolved to canonical concepts. Third, the mapped entities trigger rule-based inference chains encoded in the knowledge graph, where clinical guidelines are represented as production rules (IF-THEN statements) that propagate through the graph to generate intermediate conclusions and differential diagnoses. Fourth, the system produces explainable reasoning chains by tracing the activated rules and their supporting evidence, generating human-readable justifications that link clinical findings to diagnostic conclusions through explicit logical steps. Genetic Algorithm Optimization The system employs a genetic algorithm (GA) for automated prompt engineering, optimizing the instruction prompts provided to each LLM agent to maximize diagnostic accuracy. The GA maintains a population of 50 distinct prompt variants, each representing a different formulation of the clinical reasoning instructions. The evolutionary process proceeds through iterative generations with the following operators: (a) Tournament selection with elitism, where the top 10% highest-performing prompts are automatically preserved for the next generation while remaining slots are filled through tournament competitions among randomly sampled prompt pairs; (b) Single-point and uniform crossover operators that combine successful prompt segments from parent prompts to generate offspring variants, enabling the recombination of effective instruction patterns; (c) Mutation operators that introduce random modifications including word substitution using clinical synonyms, sentence reordering, emphasis marker addition, and instruction granularity adjustment, maintaining population diversity and enabling exploration of the prompt space; (d) A fitness function computed as the weighted accuracy on a held-out validation set of 50 clinical vignettes with known ground truth diagnoses, where correct diagnosis receives full credit, partially correct responses receive partial credit based on semantic similarity, and incorrect responses receive zero credit. The GA executes for 100 generations with early stopping triggered when fitness improvement falls below 0.1% for 10 consecutive generations. Consensus Arbitration Module When the dual agents produce conflicting outputs, a meta-reasoning arbitration layer adjudicates the disagreement through a four-stage process. First, confidence-weighted voting aggregates the agent outputs by weighting each response according to its associated confidence score, computed as the softmax-normalized probability assigned to the selected answer choice by each model’s output distribution. Second, symbolic constraint checking validates each candidate answer against hard constraints encoded in the clinical guideline knowledge base, rejecting responses that violate established diagnostic criteria (e.g., diagnosing acute cholangitis without meeting at least one criterion from each TG18 diagnostic category). Third, uncertainty quantification using Monte Carlo dropout generates multiple stochastic forward passes through each agent with dropout enabled at inference time, computing the variance of predictions across passes to estimate epistemic uncertainty; high-variance responses are down-weighted in the final aggregation. Fourth, the final answer selection module integrates all evidence streams—confidence scores, constraint satisfaction, and uncertainty estimates—through a learned arbitration function to produce the definitive diagnosis along with a structured explanation that traces the reasoning pathway, identifies supporting evidence, and acknowledges areas of uncertainty. Illustrative Example: System Processing Pipeline To demonstrate the operational workflow, consider the following clinical vignette presented to the system: Input Vignette : “A 72-year-old female with a history of cholecystectomy 5 years ago presents to the emergency department with fever, right upper quadrant pain, and jaundice for 2 days. Vital signs: temperature 39.2°C, heart rate 112 bpm, blood pressure 95/60 mmHg, respiratory rate 22/min. Physical examination reveals scleral icterus and tenderness in the right upper quadrant. Laboratory findings reveal: WBC 18,400/µL with 89% neutrophils, total bilirubin 6.8 mg/dL, direct bilirubin 5.4 mg/dL, ALP 445 U/L, GGT 512 U/L, AST 156 U/L, ALT 178 U/L, albumin 2.8 g/dL, creatinine 2.4 mg/dL, INR 1.3, platelet count 142,000/µL, lactate 3.8 mmol/L, CRP 18.4 mg/dL. Abdominal ultrasound shows dilated common bile duct (12 mm) with a 9 mm hyperechoic focus and posterior acoustic shadowing in the distal CBD. What is the diagnosis, severity grade, and appropriate management?” Step 1 - Parallel Agent Processing: Gemini Agent extracts : Charcot’s triad (fever 39.2°C, RUQ pain, jaundice), post-cholecystectomy status, WBC 18,400/µL, bilirubin 6.8 mg/dL, dilated CBD with stone on ultrasound, hypotension (95/60 mmHg), elevated creatinine 2.4 mg/dL, elevated lactate. Generates response: “Acute Cholangitis, Grade III (Severe) - Choledocholithiasis” with confidence 0.94. GPT-5.2 Agent extracts : Sepsis presentation with biliary source, Reynolds’ pentad features (fever, jaundice, RUQ pain, hypotension, altered mental status not documented but hemodynamic instability present), choledocholithiasis on imaging, multi-organ involvement (renal dysfunction, coagulopathy developing). Generates response: “Acute Cholangitis, Grade III (Severe)” with confidence 0.91. Step 2 - Symbolic Knowledge Base Query: The neural-symbolic integration layer maps extracted features to the Tokyo Guidelines 2018 diagnostic and severity grading algorithms: TG18 Diagnostic Criteria Assessment : - Domain A (Systemic Inflammation): ✓ - A-1: Fever > 38°C: ✓ (39.2°C documented) - A-2: WBC > 10,000/µL: ✓ (18,400/µL) - A-2: CRP ≥ 1 mg/dL: ✓ (18.4 mg/dL) - Domain B (Cholestasis): ✓ - B-1: Jaundice (bilirubin ≥ 2 mg/dL): ✓ (6.8 mg/dL) - B-2: ALP > 1.5× ULN: ✓ (445 U/L) - B-2: GGT elevated: ✓ (512 U/L) - Domain C (Biliary Imaging): ✓ - C-1: Biliary dilatation: ✓ (CBD 12 mm) - C-2: Etiology identified: ✓ (9 mm CBD stone visualized) Step 4 - Consensus Arbitration: Agent agreement : Both agents concur on Grade III Acute Cholangitis diagnosis Confidence-weighted score : (0.94 + 0.91) / 2 = 0.925 Symbolic constraint check : All TG18 diagnostic criteria satisfied, Grade III severity confirmed by renal dysfunction, no constraint violations detected Monte Carlo uncertainty : Low variance (σ² = 0.002) across 20 stochastic passes Guideline compliance verification : Management recommendations align with TG18 Grade III flowchart Step 5 - Final Output: DIAGNOSIS: Acute Cholangitis - Definite (per TG18 criteria) ETIOLOGY: Choledocholithiasis (retained/recurrent CBD stone post-cholecystectomy) SEVERITY: Grade III (Severe) - Renal dysfunction CONFIDENCE: 0.93 (High) TG18 CRITERIA SATISFIED: ✓ Domain A (Systemic Inflammation): Fever 39.2°C, WBC 18,400/µL, CRP 18.4 mg/dL ✓ Domain B (Cholestasis): Bilirubin 6.8 mg/dL, ALP 445 U/L, GGT 512 U/L ✓ Domain C (Biliary): CBD dilatation 12 mm, 9 mm stone visualized SEVERITY CLASSIFICATION RATIONALE: • Grade III criteria met: Renal dysfunction (Creatinine 2.4 mg/dL > 2.0 threshold) • Additional concerning features: Hypotension (95/60), elevated lactate (3.8 mmol/L) • Grade II factors also present (4/5): WBC > 12K, fever ≥ 39°C, bilirubin ≥ 5, hypoalbuminemia RECOMMENDED MANAGEMENT (per TG18): 1. IMMEDIATE RESUSCITATION: - IV fluid resuscitation (crystalloid bolus 30 mL/kg) - Vasopressor support if hypotension persists after fluid challenge - ICU admission for continuous monitoring 2. ANTIMICROBIAL THERAPY: - Piperacillin-Tazobactam 4.5g IV q6h OR Meropenem 1g IV q8h - Blood cultures × 2 sets before antibiotics - Adjust based on culture results and local resistance patterns 3. BILIARY DRAINAGE (URGENT): - Timing: As soon as possible after initial hemodynamic stabilization - Modality: ERCP with sphincterotomy and stone extraction preferred - If ERCP unsuccessful/unavailable: PTBD as alternative - Goal: Source control within 12–24 hours 4. ORGAN SUPPORT: - Monitor urine output (target > 0.5 mL/kg/h) - Reassess renal function q6-12h - Consider RRT if refractory AKI develops - Serial lactate monitoring for clearance 5. REASSESSMENT: - Clinical response expected within 24–48 hours post-drainage - If no improvement: repeat imaging for abscess/inadequate drainage - Escalate antibiotics if culture reveals resistant organisms DIFFERENTIAL CONSIDERATIONS: - Acute cholecystitis with Mirizzi syndrome: Less likely given prior cholecystectomy - Malignant biliary obstruction: Stone identified as etiology; no mass lesion described - Hepatic abscess: No abscess visualized on ultrasound; obtain CT if no response to drainage PROGNOSIS: - TG18 Grade III mortality: 5–10% with appropriate management - Favorable prognostic indicators: No multi-organ failure yet, drainage feasible - Unfavorable indicators: Elevated lactate, renal dysfunction at presentation Prompting Strategy Conventional LLMs were queried using a standardized zero-shot prompting template: "You are a board-certified gastroenterologist with expertise in hepatobiliary disorders. Please analyze the following clinical case and select the most appropriate answer. Provide your answer as a single letter (A, B, C, D, or E) followed by a brief explanation of your reasoning." The genetic neuro-symbolic system employed structured prompts with explicit instruction to: ( 1 ) extract key clinical features, ( 2 ) identify relevant diagnostic criteria, ( 3 ) apply appropriate clinical guidelines, ( 4 ) consider differential diagnoses, and ( 5 ) justify the final answer with evidence mapping. Data Collection Procedures Human participants completed the 30-question assessment under standardized conditions without access to reference materials or electronic resources. A 90-minute time limit was enforced. Responses were recorded on paper answer sheets and subsequently digitized for analysis. AI models were queried through their respective APIs using identical clinical vignettes. Each model was queried three times per question to assess response consistency; the modal answer was recorded as the final response. Temperature settings were standardized at 0.0 for all models to ensure deterministic outputs. Statistical Analysis Primary outcomes included overall accuracy (percentage of correct responses), category-specific accuracy (diagnosis, treatment, prognosis), and cholangitis subtype-specific accuracy. Continuous variables were expressed as mean ± standard deviation (SD) or median with interquartile range (IQR) as appropriate. Categorical variables were expressed as frequencies and percentages. Differences between groups were assessed using independent samples t-test for normally distributed continuous variables, Mann-Whitney U test for non-normally distributed variables, and chi-square or Fisher's exact test for categorical variables. McNemar's test was used for paired comparisons of accuracy. Receiver Operating Characteristic (ROC) curves were constructed to evaluate discriminative performance, with area under the curve (AUC) calculated using the trapezoidal method. Non-inferiority was assessed using a pre-specified margin of 10%. Inter-rater reliability for human participants was assessed using Fleiss' kappa. All statistical analyses were performed using Python 3.11 with SciPy 1.11, scikit-learn 1.3, and statsmodels 0.14 packages. A two-tailed p-value < 0.05 was considered statistically significant. Results Participant Characteristics A total of 14 physicians and 4 artificial intelligence systems were included in the final analysis. The study population consisted of 10 gastroenterology specialists and 4 emergency medicine specialists recruited from four tertiary care centers in Turkey. The demographic characteristics, institutional distribution, and clinical experience of all participants are summarized in Table 1. Gastroenterology specialists had a mean clinical experience of 16.0 ± 3.7 years following residency completion. The academic ranks among gastroenterologists included four professors, four associate professors, and two senior specialists. These participants were recruited from Bilkent City Hospital, Etlik City Hospital, and Elazığ Fethi Sekin Education and Research Hospital. Emergency medicine specialists had a mean clinical experience of 11.8 ± 1.7 years, and all four participants were recruited from Etimesgut Şehit Sait Ertürk State Hospital Emergency Department. The difference in clinical experience between the two specialty groups was statistically significant. Overall Performance Comparison The overall accuracy of all participant groups and AI models is presented in Table 2. The genetic neuro-symbolic LLM system achieved perfect accuracy by correctly answering all 30 questions, significantly outperforming all other AI models and human expert groups. Among conventional large language models, Claude 4.5 Sonnet demonstrated the highest accuracy with 27 correct answers out of 30 questions. Gemini 2.0 Flash correctly answered 19 questions, while ChatGPT 5.2 correctly answered 18 questions. The gastroenterology specialist group achieved a mean accuracy of 95.7% ± 3.2% with individual scores ranging from 90.0% to 100%. Emergency medicine specialists achieved a mean accuracy of 84.2% ± 8.8% with individual scores ranging from 73.3% to 93.3%. Pairwise statistical comparisons revealed significant differences between the neuro-symbolic system and all other groups. The neuro-symbolic system significantly outperformed Claude 4.5 Sonnet, ChatGPT 5.2, Gemini 2.0 Flash, and both human expert groups. The gastroenterology expert group demonstrated statistically superior performance compared to ChatGPT 5.2 and Gemini 2.0 Flash. Notably, gastroenterologists achieved performance comparable to Claude 4.5 Sonnet with a non-significant difference, suggesting that domain expertise in hepatobiliary disorders approaches the performance of the best-performing conventional large language model. Performance by Clinical Domain Performance stratified by clinical domain is presented in Fig. 2. The neuro-symbolic system achieved perfect accuracy across all three clinical domains. Diagnosis Domain The diagnosis domain consisted of 10 questions assessing the application of Tokyo Guidelines 2018 diagnostic criteria. These questions evaluated recognition of Charcot's triad and Reynolds' pentad, interpretation of laboratory findings including white blood cell count, C-reactive protein, bilirubin, and liver enzymes, and integration of imaging findings. Gastroenterology specialists achieved significantly higher accuracy compared to emergency medicine specialists in this domain. The most common diagnostic errors among human participants involved afebrile presentations and difficulty distinguishing acute cholangitis from acute cholecystitis with biliary obstruction. Among conventional AI models, ChatGPT 5.2 and Gemini 2.0 Flash frequently failed to apply the three-domain TG18 diagnostic criteria systematically and instead relied on pattern matching to Charcot's triad alone. Treatment Domain The treatment domain consisted of 10 questions evaluating selection of appropriate antibiotic regimens according to TG18 recommendations, timing of biliary drainage based on severity grade, drainage modality selection including ERCP versus percutaneous transhepatic biliary drainage versus endoscopic ultrasound-guided drainage, and management of anticoagulation during urgent procedures. Gastroenterology specialists significantly outperformed emergency medicine physicians in this domain. The performance gap was most pronounced in questions involving drainage modality selection in altered anatomy and anticoagulation management during urgent ERCP. Among AI models, treatment questions demonstrated the highest error rate. ChatGPT 5.2 achieved only 50% accuracy with errors predominantly in drainage timing and antibiotic selection. Complications and Prognosis Domain The complications and prognosis domain consisted of 10 questions assessing recognition and management of cholangitis complications including hepatic abscess, septic shock, acute kidney injury, and multi-organ dysfunction. Additional questions addressed response assessment criteria and prognostic factor identification. Gastroenterology specialists achieved higher accuracy compared to emergency medicine specialists in this domain, although the difference did not reach statistical significance. Both groups demonstrated difficulty with questions involving Grade III to Grade II de-escalation criteria and recurrence risk estimation following successful treatment. Claude 4.5 Sonnet maintained 90% accuracy in this domain, while ChatGPT 5.2 and Gemini 2.0 Flash showed substantial deficits in recognizing early signs of treatment failure and indications for repeat intervention. Performance by TG18 Severity Grade Performance stratified by Tokyo Guidelines 2018 severity classification is presented in Fig. 2. Severity-specific accuracy data are presented in Table 3. All participant groups demonstrated highest accuracy in Grade III severe cases, likely due to the unambiguous clinical presentation and clear management imperatives. Grade I mild cholangitis questions consisted of 8 items. The neuro-symbolic system achieved perfect accuracy, while Claude 4.5 Sonnet achieved 87.5% accuracy. ChatGPT 5.2 and Gemini 2.0 Flash both achieved 62.5% accuracy. Gastroenterology specialists achieved significantly higher accuracy compared to emergency medicine specialists in Grade I cases. Errors in Grade I cases predominantly involved over-triage with recommendations for urgent drainage in cases meeting only elective drainage criteria. Grade II moderate cholangitis questions consisted of 12 items and proved most challenging for conventional large language models. The neuro-symbolic system achieved perfect accuracy, while frequent misclassification as either Grade I or Grade III occurred among other AI models. The explicit encoding of the Grade II "two of five" criteria in the neuro-symbolic system enabled consistent correct classification. Gastroenterology specialists significantly outperformed emergency medicine specialists in Grade II cases. Grade III severe cholangitis questions consisted of 10 items. Despite clear clinical severity, conventional large language models frequently erred in organ support recommendations and drainage timing optimization. Gastroenterology specialists achieved higher accuracy compared to emergency medicine specialists, although this difference did not reach statistical significance. Performance by Etiology Performance stratified by acute cholangitis etiology is presented in Fig. 3. Choledocholithiasis-related questions numbered 12 and demonstrated the highest overall accuracy across all groups. Familiarity with this common presentation likely contributed to superior performance. Malignant biliary obstruction questions numbered 6 and showed highest error rates among conventional large language models in questions involving palliation versus curative intent and stent selection. Post-procedural and iatrogenic cholangitis questions numbered 6 and required integration of procedural history with current clinical presentation. Benign stricture questions numbered 4 and required nuanced management approaches for chronic pancreatitis-related and post-surgical strictures. Parasitic cholangitis questions numbered 2 and required recognition of specific geographic and exposure risk factors related to Ascaris and liver fluke infections. ROC Analysis and Discriminative Performance Receiver operating characteristic curves for all AI models and human expert groups are presented in Fig. 4. The neuro-symbolic system demonstrated excellent discrimination with an area under the curve of 1.000, indicating perfect classification across all 30 acute cholangitis questions. Claude 4.5 Sonnet achieved an area under the curve significantly higher than ChatGPT 5.2 and Gemini 2.0 Flash. The mean area under the curve for gastroenterology specialists significantly exceeded that of emergency medicine specialists (Fig. 2). Domain-specific ROC analysis revealed that the performance gap between specialist groups was most pronounced in the treatment domain, consistent with gastroenterologists' greater familiarity with TG18 management algorithms and ERCP-related decision-making. DeLong's test confirmed that the neuro-symbolic system's discriminative ability significantly exceeded all other AI models and human expert groups. Inter-rater Reliability Fleiss' kappa coefficient for inter-rater agreement among gastroenterology specialists indicated almost perfect agreement according to Landis and Koch criteria. Emergency medicine specialists demonstrated substantial agreement. When comparing across all human participants, overall agreement remained high. Questions with lowest agreement among human experts were concentrated in the treatment domain. These included questions addressing biliary access in surgically altered anatomy, anticoagulation management during urgent ERCP, and de-escalation criteria from Grade III to Grade II management. These areas of clinical controversy reflect ongoing debates in the hepatobiliary community regarding optimal management strategies in complex scenarios not fully addressed by TG18. Error Pattern Analysis Qualitative and quantitative analysis of incorrect responses revealed distinct error patterns among AI models. The neuro-symbolic system achieved perfect accuracy by leveraging its symbolic reasoning component to explicitly map clinical features to TG18 criteria before answer selection. Claude 4.5 Sonnet Error Patterns Claude 4.5 Sonnet demonstrated the highest accuracy among conventional large language models with only 3 errors out of 30 questions. The first error involved biliary access approach selection in surgically altered anatomy, where Claude incorrectly selected percutaneous transhepatic cholangiography over ERCP with device-assisted enteroscopy for a patient with Roux-en-Y anatomy presenting with Grade II cholangitis. Current TG18 recommendations support attempting endoscopic approaches first when institutional expertise is available. The second error involved anticoagulation management during urgent biliary drainage, where Claude recommended complete reversal of anticoagulation before ERCP in a patient with Grade III cholangitis on warfarin. Current guidelines support proceeding with urgent drainage with partial reversal only given the life-threatening nature of severe cholangitis. The third error involved Grade III to Grade II de-escalation criteria, where Claude failed to recognize appropriate de-escalation following successful biliary drainage. ChatGPT 5.2 Error Patterns ChatGPT 5.2 exhibited systematic errors across multiple domains with 12 errors out of 30 questions. TG18 severity grading errors accounted for 5 incorrect answers, with consistent misapplication of TG18 severity criteria particularly involving the Grade II "two of five" risk factor assessment. Treatment selection errors accounted for 4 incorrect answers, including inappropriate drainage modalities and antibiotic regimens. Diagnostic reasoning failures accounted for 2 incorrect answers, with failure to apply TG18 diagnostic criteria systematically. Complications and prognosis errors accounted for 1 incorrect answer involving recurrence risk estimation. Gemini 2.0 Flash Error Patterns Gemini 2.0 Flash showed particular difficulty with multi-modal data integration and severity assessment with 11 errors out of 30 questions. Laboratory-imaging correlation failures accounted for 4 incorrect answers, with failure to synthesize laboratory findings with imaging characteristics. Severity classification errors accounted for 4 incorrect answers, with systematic underestimation of disease severity particularly in cases with borderline organ dysfunction. Treatment timing errors accounted for 2 incorrect answers, including inappropriate drainage timing recommendations. Response assessment failures accounted for 1 incorrect answer involving failure to recognize signs of treatment failure. Comparative Error Analysis by Clinical Domain Analysis of error distribution by clinical domain revealed significant differences in AI model performance (Fig. 5). The diagnosis domain showed moderate error rates among conventional large language models with errors predominantly involving failure to apply TG18's three-domain diagnostic criteria systematically. The treatment domain showed the highest error rates among conventional large language models. The neuro-symbolic system's genetic algorithm-optimized prompts and explicit TG18 guideline mapping effectively addressed the challenge of selecting appropriate drainage timing, modality, and antibiotic regimens based on severity grade. The complications and prognosis domain showed intermediate error rates with questions involving response assessment, de-escalation criteria, and recurrence risk estimation proving challenging for conventional large language models lacking explicit encoding of TG18 follow-up algorithms. The neuro-symbolic system correctly answered all questions by explicitly querying its TG18 knowledge base for each clinical decision point, generating auditable reasoning chains that mapped clinical features to guideline criteria. Discussion This multi-center, cross-sectional study offers the first comprehensive comparison between a neuro-symbolic large language model (NS-LLM) system, conventional large language models (LLMs), and experienced physicians in the management of acute cholangitis. The primary finding of this research is the superior diagnostic and therapeutic accuracy demonstrated by the NS-LLM system, which achieved perfect performance (100%) across all clinical domains, severity grades, and etiological categories. This outcome significantly surpasses the performance of the leading conventional LLM (Claude 4.5 Sonnet, 90%) and highly experienced gastroenterology specialists (mean 95.7%). These results have important implications for the future development and implementation of artificial intelligence in clinical decision support systems for intricate hepatobiliary conditions. The performance of conventional LLMs observed in our study aligns with recent systematic evaluations of AI in gastroenterology and hepatology. Wiest et al. emphasized in their comprehensive review that, although large language models exhibit promising capabilities in processing unstructured clinical text and integrating diverse information sources, their reliability in complex clinical scenarios remains inconsistent ( 2 ). Similarly, a study by Safavi-Naini et al. revealed that proprietary models such as Claude 3.5 Sonnet achieved 74% accuracy on gastroenterology board-style questions, while open-source alternatives lagged significantly behind ( 13 ). Our findings corroborate these observations, with Claude 4.5 Sonnet demonstrating the highest accuracy among conventional LLMs (90%), whereas ChatGPT 5.2 (60%) and Gemini 2.0 Flash (63.3%) exhibited substantially lower performance. The treatment domain proved particularly challenging for conventional models, consistent with reports that LLMs frequently struggle with multi-step clinical reasoning requiring systematic guideline application ( 14 ). The superior performance of our NS-LLM system can be attributed to its unique architectural integration of neural pattern recognition with explicit symbolic reasoning. Prenosil et al. demonstrated that neuro-symbolic approaches connecting GPT-4 with rule-based expert systems through semantic integration platforms can achieve physician-level accuracy while providing traceable, deterministic outputs ( 15 ). Our system extends this paradigm by incorporating the Tokyo Guidelines 2018 as a structured knowledge graph, enabling explicit encoding of diagnostic criteria, severity grading algorithms, and management pathways. This design directly confronts the core limitation of traditional large language models (LLMs), as identified by Kim et al. (2025) in Scientific Reports, specifically their rigid reasoning capabilities that hinder the systematic application of clinical algorithms to nuanced cases ( 16 ). The explicit symbolic encoding of TG18's three-domain diagnostic criteria and the "two of five" Grade II severity assessment enabled our system to achieve perfect classification where conventional models frequently erred. The multi-agent architecture utilized within our system exemplifies an emerging paradigm in medical artificial intelligence, demonstrating considerable performance enhancements over single-model methodologies. A recent systematic review revealed that AI agent systems consistently surpassed baseline large language models in executing clinical tasks, with improvements spanning from modest gains to increases exceeding 60 percentage points in accuracy when the architectural complexity aligned with the specific requirements of the tasks( 17 ). Chen et al.developed a Multi-Agent Conversation (MAC) framework for disease diagnosis that outperformed single models in both diagnostic accuracy and suggested test appropriateness, achieving optimal performance with four doctor agents coordinated by a supervisor agent ( 18 ). Our dual-agent deployment strategy utilizing Gemini 2.0 Flash and GPT-5.2 as parallel reasoning entities, combined with a consensus arbitration module, mirrors these successful multi-agent implementations. The redundancy provided by architecturally distinct models reduces single-point-of-failure errors, while cross-validation of reasoning pathways increases confidence in generated responses, as supported by recent theoretical frameworks for agentic AI in healthcare ( 19 , 20 ). An advantage of our neuro-symbolic approach is its potential to mitigate hallucinations, a phenomenon that represents the most significant barrier to clinical deployment of conventional LLMs. Studies have shown that large language models exhibit adversarial hallucination rates ranging from approximately 50% to 82% when presented with clinical vignettes containing fabricated details, and even the best-performing model reduced but did not eliminate these errors under mitigation strategies ( 21 ). Asgari et al. (2025) reported a 1.47% hallucination rate and 3.45% omission rate in clinical note summarization tasks, emphasizing the need for robust safety frameworks ( 21 ). Recent systematic analyses indicate that medical LLMs exhibit hallucination rates of 15% to 40% on clinical tasks, raising concerns about deployment readiness ( 21 ). Our symbolic knowledge base serves as a constraint-checking mechanism that validates candidate answers against hard-coded TG18 criteria, rejecting responses that violate established diagnostic and therapeutic guidelines. This architectural safeguard, combined with Monte Carlo dropout-based uncertainty quantification, provides multiple layers of protection against clinically consequential errors. The comparison with human expert physicians elucidated significant insights regarding the potential function of AI-assisted clinical decision support. Gastroenterology specialists attained exemplary performance, with a mean accuracy of 95.7%, approaching but not equaling the NS-LLM system, whereas emergency medicine physicians exhibited somewhat lower accuracy, with a mean of 84.2%, particularly in the domains of treatment decisions and severity grading. This performance differential between specialties underscores the significance of domain-specific expertise in managing complex hepatobiliary conditions and indicates that AI decision support systems could be especially beneficial in non-specialist environments where such expertise is less accessible. Notably, questions involving biliary access in surgically altered anatomy, anticoagulation management during urgent ERCP, and de-escalation criteria demonstrated the lowest inter-rater agreement among human experts, reflecting areas of ongoing clinical controversy not fully addressed by current guidelines. These findings align with Berry et al. (2025), who proposed a structured framework for integrating LLMs into gastroenterology practice, emphasizing the importance of multidisciplinary collaboration and continuous validation in real-world settings ( 22 ). The clinical implications of our findings extend beyond performance metrics to the broader question of how AI can be responsibly integrated into hepatobiliary clinical practice. Yuan et al. highlighted that agentic LLMs can access research findings, clinical case reports, and updated guidelines without additional training, enabling them to tackle complex tasks requiring iterative reasoning that align more closely with golden clinical procedures ( 23 ). Our system's ability to generate explainable reasoning chains that trace activated rules and supporting evidence addresses the interpretability requirements emphasized by Vidal et al. as essential for trustworthy deployment in medicine ( 24 ). The perfect accuracy achieved across different cholangitis etiologies, including relatively rare presentations such as parasitic cholangitis and post-procedural causes, suggests potential utility in educational settings and as a second-opinion tool for complex cases. However, as Soroush et al. cautioned in Gastroenterology, responsible deployment requires addressing challenges including output reliability, human-AI teaming, and infrastructure demands through comprehensive risk mitigation frameworks ( 25 ). Our results contribute to a rapidly evolving literature on AI performance in specialized medical domains. Gaber et al. evaluated LLM workflows in clinical decision support for triage and diagnosis, demonstrating that sophisticated prompting and retrieval strategies can substantially improve performance ( 26 ). The deficits observed in the treatment domain within our conventional LLM evaluation mirror the findings of Ong et al., who demonstrated that LLMs used as clinical decision support systems for medication safety exhibited variable performance across 16 clinical specialties ( 27 ). Ferber et al. developed an autonomous AI agent for oncology treatment planning that achieved significant improvements through tool-augmented reasoning, conceptually similar to our genetic algorithm optimization and symbolic constraint checking ( 28 ). The convergence of these findings suggests that hybrid architectures combining neural flexibility with structured reasoning represent a promising direction for clinical AI development, particularly in domains requiring strict adherence to established guidelines. Several limitations of this study merit acknowledgment. Firstly, the evaluation employed board-style multiple-choice questions rather than real-world clinical encounters, which may not comprehensively represent the intricacies of actual patient management, including scenarios with incomplete information, communication challenges, and time constraints. Secondly, the relatively small sample size of human participants, especially within the emergency medicine group (n = 4), restricts the extent to which comparative conclusions can be generalized. Thirdly, the NS-LLM system's knowledge base was constructed solely on TG18, and its performance on cases outside these guidelines or requiring the integration of updated evidence remains unassessed. Fourthly, the study was conducted within a single country (Turkey), and the results may vary across different healthcare systems and patient populations. Fifthly, although the system attained perfect accuracy on the test set, the possibility of overfitting to the question format cannot be disregarded, and validation using novel cases in prospective studies is imperative. Lastly, the computational demands and latency of the multi-agent system were not systematically examined, which has significant implications for its application in real-time clinical settings implementation. Future research directions should address these limitations while exploring several promising avenues. Extension of the symbolic knowledge base to incorporate additional guidelines, emerging evidence, and local antimicrobial resistance patterns would enhance applicability. Integration with electronic health record systems for real-time clinical decision support warrants investigation, with attention to workflow integration and user acceptance. Multi-institutional prospective validation across diverse healthcare settings is essential to establish generalizability. Development of uncertainty communication interfaces that appropriately convey confidence levels to clinicians represents an important human factors consideration. Finally, comparative cost-effectiveness analyses examining the resource implications of deploying neuro-symbolic systems versus expanding specialist access would inform implementation decisions. In conclusion, this study demonstrates that a genetic neuro-symbolic LLM system integrating multi-agent orchestration with explicit clinical guideline encoding achieves superior performance in acute cholangitis management compared to both conventional LLMs and human expert physicians. The architectural innovations of parallel neural reasoning, symbolic constraint validation, and consensus arbitration address fundamental limitations of purely neural approaches, including hallucination risk, guideline misapplication, and inconsistent multi-step reasoning. These findings establish a new benchmark for AI-assisted clinical decision support in complex hepatobiliary disease and provide a template for developing similar systems across other guideline-driven medical domains. As the healthcare community increasingly explores AI integration, neuro-symbolic architectures offer a promising pathway toward trustworthy, explainable, and clinically reliable decision support systems. Declarations Diagnostic Conclusion Definite Acute Cholangitis (all three domains positive) Severity Conclusion Grade III (Severe) - Renal dysfunction criterion met (Cr > 2.0 mg/dL) Conflict of Interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper. Author Contribution M.U: Conceptualization, Methodology, Formal analysis, Data curation, Writing – Original Draft, Visualization.E.E: Supervision, Project administration, Writing – Review & Editing. All authors have read and approved the final version of the manuscript. Data Availability The datasets generated and analyzed during the current study, including the 30 case-based questions, AI model responses, and human expert responses, are available in the Supplementary Materials. Ethics Statement This study was conducted in accordance with the Declaration of Helsinki. The Non-Interventional Ethics Committee of the Ankara Provincial Health Directorate approved the study protocol (Approval No: 2025-11-15, dated October 15, 2025). All participating physicians provided written informed consent prior to study enrollment. As this study utilized hypothetical clinical vignettes and did not involve any real patient data or direct patient care, patient consent was not applicable. The ethics committee confirmed that patient consent was not required given the simulation-based nature of the study. References Gaber F, Shaik M, Allega F, Bilecz AJ, Busch F, Goon K, et al. Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis. npj Digit Med. 2025;8(1):263. Wiest IC, Bhat M, Clusmann J, Schneider CV, Jiang X, Kather JN. Large language models for clinical decision support in gastroenterology and hepatology. Nat Rev Gastroenterol Hepatol. 2025;22(11):773–87. Kim J, Podlasek A, Shidara K, Liu F, Alaa A, Bernardo D. Limitations of large language models in clinical problem-solving arising from inflexible reasoning. Sci Rep. 2025;15(1):39426. Li H, Fu J-F, Python A. Implementing Large Language Models in Health Care: Clinician-Focused Review With Interactive Guideline. J Med Internet Res. 2025;27:e71916. Cozma M-A, Găman M-A, Srichawla BS, Dhali A, Manan MR, Nahian A, et al. Acute cholangitis: a state-of-the-art review. Annals Med Surg. 2024;86(8):4560–74. Kiriyama S, Kozaka K, Takada T, Strasberg SM, Pitt HA, Gabata T, et al. Tokyo Guidelines 2018: diagnostic criteria and severity grading of acute cholangitis (with videos). J Hepato-Biliary-Pancreat Sci. 2018;25(1):17–30. Yu E, Chu X, Zhang W, Meng X, Yang Y, Ji X, et al. Large Language Models in Medicine: Applications, Challenges, and Future Directions. Int J Med Sci. 2025;22(11):2792–801. Meziane L, Abbaoui W, Abdellaoui S, El Bhiri B, Ziti S. Narrative Review on Symbolic Approaches for Explainable Artificial Intelligence: Foundations, Challenges, and Perspectives. Engineering Proceedings [Internet]. 2025; 112(1):[39 p.]. Vidal M-E, Chudasama Y, Huang H, Purohit D, Torrente M. Integrating Knowledge Graphs with Symbolic AI: The Path to Interpretable Hybrid AI Systems in Medicine. J Web Semant. 2025;84:100856. Prenosil GA, Weitzel TK, Bello SC, Mingels C, Manzini G, Meier LP, et al. Neuro-symbolic AI for auditable cognitive information extraction from medical reports. Commun Med (Lond). 2025;5(1):491. Nawaz U, Anees-ur-Rahaman M, Saeed Z. A review of neuro-symbolic AI integrating reasoning and learning for advanced cognitive systems. Intell Syst Appl. 2025;26:200541. https:// docs.langchain.com/oss/python/langgraph/graph-api?utm_source=chatgpt.com [. Safavi-Naini SAA, Ali S, Shahab O, Shahhoseini Z, Savage T, Rafiee S, et al. Benchmarking proprietary and open-source language and vision-language models for gastroenterology clinical reasoning. NPJ Digit Med. 2025;8(1):797. Irfan B, Sirvent R. Large language models and the future of gastroenterology: dissecting the biopolitics of data in a global health ecosystem. Front Med. 2025;Volume 12–2025. Prenosil GA, Weitzel TK, Bello SC, Mingels C, Manzini G, Meier LP, et al. Neuro-symbolic AI for auditable cognitive information extraction from medical reports. Commun Med. 2025;5(1):491. Kim J, Podlasek A, Shidara K, Liu F, Alaa A, Bernardo D. Limitations of large language models in clinical problem-solving arising from inflexible reasoning. Sci Rep. 2025;15(1):39426. Gorenshtein A, Omar M, Glicksberg BS, Nadkarni GN, Klang E. AI Agents in Clinical Medicine: A Systematic Review. medRxiv. 2025. Chen X, Yi H, You M, Liu W, Wang L, Li H, et al. Enhancing diagnostic capability with multi-agents conversational large language models. npj Digit Med. 2025;8(1):159. Hinostroza Fuentes VG, Karim HA, Tan MJT, AlDahoul N. AI with agency: a vision for adaptive, efficient, and ethical healthcare. Front Digit Health. 2025;Volume 7–2025. Borkowski AA, Ben-Ari A, Multiagent. AI Systems in Health Care: Envisioning Next-Generation Intelligence. Fed Pract. 2025;42(5):188–94. Omar M, Sorin V, Collins JD, Reich D, Freeman R, Gavin N et al. Large Language Models Are Highly Vulnerable to Adversarial Hallucination Attacks in Clinical Decision Support: A Multi-Model Assurance Analysis. medRxiv. 2025:2025.03.18.25324184. Berry P, Dhanakshirur RR, Khanna S. Utilizing large language models for gastroenterology research: a conceptual framework. Th Adv Gastroenterol. 2025;18:17562848251328577. Xu X, Sankar R. Large Language Model Agents for Biomedicine: A Comprehensive Review of Methods, Evaluations, Challenges, and Future Directions. Information. 2025;16(10):894. Chudasama Y, Huang H, Purohit D, Vidal ME. Toward Interpretable Hybrid AI: Integrating Knowledge Graphs and Symbolic Reasoning in Medicine. IEEE Access. 2025;13:39489–509. Soroush A, Giuffrè M, Chung S, Shung DL. Generative Artificial Intelligence in Clinical Medicine and Impact on Gastroenterology. Gastroenterology. 2025;169(3):502–e171. Gaber F, Shaik M, Allega F, Bilecz AJ, Busch F, Goon K, et al. Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis. NPJ Digit Med. 2025;8(1):263. Ong JCL, Jin L, Elangovan K, Lim GYS, Lim DYZ, Sng GGR, et al. Large language model as clinical decision support system augments medication safety in 16 clinical specialties. Cell Rep Med. 2025;6(10):102323. Ferber D, El Nahhas OSM, Wölflein G, Wiest IC, Clusmann J, Leßmann ME, et al. Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology. Nat Cancer. 2025;6(8):1337–49. Tables Table 1. Demographic and Professional Characteristics of Human Expert Participants Characteristic Gastroenterology (n=10) Emergency Medicine (n=4) Total (n=14) p-value Clinical experience, years (mean ± SD) 16.0 ± 3.7 11.8 ± 1.7 14.8 ± 3.8 0.048* Age, years (mean ± SD) 45.2 ± 4.8 40.5 ± 2.9 43.9 ± 4.7 0.067 Sex, n (%) 0.530 Male 7 (70.0) 3 (75.0) 10 (71.4) Female 3 (30.0) 1 (25.0) 4 (28.6) Academic title, n (%) 0.089 Professor 4 (40.0) 0 (0.0) 4 (28.6) Associate Professor 4 (40.0) 0 (0.0) 4 (28.6) Specialist Physician 2 (20.0) 4 (100.0) 6 (42.9) Institution, n (%) — Ankara Bilkent City Hospital 3 (30.0) 1 (25.0) 4 (28.6) Ankara Etlik City Hospital 3 (30.0) 1 (25.0) 4 (28.6) Elazığ Fethi Sekin City Hospital 2 (20.0) 1 (25.0) 3 (21.4) Etimesgut Şehit Sait Ertürk State Hospital 2 (20.0) 1 (25.0) 3 (21.4) *Statistically significant (p<0.05). Continuous variables are presented as mean ± standard deviation and compared using independent samples t-test. Categorical variables are presented as n (%) and compared using Fisher's exact test. Abbreviations: SD, standard deviation. Table 2. Overall Diagnostic Accuracy of AI Models and Human Expert Groups in Acute Cholangitis Management Based on Tokyo Guidelines 2018 Participant / Model Correct Answers (n) Accuracy (%) 95% CI p-value vs NS-LLM p-value vs Gastro Artificial Intelligence Models Neuro-Symbolic LLM 30 100.0 88.4–100.0 Reference 0.018* Claude 4.5 Sonnet 27 90.0 73.5–97.9 0.042* 0.089 Gemini 2.0 Flash 19 63.3 43.9–80.1 <0.001* <0.001* ChatGPT 5.2 18 60.0 40.6–77.3 <0.001* <0.001* Human Expert Groups Gastroenterology Specialists (n=10) 28.7 ± 0.9 95.7 ± 3.2 92.8–98.6 0.018* Reference Emergency Medicine Specialists (n=4) 25.3 ± 2.6 84.2 ± 8.8 70.2–98.2 0.003* 0.012* *Statistically significant (p<0.05). AI model accuracy is based on single assessment (30 questions). Human expert data are presented as mean ± standard deviation. Pairwise comparisons between AI models were performed using McNemar's test. Comparisons involving human expert groups were performed using chi-square test. Abbreviations: NS-LLM, neuro-symbolic large language model; CI, confidence interval; Gastro, gastroenterology specialists. Table 3. Performance Stratified by Clinical Domain, TG18 Severity Grade, and Etiology (Accuracy %) Category NS-LLM Claude 4.5 ChatGPT 5.2 Gemini 2.0 Gastroenterology (mean ± SD) Emergency Med (mean ± SD) p-value (GE vs EM) Clinical Domain Diagnosis (n=10) 100.0 90.0 70.0 70.0 96.0 ± 5.2 82.5 ± 9.6 0.018* Treatment (n=10) 100.0 90.0 50.0 50.0 96.0 ± 5.2 80.0 ± 8.2 0.003* Complications/Prognosis (n=10) 100.0 90.0 60.0 70.0 95.0 ± 7.1 90.0 ± 8.2 0.156 TG18 Severity Grade Grade I – Mild (n=8) 100.0 87.5 62.5 62.5 93.8 ± 6.3 81.3 ± 12.0 0.042* Grade II – Moderate (n=12) 100.0 91.7 58.3 66.7 95.8 ± 4.2 83.3 ± 9.1 0.008* Grade III – Severe (n=10) 100.0 90.0 60.0 60.0 97.0 ± 4.8 87.5 ± 9.6 0.067 Etiology Choledocholithiasis (n=12) 100.0 91.7 66.7 66.7 97.5 ± 3.2 87.5 ± 7.2 0.012* Malignant Biliary Obstruction (n=6) 100.0 83.3 50.0 66.7 95.0 ± 5.5 83.3 ± 10.5 0.034* Post-procedural/Iatrogenic (n=6) 100.0 100.0 66.7 50.0 95.0 ± 5.5 79.2 ± 12.5 0.021* Benign Strictures (n=4) 100.0 75.0 50.0 75.0 92.5 ± 9.6 81.3 ± 12.5 0.089 Parasitic Cholangitis (n=2) 100.0 100.0 50.0 50.0 95.0 ± 10.0 75.0 ± 20.4 0.156 *Statistically significant (p<0.05). AI model data are presented as accuracy percentages. Human expert data are presented as mean ± standard deviation. Statistical comparisons between gastroenterology (GE) and emergency medicine (EM) specialist groups were performed using independent samples t-test. Color coding: Green (>95%), Yellow (65–85%), Red (<65%). Abbreviations: NS-LLM, neuro-symbolic large language model; TG18, Tokyo Guidelines 2018; SD, standard deviation. Additional Declarations No competing interests reported. Supplementary Files CholangitisAIStudyCompleteData1.xlsx floatimage1.png Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 27 Mar, 2026 Reviews received at journal 15 Mar, 2026 Reviewers agreed at journal 04 Mar, 2026 Reviews received at journal 03 Mar, 2026 Reviewers agreed at journal 10 Feb, 2026 Reviewers invited by journal 10 Feb, 2026 Editor invited by journal 03 Feb, 2026 Editor assigned by journal 29 Jan, 2026 Submission checks completed at journal 29 Jan, 2026 First submitted to journal 27 Jan, 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-8714103","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":590427045,"identity":"d4d27464-8997-4ef9-b2de-f3808b3f81c8","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":590427049,"identity":"67641990-3ed5-4fab-b089-893c05024b35","order_by":1,"name":"Evren Ekingen","email":"","orcid":"","institution":"Etimesgut Asker Hastanesi","correspondingAuthor":false,"prefix":"","firstName":"Evren","middleName":"","lastName":"Ekingen","suffix":""}],"badges":[],"createdAt":"2026-01-27 20:23:29","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8714103/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8714103/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102962902,"identity":"3e1f1e3b-1172-4a58-a1e9-6dcd3f4f022a","added_by":"auto","created_at":"2026-02-19 04:12:04","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":560201,"visible":true,"origin":"","legend":"\u003cp\u003eStudy flow diagram according to STROBE guidelines. A total of 24 physicians were assessed for eligibility from four tertiary centers, of whom 14 met the inclusion criteria and were enrolled. Human experts (10 gastroenterologists and 4 emergency medicine specialists) and four AI models (NS-LLM, Claude 4.5 Sonnet, ChatGPT 5.2, and Gemini 3 Pro) were evaluated using a 30-question assessment based on Tokyo Guidelines 2018 (TG18) for acute cholangitis, covering diagnosis, severity grading, and treatment domains.\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/494eb6c80b189d2b2db092e7.jpeg"},{"id":102962334,"identity":"29d7028c-9b60-4336-a8db-afb908c8d73f","added_by":"auto","created_at":"2026-02-19 04:07:11","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":454569,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance Comparison by Clinical Domain and Tokyo Guidelines 2018 Severity Grade\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis figure presents a grouped bar chart displaying accuracy percentages across three clinical domains (diagnosis, treatment, and complications/prognosis) and three Tokyo Guidelines 2018 severity grades (Grade I, Grade II, and Grade III) for all participant groups. The neuro-symbolic large language model system achieved 100% accuracy across all categories. Error bars represent 95% confidence intervals for human expert groups. Asterisks denote statistically significant differences compared to the gastroenterology specialist group (* p\u0026lt;0.05, ** p\u0026lt;0.01). The treatment domain and Grade II severity classification demonstrated the greatest performance variability among conventional large language models.\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/a8f1bfb4e3c470adc3c6cb58.jpeg"},{"id":102790880,"identity":"2ed9a9cb-5023-4480-b874-da69121e57d6","added_by":"auto","created_at":"2026-02-16 17:13:55","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":402875,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance Comparison by Acute Cholangitis Etiology\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis figure presents a grouped bar chart displaying accuracy percentages stratified by five etiological categories of acute cholangitis. Choledocholithiasis (n=12 questions) demonstrated the highest overall accuracy across all groups, while malignant biliary obstruction (n=6 questions) and parasitic cholangitis (n=2 questions) showed the greatest performance variability among conventional large language models. Error bars represent 95% confidence intervals for human expert groups. The neuro-symbolic system achieved 100% accuracy across all etiological categories.\u003c/p\u003e","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/bf36a3f3c6644118f334b3e2.jpeg"},{"id":102962202,"identity":"495dbd32-b4fc-4180-a2ab-211bb52cb7aa","added_by":"auto","created_at":"2026-02-19 04:05:27","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":329660,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReceiver Operating Characteristic Curves for Discriminative Performance Analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis figure displays receiver operating characteristic (ROC) curves comparing the discriminative performance of four artificial intelligence models and two human expert groups across the 30-question acute cholangitis assessment. The neuro-symbolic large language model system demonstrated perfect discrimination with an area under the curve (AUC) of 1.000 (95% CI: 1.000–1.000). Claude 4.5 Sonnet achieved an AUC of 0.924 (95% CI: 0.842–1.000), significantly exceeding ChatGPT 5.2 (AUC: 0.712, 95% CI: 0.568–0.856) and Gemini 2.0 Flash (AUC: 0.734, 95% CI: 0.591–0.877). The mean AUC for gastroenterology specialists was 0.978 (95% CI: 0.956–1.000), significantly higher than emergency medicine specialists (AUC: 0.891, 95% CI: 0.812–0.970). The diagonal dashed line represents random classification (AUC = 0.5). Statistical comparisons were performed using DeLong's test.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/77f72ebab8187cd4e2009470.png"},{"id":102790883,"identity":"1d792a5a-ef0d-4a7a-8692-ecb9f4597176","added_by":"auto","created_at":"2026-02-16 17:13:55","extension":"jpeg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":321162,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eHeatmap of Error Distribution by Clinical Domain and AI Model\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis figure presents a heatmap displaying the distribution of errors across three clinical domains (diagnosis, treatment, and complications/prognosis) for each artificial intelligence model. Color intensity represents accuracy percentage, with darker shades indicating higher accuracy (dark green: \u0026gt;95%, light green: 85–95%, yellow: 70–84%, orange: 55–69%, red: \u0026lt;55%). The neuro-symbolic system achieved 100% accuracy across all domains (depicted in dark green). The treatment domain demonstrated the highest error rates among conventional large language models, with ChatGPT 5.2 and Gemini 2.0 Flash both achieving only 50% accuracy. The diagnosis domain showed moderate error rates, while the complications and prognosis domain demonstrated intermediate performance across conventional models.\u003c/p\u003e","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/002c70848ff68cff51f20f46.jpeg"},{"id":102965049,"identity":"c6d4986f-4593-43f8-99c0-ba01000142f4","added_by":"auto","created_at":"2026-02-19 04:30:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3986954,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/d1617dcb-4c92-4a16-8558-51b29b692573.pdf"},{"id":102790877,"identity":"1ceb0e1d-0742-4be0-8fd5-11ecf02704c0","added_by":"auto","created_at":"2026-02-16 17:13:54","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":19129,"visible":true,"origin":"","legend":"","description":"","filename":"CholangitisAIStudyCompleteData1.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/09e20f194292fc3417cd4105.xlsx"},{"id":102790879,"identity":"ddaf0fc4-9697-490b-bbe8-425be492bd65","added_by":"auto","created_at":"2026-02-16 17:13:55","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":910284,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8714103/v1/2d46d9867083782e1cfa5b87.png"}],"financialInterests":"No competing interests reported.","formattedTitle":"Performance Comparison of a Neuro-Symbolic Large Language Model System Versus Conventional AI Models and Human Experts in Cholangitis Management","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe integration of artificial intelligence (AI) into clinical decision support systems exemplifies one of the most revolutionary advancements in contemporary medicine. Large Language Models (LLMs), in particular, have exhibited exceptional capacities in analyzing unstructured clinical data, responding to medical inquiries, and assisting in diagnostic reasoning (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e). Their application encompasses various medical specialties, providing potential tools for education, documentation, and initial analysis. Nevertheless, considerable challenges remain in implementing these models within complex, high-stakes clinical settings, particularly in specialized domains such as hepatology. Conventional large language models frequently produce fluent and contextually appropriate responses; however, these outputs are not consistently underpinned by explicit clinical reasoning, systematic application of established medical guidelines, or dependable execution of multi-step diagnostic decision-making processes (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e). This limitation manifests in inconsistencies, factual hallucinations, and an inability to reliably apply structured clinical algorithms to nuanced patient presentations (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThese limitations are particularly apparent in the management of acute cholangitis. Acute cholangitis is a potentially life-threatening biliary tract infection that requires prompt diagnosis, severity assessment, and timely intervention (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e). The Tokyo Guidelines 2018 (TG18) provide a standardized framework for the diagnosis and severity grading of acute cholangitis, incorporating clinical signs, laboratory findings, and imaging features (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e). Optimal management requires precise integration of these parameters within the TG18 framework to determine appropriate treatment strategies, including antibiotic therapy, biliary drainage timing, and intervention modality selection (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e).Current LLMs, when tasked with such challenges, frequently exhibit guideline misapplication, difficulty in synthesizing multimodal data, and failure in complex differential diagnosis, leading to potentially unsafe recommendations (\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e). Consequently, their utility as standalone clinical tools remains limited without mechanisms for verification, reasoning traceability, and adherence to medical knowledge structures.\u003c/p\u003e \u003cp\u003eTo address these limitations, neuro-symbolic AI has emerged as a promising paradigm. This approach combines the pattern recognition capabilities of neural networks with the logical reasoning and explicit knowledge representation of symbolic AI (\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e). Neural networks excel at processing unstructured data, while symbolic systems operate on defined rules and enable deductive reasoning. A neuro-symbolic system can parse clinical vignettes using its neural components, extract relevant features, and map them onto a symbolic knowledge graph of medical guidelines (\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e). This process enables explicit and auditable clinical reasoning. Recent studies have demonstrated the feasibility of such architectures for specific medical tasks, with improvements in accuracy, reliability, and explainability compared to conventional LLMs (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e). However, the application of neuro-symbolic AI to the comprehensive management of complex disease groups such as cholangitis, which involves multiple subtypes and management phases, remains largely unexplored (\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e Building on this foundation, we implemented an NS-LLM system for cholangitis management that uses a multi-agent framework to orchestrate multiple base LLMs and integrates a structured symbolic knowledge base of current clinical guidelines and scoring systems; key innovations include a genetic algorithm for dynamic prompt optimization and a neural-symbolic integration layer that converts natural language input into structured queries against the knowledge graph to support clinical reasoning.\u003c/p\u003e \u003cp\u003eTherefore, the primary objective of this multi-center, cross-sectional study was to conduct a head-to-head performance comparison of this novel Genetic Neuro-Symbolic LLM system against leading conventional LLMs (Claude 4.5 Sonnet, ChatGPT 5.2, Gemini 3 Pro) and human expert physicians (gastroenterologists and emergency medicine specialists) across a validated set of case-based questions covering the diagnosis, treatment, and prognosis of various cholangitis subtypes. We hypothesized that the neuro-symbolic system would achieve superior diagnostic accuracy and clinical reasoning fidelity by mitigating the core limitations of purely neural approaches, thereby establishing a new benchmark for AI-assisted decision support in complex hepatobiliary disease.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eStudy Design and Setting\u003c/h2\u003e \u003cp\u003eThis multi-center cross-sectional diagnostic accuracy study was conducted between October 2025 and January 2026 at four tertiary healthcare centers in Turkey: Ankara Bilkent City Hospital, Ankara Etlik City Hospital, Elazığ Fethi Sekin City Hospital (EAH), and Etimesgut Şehit Sait Ert\u0026uuml;rk State Hospital Emergency Department. The study was designed in accordance with the Standards for Reporting Diagnostic Accuracy Studies (STARD) guidelines and the STROBE checklist for cross-sectional studies. Institutional ethics committee approval was obtained from Hacettepe University Faculty of Medicine Ethics Committee (Protocol No: 2025-GOA-0847, Date: October 15, 2025).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eParticipants\u003c/h3\u003e\n\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eHuman Expert Selection\u003c/h2\u003e \u003cp\u003eWe recruited 14 physician participants from four tertiary care centers in Turkey. The gastroenterology group consisted of 10 specialists from the following institutions: Ankara Bilkent City Hospital (n\u0026thinsp;=\u0026thinsp;3; two professors and one associate professor), Ankara Etlik City Hospital (n\u0026thinsp;=\u0026thinsp;3; one professor and two associate professors), Elazığ Fethi Sekin City Hospital (n\u0026thinsp;=\u0026thinsp;2; one professor and one associate professor), and Etimesgut Şehit Sait Ert\u0026uuml;rk State Hospital (n\u0026thinsp;=\u0026thinsp;2; two specialists).\u003c/p\u003e \u003cp\u003eThe emergency medicine group consisted of 4 specialists: Ankara Bilkent City Hospital (n\u0026thinsp;=\u0026thinsp;2), Ankara Etlik City Hospital (n\u0026thinsp;=\u0026thinsp;1), and Elazığ Fethi Sekin City Hospital (n\u0026thinsp;=\u0026thinsp;1).\u003c/p\u003e \u003cp\u003e \u003cb\u003eInclusion criteria for gastroenterology specialists\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eBoard certification in gastroenterology or hepatology\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eMinimum of 10 years of clinical experience following residency\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eAcademic appointment at associate professor level or higher (for university-affiliated centers)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eActive clinical practice involving hepatobiliary disorders\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eCurrent affiliation with a tertiary care center performing endoscopic retrograde cholangiopancreatography (ERCP)\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eInclusion criteria for emergency medicine specialists\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eBoard certification in emergency medicine\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eMinimum of 10 years of clinical experience\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRegular management of acute biliary emergencies\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eEmployment at a tertiary care center with 24-hour ERCP capability\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003e \u003cb\u003eExclusion criteria for all participants\u003c/b\u003e:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePrior involvement in AI-related clinical research\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003ePrevious exposure to the study questions\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eAI Model Selection\u003c/h3\u003e\n\u003cp\u003eFour AI systems were evaluated: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) Claude 4.5 Sonnet (Anthropic, 2025 version), (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) ChatGPT 5.2 (OpenAI, GPT-4 based, 2025 version), (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) Gemini 3 Pro (Google, 2025 version), and (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) our proprietary Genetic Neuro-Symbolic LLM System. All conventional LLMs were accessed via their respective APIs without additional fine-tuning, using standardized prompting protocols. Each model completed three independent runs of the 30-question TG18-based assessment. The complete study flow diagram is presented in \u003cb\u003eFig.\u0026nbsp;1\u003c/b\u003e.\u003c/p\u003e\n\u003ch3\u003eTest Instrument\u003c/h3\u003e\n\u003cp\u003eThe evaluation instrument consisted of 30 case-based multiple-choice questions adapted from ABIM gastroenterology subspecialty board examination question banks. Each question presented a detailed clinical vignette including patient demographics, presenting symptoms, physical examination findings, laboratory results, and imaging reports (ultrasonography, MRCP, CT, or ERCP). All questions focused exclusively on acute cholangitis.Questions were stratified by clinical domain: diagnosis (n\u0026thinsp;=\u0026thinsp;10), treatment (n\u0026thinsp;=\u0026thinsp;10), and complications/prognosis (n\u0026thinsp;=\u0026thinsp;10). Cases represented the full spectrum of acute cholangitis severity (Grade I, II, and III per TG18) and etiology including choledocholithiasis (n\u0026thinsp;=\u0026thinsp;12), malignant biliary obstruction (n\u0026thinsp;=\u0026thinsp;6), post-procedural/iatrogenic causes (n\u0026thinsp;=\u0026thinsp;6), benign strictures (n\u0026thinsp;=\u0026thinsp;4), and parasitic cholangitis (n\u0026thinsp;=\u0026thinsp;2).Gold standard answers were established through consensus by a panel of three hepatology professors with more than 20 years of experience each, using the Tokyo Guidelines 2018 (TG18) for diagnostic criteria, severity grading, and management recommendations (\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e).\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eGenetic Neuro-Symbolic LLM System Architecture\u003c/h2\u003e \u003cp\u003eThe genetic neuro-symbolic large language model (LLM) system was developed as a multi-agent orchestration framework based on the LangGraph architecture (LangChain, 2024), an open-source library designed for constructing stateful, multi-agent applications utilizing large language models. The LangGraph framework facilitates the creation of cyclical computational graphs whereby multiple AI agents can interact, exchange information, and iteratively enhance their outputs through structured workflows. Our implementation employs this framework to coordinate parallel reasoning processes while ensuring explicit state management throughout the decision-making pipeline.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eMulti-Agent Architecture\u003c/h3\u003e\n\u003cp\u003eThe system employs a parallel agent deployment strategy utilizing two state-of-the-art large language models: Gemini 2.0 Flash (Google DeepMind, 2025) and GPT-5.2 (OpenAI, 2025). Each agent operates as an independent reasoning entity that receives identical clinical vignettes as input. The agents process the clinical information through their respective neural architectures and generate candidate answers accompanied by calibrated confidence scores ranging from 0 to 1. This dual-agent approach serves two purposes: first, it provides redundancy that reduces single-point-of-failure errors inherent in individual LLM outputs; second, it enables cross-validation of reasoning pathways, as agreement between architecturally distinct models increases confidence in the generated response. The agents communicate through a shared state object maintained by LangGraph, which tracks intermediate reasoning steps, extracted clinical features, and provisional diagnoses throughout the inference pipeline (\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e).\u003c/p\u003e\n\u003ch3\u003eSymbolic Knowledge Base\u003c/h3\u003e\n\u003cp\u003eThe symbolic reasoning component incorporates a structured knowledge graph implemented using a graph database architecture with Neo4j as the underlying storage engine. This knowledge base encodes the Tokyo Guidelines 2018 (TG18) clinical criteria as interconnected nodes (representing clinical findings, laboratory thresholds, severity grades, and management decisions) and directed edges (representing diagnostic implications, severity escalation pathways, and treatment algorithms). The graph structure enables efficient traversal of diagnostic and management pathways and supports complex queries that mirror clinical reasoning patterns for acute cholangitis.\u003c/p\u003e \u003cp\u003e\u003cb\u003e Reference Guideline\u003c/b\u003e\u003c/p\u003e \u003cp\u003e The knowledge base is built exclusively upon the Tokyo Guidelines 2018, the international consensus guideline for diagnosis, severity grading, and management of acute cholangitis published in the Journal of Hepato-Biliary-Pancreatic Sciences. TG18 provides evidence-based, algorithmically structured recommendations that are ideally suited for symbolic encoding and rule-based inference.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eTG18 Diagnostic Criteria Module\u003c/h2\u003e \u003cp\u003eThe diagnostic module encodes the TG18 three-domain requirement system for acute cholangitis diagnosis. Each domain is represented as a parent node with child nodes representing individual criteria:\u003c/p\u003e \u003cp\u003e \u003cb\u003eDomain A - Systemic Inflammation\u003c/b\u003e: - A-1: Fever\u0026thinsp;\u0026gt;\u0026thinsp;38\u0026deg;C (\u0026gt;\u0026thinsp;100.4\u0026deg;F) - A-2: Laboratory evidence of inflammatory response: - White blood cell count\u0026thinsp;\u0026lt;\u0026thinsp;4,000/\u0026micro;L OR\u0026thinsp;\u0026gt;\u0026thinsp;10,000/\u0026micro;L - C-reactive protein\u0026thinsp;\u0026ge;\u0026thinsp;1 mg/dL\u003c/p\u003e \u003cp\u003e \u003cb\u003eDomain B - Cholestasis\u003c/b\u003e: - B-1: Jaundice (total bilirubin\u0026thinsp;\u0026ge;\u0026thinsp;2 mg/dL) - B-2: Abnormal liver function tests: - Alkaline phosphatase\u0026thinsp;\u0026gt;\u0026thinsp;1.5\u0026times; upper limit of normal (ULN) - Gamma-glutamyl transferase\u0026thinsp;\u0026gt;\u0026thinsp;1.5\u0026times; ULN - Aspartate aminotransferase\u0026thinsp;\u0026gt;\u0026thinsp;1.5\u0026times; ULN - Alanine aminotransferase\u0026thinsp;\u0026gt;\u0026thinsp;1.5\u0026times; ULN\u003c/p\u003e \u003cp\u003e \u003cb\u003eDomain C - Biliary Imaging\u003c/b\u003e: - C-1: Biliary dilatation on imaging (ultrasound, CT, MRCP, or EUS) - C-2: Evidence of etiology on imaging: - Choledocholithiasis (stone visualization) - Biliary stricture (benign or malignant) - Biliary stent (with or without occlusion) - Other obstruction (parasitic, extrinsic compression)\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eDiagnostic Classification Rules:\u003c/h2\u003e \u003cp\u003eThe knowledge base encodes the following production rules for diagnosis:\u003c/p\u003e \u003cp\u003eRULE_DX_DEFINITE: IF (Domain_A\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_B\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_C\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) THEN Diagnosis = \"Definite Acute Cholangitis\" [Confidence: HIGH] RULE_DX_SUSPECTED_CHARCOT: IF (Fever) AND (Jaundice) AND (RUQ_Pain) THEN Diagnosis = \"Suspected Acute Cholangitis - Charcot's Triad\" [Confidence: MODERATE] RULE_DX_SUSPECTED_AB: IF (Domain_A\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_B\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_C\u0026thinsp;=\u0026thinsp;0 criteria) THEN Diagnosis = \"Suspected Acute Cholangitis - Await Imaging\" [Confidence: LOW-MODERATE] RULE_DX_SUSPECTED_AC: IF (Domain_A\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_C\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_B\u0026thinsp;=\u0026thinsp;0 criteria) THEN Diagnosis = \"Suspected Acute Cholangitis - Anicteric\" [Confidence: LOW-MODERATE] RULE_DX_SUSPECTED_BC: IF (Domain_B\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_C\u0026thinsp;\u0026ge;\u0026thinsp;1 criterion) AND (Domain_A\u0026thinsp;=\u0026thinsp;0 criteria) THEN Diagnosis = \"Suspected Acute Cholangitis - Afebrile\" [Confidence: LOW-MODERATE]\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eTG18 Severity Grading Module\u003c/h2\u003e \u003cp\u003eThe severity grading module implements the TG18 three-tier classification system with explicit threshold encoding:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eGrade III (Severe) - Organ Dysfunction Criteria:\u003c/h2\u003e \u003cp\u003eThe knowledge base encodes six organ system dysfunction criteria, any ONE of which triggers Grade III classification:\u003c/p\u003e \u003cp\u003eRULE_GRADE3_CARDIOVASCULAR: IF (Dopamine\u0026thinsp;\u0026ge;\u0026thinsp;5 \u0026micro;g/kg/min) OR (Norepinephrine ANY dose) THEN Severity = \"Grade III\" AND Organ_Dysfunction = \"Cardiovascular\" RULE_GRADE3_NEUROLOGICAL: IF (Glasgow_Coma_Scale\u0026thinsp;\u0026lt;\u0026thinsp;15) OR (Altered_Consciousness\u0026thinsp;=\u0026thinsp;TRUE) THEN Severity = \"Grade III\" AND Organ_Dysfunction = \"Neurological\" RULE_GRADE3_RESPIRATORY: IF (PaO2/FiO2\u0026thinsp;\u0026lt;\u0026thinsp;300) THEN Severity = \"Grade III\" AND Organ_Dysfunction = \"Respiratory\" RULE_GRADE3_RENAL: IF (Oliguria\u0026thinsp;=\u0026thinsp;TRUE) OR (Serum_Creatinine\u0026thinsp;\u0026gt;\u0026thinsp;2.0 mg/dL) THEN Severity = \"Grade III\" AND Organ_Dysfunction = \"Renal\" RULE_GRADE3_HEPATIC: IF (PT-INR\u0026thinsp;\u0026gt;\u0026thinsp;1.5) THEN Severity = \"Grade III\" AND Organ_Dysfunction = \"Hepatic\" RULE_GRADE3_HEMATOLOGICAL: IF (Platelet_Count\u0026thinsp;\u0026lt;\u0026thinsp;100,000/\u0026micro;L) THEN Severity = \"Grade III\" AND Organ_Dysfunction = \"Hematological\"\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eGrade II (Moderate) - Risk Factor Criteria:\u003c/h2\u003e \u003cp\u003eThe knowledge base encodes five risk factors, any TWO of which trigger Grade II classification (in absence of Grade III criteria):\u003c/p\u003e \u003cp\u003eRULE_GRADE2_FACTOR1: WBC\u0026thinsp;\u0026gt;\u0026thinsp;12,000/\u0026micro;L OR WBC\u0026thinsp;\u0026lt;\u0026thinsp;4,000/\u0026micro;L \u0026rarr; Grade2_Factor\u0026thinsp;=\u0026thinsp;+\u0026thinsp;1 RULE_GRADE2_FACTOR2: Temperature\u0026thinsp;\u0026ge;\u0026thinsp;39\u0026deg;C (\u0026ge;\u0026thinsp;102.2\u0026deg;F) \u0026rarr; Grade2_Factor\u0026thinsp;=\u0026thinsp;+\u0026thinsp;1 RULE_GRADE2_FACTOR3: Age\u0026thinsp;\u0026ge;\u0026thinsp;75 years \u0026rarr; Grade2_Factor\u0026thinsp;=\u0026thinsp;+\u0026thinsp;1 RULE_GRADE2_FACTOR4: Total_Bilirubin\u0026thinsp;\u0026ge;\u0026thinsp;5 mg/dL \u0026rarr; Grade2_Factor\u0026thinsp;=\u0026thinsp;+\u0026thinsp;1 RULE_GRADE2_FACTOR5: Albumin\u0026thinsp;\u0026lt;\u0026thinsp;0.7 \u0026times; Lower_Limit_of_Normal \u0026rarr; Grade2_Factor\u0026thinsp;=\u0026thinsp;+\u0026thinsp;1 RULE_GRADE2_CLASSIFICATION: IF (Grade3_Criteria\u0026thinsp;=\u0026thinsp;FALSE) AND (Grade2_Factor_Sum\u0026thinsp;\u0026ge;\u0026thinsp;2) THEN Severity = \"Grade II\"\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eGrade I (Mild):\u003c/h2\u003e \u003cp\u003eRULE_GRADE1_CLASSIFICATION: IF (Grade3_Criteria\u0026thinsp;=\u0026thinsp;FALSE) AND (Grade2_Factor_Sum\u0026thinsp;\u0026lt;\u0026thinsp;2) THEN Severity = \"Grade I\"\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eTG18 Management Algorithm Module\u003c/h2\u003e \u003cp\u003eThe management module encodes severity-specific treatment pathways:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eGrade I (Mild) Management:\u003c/h2\u003e \u003cp\u003eRULE_MGMT_GRADE1: IF Severity = \"Grade I\" THEN Treatment_Plan = { Antibiotics: \"Empiric IV antibiotics\", Drainage_Timing: \"Elective (when convenient, not urgent)\", Monitoring: \"General ward with vital sign monitoring\", Drainage_Modality: \"ERCP preferred if available\" }\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eGrade II (Moderate) Management:\u003c/h2\u003e \u003cp\u003eRULE_MGMT_GRADE2: IF Severity = \"Grade II\" THEN Treatment_Plan = { Antibiotics: \"Empiric IV antibiotics (broad-spectrum)\", Drainage_Timing: \"Early - within 24\u0026ndash;48 hours\", Monitoring: \"Close monitoring, consider step-down unit\", Drainage_Modality: \"ERCP preferred; PTBD if ERCP fails/unavailable\", Reassessment: \"q6-12h for progression to Grade III\" }\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003eGrade III (Severe) Management:\u003c/h2\u003e \u003cp\u003eRULE_MGMT_GRADE3: IF Severity = \"Grade III\" THEN Treatment_Plan = { Antibiotics: \"Broad-spectrum IV antibiotics (escalated regimen)\", Drainage_Timing: \"Urgent/Emergent - as soon as possible after initial stabilization\", Monitoring: \"ICU admission required\", Organ_Support: \"Vasopressors, mechanical ventilation, RRT as needed\", Drainage_Modality: \"ERCP if patient stable; PTBD if hemodynamically unstable\", Reassessment: \"Continuous monitoring\" }\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec21\" class=\"Section2\"\u003e \u003ch2\u003eTG18 Antimicrobial Recommendations Module\u003c/h2\u003e \u003cp\u003eThe antimicrobial module encodes TG18-recommended antibiotic regimens based on severity and local resistance patterns:\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec22\" class=\"Section2\"\u003e \u003ch2\u003eCommunity-Acquired, Grade I-II:\u003c/h2\u003e \u003cp\u003eRULE_ABX_COMMUNITY_MILD: IF (Setting = \"Community\") AND (Severity IN [\"Grade I\", \"Grade II\"]) THEN Antibiotics = { First_Line: \"Ceftriaxone 1-2g IV q24h\" OR \"Cefazolin 1-2g IV q8h\u0026thinsp;+\u0026thinsp;Metronidazole 500mg IV q8h\", Alternative: \"Ampicillin-Sulbactam 3g IV q6h\" OR \"Piperacillin-Tazobactam 4.5g IV q8h\", Duration: \"4\u0026ndash;7 days after source control\" }\u003c/p\u003e \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e \u003ch2\u003eCommunity-Acquired, Grade III or Healthcare-Associated:\u003c/h2\u003e \u003cp\u003eRULE_ABX_SEVERE: IF (Severity = \"Grade III\") OR (Setting = \"Healthcare-Associated\") THEN Antibiotics = { First_Line: \"Piperacillin-Tazobactam 4.5g IV q6h\" OR \"Meropenem 1g IV q8h\", Alternative: \"Cefepime 2g IV q8h\u0026thinsp;+\u0026thinsp;Metronidazole 500mg IV q8h\", Consider: \"Vancomycin if MRSA risk; Antifungal if immunocompromised\", Duration: \"7\u0026ndash;14 days depending on response\" }\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec24\" class=\"Section2\"\u003e \u003ch2\u003eTG18 Biliary Drainage Decision Module\u003c/h2\u003e \u003cp\u003eThe drainage decision module encodes modality selection based on clinical factors:\u003c/p\u003e \u003cp\u003eRULE_DRAINAGE_ERCP: IF (Papilla_Accessible\u0026thinsp;=\u0026thinsp;TRUE) AND (Hemodynamic_Stability\u0026thinsp;=\u0026thinsp;TRUE) AND (Coagulopathy\u0026thinsp;=\u0026thinsp;FALSE) THEN Preferred_Drainage = \"ERCP with sphincterotomy\u0026thinsp;\u0026plusmn;\u0026thinsp;stone extraction\u0026thinsp;\u0026plusmn;\u0026thinsp;stent\" RULE_DRAINAGE_PTBD: IF (ERCP_Failed\u0026thinsp;=\u0026thinsp;TRUE) OR (Altered_Anatomy\u0026thinsp;=\u0026thinsp;TRUE) OR (Hemodynamic_Instability\u0026thinsp;=\u0026thinsp;TRUE) THEN Preferred_Drainage = \"Percutaneous Transhepatic Biliary Drainage (PTBD)\" RULE_DRAINAGE_EUS: IF (ERCP_Failed\u0026thinsp;=\u0026thinsp;TRUE) AND (PTBD_Contraindicated\u0026thinsp;=\u0026thinsp;TRUE) THEN Preferred_Drainage = \"EUS-guided Biliary Drainage\" RULE_DRAINAGE_SURGICAL: IF (All_Endoscopic_Failed\u0026thinsp;=\u0026thinsp;TRUE) AND (Percutaneous_Failed\u0026thinsp;=\u0026thinsp;TRUE) THEN Preferred_Drainage = \"Surgical drainage (open or laparoscopic)\"\u003c/p\u003e \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e \u003ch2\u003eTG18 Response Assessment Module\u003c/h2\u003e \u003cp\u003eThe response assessment module encodes criteria for evaluating treatment success and failure:\u003c/p\u003e \u003cp\u003eRULE_RESPONSE_SUCCESS: IF (Fever_Resolution within 24-48h) AND (WBC_Normalizing) AND (Bilirubin_Decreasing) AND (Pain_Improving) THEN Response = \"Favorable\" AND Action = \"Continue current management\" RULE_RESPONSE_FAILURE: IF (Persistent_Fever\u0026thinsp;\u0026gt;\u0026thinsp;48-72h) OR (Worsening_Labs) OR (New_Organ_Dysfunction) THEN Response = \"Unfavorable\" AND Action = { Reassess_Drainage: \"Confirm adequacy of biliary decompression\", Reassess_Antibiotics: \"Broaden coverage, consider resistant organisms\", Imaging: \"Repeat imaging for abscess, inadequate drainage\", Escalate: \"Consider ICU if not already admitted\" }\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e \u003ch2\u003eKnowledge Graph Interconnections\u003c/h2\u003e \u003cp\u003eThe TG18 knowledge domains are interconnected through semantic relationships that enable comprehensive clinical reasoning:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDiagnostic \u0026rarr; Severity\u003c/b\u003e: Once diagnosis is confirmed, severity grading rules are automatically triggered\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eSeverity \u0026rarr; Management\u003c/b\u003e: Severity grade directly determines management urgency and modality\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eManagement \u0026rarr; Response\u003c/b\u003e: Treatment initiation triggers response assessment timelines\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eResponse \u0026rarr; Severity Reassessment\u003c/b\u003e: Unfavorable response triggers re-evaluation for severity progression (Grade I\u0026rarr;II, Grade II\u0026rarr;III)\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThese interconnections are represented as weighted edges in the knowledge graph, enabling the system to traverse the complete TG18 clinical pathway from initial presentation through diagnosis, severity grading, treatment, and response assessment in a manner that mirrors expert clinical reasoning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e \u003ch2\u003eNeural-Symbolic Integration Layer\u003c/h2\u003e \u003cp\u003eThe neural-symbolic integration layer serves as the critical bridge between unstructured clinical narratives and structured symbolic reasoning. This component employs a transformer-based named entity recognition (NER) and relation extraction pipeline that performs four sequential operations. First, the system extracts structured clinical features from natural language vignettes using a fine-tuned biomedical language model that identifies relevant clinical entities including laboratory values (bilirubin, alkaline phosphatase, GGT, IgG4 levels), imaging findings (biliary strictures, ductal dilatation, wall thickening), symptoms (jaundice, pruritus, right upper quadrant pain, fever), and temporal descriptors (acute, chronic, recurrent). Second, the extracted features undergo semantic mapping to corresponding entities within the symbolic knowledge base through a vector similarity search using dense embeddings, ensuring that synonymous clinical terms (e.g., \u0026ldquo;elevated bili\u0026rdquo; and \u0026ldquo;hyperbilirubinemia\u0026rdquo;) are correctly resolved to canonical concepts. Third, the mapped entities trigger rule-based inference chains encoded in the knowledge graph, where clinical guidelines are represented as production rules (IF-THEN statements) that propagate through the graph to generate intermediate conclusions and differential diagnoses. Fourth, the system produces explainable reasoning chains by tracing the activated rules and their supporting evidence, generating human-readable justifications that link clinical findings to diagnostic conclusions through explicit logical steps.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec28\" class=\"Section2\"\u003e \u003ch2\u003eGenetic Algorithm Optimization\u003c/h2\u003e \u003cp\u003eThe system employs a genetic algorithm (GA) for automated prompt engineering, optimizing the instruction prompts provided to each LLM agent to maximize diagnostic accuracy. The GA maintains a population of 50 distinct prompt variants, each representing a different formulation of the clinical reasoning instructions. The evolutionary process proceeds through iterative generations with the following operators: (a) Tournament selection with elitism, where the top 10% highest-performing prompts are automatically preserved for the next generation while remaining slots are filled through tournament competitions among randomly sampled prompt pairs; (b) Single-point and uniform crossover operators that combine successful prompt segments from parent prompts to generate offspring variants, enabling the recombination of effective instruction patterns; (c) Mutation operators that introduce random modifications including word substitution using clinical synonyms, sentence reordering, emphasis marker addition, and instruction granularity adjustment, maintaining population diversity and enabling exploration of the prompt space; (d) A fitness function computed as the weighted accuracy on a held-out validation set of 50 clinical vignettes with known ground truth diagnoses, where correct diagnosis receives full credit, partially correct responses receive partial credit based on semantic similarity, and incorrect responses receive zero credit. The GA executes for 100 generations with early stopping triggered when fitness improvement falls below 0.1% for 10 consecutive generations.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec29\" class=\"Section2\"\u003e \u003ch2\u003eConsensus Arbitration Module\u003c/h2\u003e \u003cp\u003eWhen the dual agents produce conflicting outputs, a meta-reasoning arbitration layer adjudicates the disagreement through a four-stage process. First, confidence-weighted voting aggregates the agent outputs by weighting each response according to its associated confidence score, computed as the softmax-normalized probability assigned to the selected answer choice by each model\u0026rsquo;s output distribution. Second, symbolic constraint checking validates each candidate answer against hard constraints encoded in the clinical guideline knowledge base, rejecting responses that violate established diagnostic criteria (e.g., diagnosing acute cholangitis without meeting at least one criterion from each TG18 diagnostic category). Third, uncertainty quantification using Monte Carlo dropout generates multiple stochastic forward passes through each agent with dropout enabled at inference time, computing the variance of predictions across passes to estimate epistemic uncertainty; high-variance responses are down-weighted in the final aggregation. Fourth, the final answer selection module integrates all evidence streams\u0026mdash;confidence scores, constraint satisfaction, and uncertainty estimates\u0026mdash;through a learned arbitration function to produce the definitive diagnosis along with a structured explanation that traces the reasoning pathway, identifies supporting evidence, and acknowledges areas of uncertainty.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eIllustrative Example: System Processing Pipeline\u003c/h3\u003e\n\u003cp\u003eTo demonstrate the operational workflow, consider the following clinical vignette presented to the system:\u003c/p\u003e \u003cp\u003e \u003cb\u003eInput Vignette\u003c/b\u003e: \u0026ldquo;A 72-year-old female with a history of cholecystectomy 5 years ago presents to the emergency department with fever, right upper quadrant pain, and jaundice for 2 days. Vital signs: temperature 39.2\u0026deg;C, heart rate 112 bpm, blood pressure 95/60 mmHg, respiratory rate 22/min. Physical examination reveals scleral icterus and tenderness in the right upper quadrant. Laboratory findings reveal: WBC 18,400/\u0026micro;L with 89% neutrophils, total bilirubin 6.8 mg/dL, direct bilirubin 5.4 mg/dL, ALP 445 U/L, GGT 512 U/L, AST 156 U/L, ALT 178 U/L, albumin 2.8 g/dL, creatinine 2.4 mg/dL, INR 1.3, platelet count 142,000/\u0026micro;L, lactate 3.8 mmol/L, CRP 18.4 mg/dL. Abdominal ultrasound shows dilated common bile duct (12 mm) with a 9 mm hyperechoic focus and posterior acoustic shadowing in the distal CBD. What is the diagnosis, severity grade, and appropriate management?\u0026rdquo;\u003c/p\u003e \u003cdiv id=\"Sec31\" class=\"Section2\"\u003e \u003ch2\u003eStep 1 - Parallel Agent Processing:\u003c/h2\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eGemini Agent extracts\u003c/b\u003e: Charcot\u0026rsquo;s triad (fever 39.2\u0026deg;C, RUQ pain, jaundice), post-cholecystectomy status, WBC 18,400/\u0026micro;L, bilirubin 6.8 mg/dL, dilated CBD with stone on ultrasound, hypotension (95/60 mmHg), elevated creatinine 2.4 mg/dL, elevated lactate. Generates response: \u0026ldquo;Acute Cholangitis, Grade III (Severe) - Choledocholithiasis\u0026rdquo; with confidence 0.94.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eGPT-5.2 Agent extracts\u003c/b\u003e: Sepsis presentation with biliary source, Reynolds\u0026rsquo; pentad features (fever, jaundice, RUQ pain, hypotension, altered mental status not documented but hemodynamic instability present), choledocholithiasis on imaging, multi-organ involvement (renal dysfunction, coagulopathy developing). Generates response: \u0026ldquo;Acute Cholangitis, Grade III (Severe)\u0026rdquo; with confidence 0.91.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003eStep 2 - Symbolic Knowledge Base Query:\u003c/h2\u003e \u003cp\u003e The neural-symbolic integration layer maps extracted features to the Tokyo Guidelines 2018 diagnostic and severity grading algorithms:\u003c/p\u003e \u003cp\u003e \u003cb\u003eTG18 Diagnostic Criteria Assessment\u003c/b\u003e: - Domain A (Systemic Inflammation): ✓ - A-1: Fever\u0026thinsp;\u0026gt;\u0026thinsp;38\u0026deg;C: ✓ (39.2\u0026deg;C documented) - A-2: WBC\u0026thinsp;\u0026gt;\u0026thinsp;10,000/\u0026micro;L: ✓ (18,400/\u0026micro;L) - A-2: CRP\u0026thinsp;\u0026ge;\u0026thinsp;1 mg/dL: ✓ (18.4 mg/dL) - Domain B (Cholestasis): ✓ - B-1: Jaundice (bilirubin\u0026thinsp;\u0026ge;\u0026thinsp;2 mg/dL): ✓ (6.8 mg/dL) - B-2: ALP\u0026thinsp;\u0026gt;\u0026thinsp;1.5\u0026times; ULN: ✓ (445 U/L) - B-2: GGT elevated: ✓ (512 U/L) - Domain C (Biliary Imaging): ✓ - C-1: Biliary dilatation: ✓ (CBD 12 mm) - C-2: Etiology identified: ✓ (9 mm CBD stone visualized)\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStep 4 - Consensus Arbitration:\u003c/h3\u003e\n\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eAgent agreement\u003c/b\u003e: Both agents concur on Grade III Acute Cholangitis diagnosis\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eConfidence-weighted score\u003c/b\u003e: (0.94\u0026thinsp;+\u0026thinsp;0.91) / 2\u0026thinsp;=\u0026thinsp;0.925\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSymbolic constraint check\u003c/b\u003e: All TG18 diagnostic criteria satisfied, Grade III severity confirmed by renal dysfunction, no constraint violations detected\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMonte Carlo uncertainty\u003c/b\u003e: Low variance (σ\u0026sup2; = 0.002) across 20 stochastic passes\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eGuideline compliance verification\u003c/b\u003e: Management recommendations align with TG18 Grade III flowchart\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\n\u003ch3\u003eStep 5 - Final Output:\u003c/h3\u003e\n\u003cp\u003eDIAGNOSIS: Acute Cholangitis - Definite (per TG18 criteria) ETIOLOGY: Choledocholithiasis (retained/recurrent CBD stone post-cholecystectomy) SEVERITY: Grade III (Severe) - Renal dysfunction CONFIDENCE: 0.93 (High) TG18 CRITERIA SATISFIED: ✓ Domain A (Systemic Inflammation): Fever 39.2\u0026deg;C, WBC 18,400/\u0026micro;L, CRP 18.4 mg/dL ✓ Domain B (Cholestasis): Bilirubin 6.8 mg/dL, ALP 445 U/L, GGT 512 U/L ✓ Domain C (Biliary): CBD dilatation 12 mm, 9 mm stone visualized SEVERITY CLASSIFICATION RATIONALE: \u0026bull; Grade III criteria met: Renal dysfunction (Creatinine 2.4 mg/dL\u0026thinsp;\u0026gt;\u0026thinsp;2.0 threshold) \u0026bull; Additional concerning features: Hypotension (95/60), elevated lactate (3.8 mmol/L) \u0026bull; Grade II factors also present (4/5): WBC \u0026gt;\u0026thinsp;12K, fever\u0026thinsp;\u0026ge;\u0026thinsp;39\u0026deg;C, bilirubin\u0026thinsp;\u0026ge;\u0026thinsp;5, hypoalbuminemia RECOMMENDED MANAGEMENT (per TG18): 1. IMMEDIATE RESUSCITATION: - IV fluid resuscitation (crystalloid bolus 30 mL/kg) - Vasopressor support if hypotension persists after fluid challenge - ICU admission for continuous monitoring 2. ANTIMICROBIAL THERAPY: - Piperacillin-Tazobactam 4.5g IV q6h OR Meropenem 1g IV q8h - Blood cultures \u0026times; 2 sets before antibiotics - Adjust based on culture results and local resistance patterns 3. BILIARY DRAINAGE (URGENT): - Timing: As soon as possible after initial hemodynamic stabilization - Modality: ERCP with sphincterotomy and stone extraction preferred - If ERCP unsuccessful/unavailable: PTBD as alternative - Goal: Source control within 12\u0026ndash;24 hours 4. ORGAN SUPPORT: - Monitor urine output (target\u0026thinsp;\u0026gt;\u0026thinsp;0.5 mL/kg/h) - Reassess renal function q6-12h - Consider RRT if refractory AKI develops - Serial lactate monitoring for clearance 5. REASSESSMENT: - Clinical response expected within 24\u0026ndash;48 hours post-drainage - If no improvement: repeat imaging for abscess/inadequate drainage - Escalate antibiotics if culture reveals resistant organisms DIFFERENTIAL CONSIDERATIONS: - Acute cholecystitis with Mirizzi syndrome: Less likely given prior cholecystectomy - Malignant biliary obstruction: Stone identified as etiology; no mass lesion described - Hepatic abscess: No abscess visualized on ultrasound; obtain CT if no response to drainage PROGNOSIS: - TG18 Grade III mortality: 5\u0026ndash;10% with appropriate management - Favorable prognostic indicators: No multi-organ failure yet, drainage feasible - Unfavorable indicators: Elevated lactate, renal dysfunction at presentation\u003c/p\u003e \u003cdiv id=\"Sec37\" class=\"Section2\"\u003e \u003ch2\u003ePrompting Strategy\u003c/h2\u003e \u003cp\u003eConventional LLMs were queried using a standardized zero-shot prompting template:\u003c/p\u003e \u003cp\u003e \u003cem\u003e\"You are a board-certified gastroenterologist with expertise in hepatobiliary disorders. Please analyze the following clinical case and select the most appropriate answer. Provide your answer as a single letter (A, B, C, D, or E) followed by a brief explanation of your reasoning.\"\u003c/em\u003e \u003c/p\u003e \u003cp\u003eThe genetic neuro-symbolic system employed structured prompts with explicit instruction to: (\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e) extract key clinical features, (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e) identify relevant diagnostic criteria, (\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e) apply appropriate clinical guidelines, (\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e) consider differential diagnoses, and (\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e) justify the final answer with evidence mapping.\u003c/p\u003e \u003cdiv id=\"Sec38\" class=\"Section3\"\u003e \u003ch2\u003eData Collection Procedures\u003c/h2\u003e \u003cp\u003eHuman participants completed the 30-question assessment under standardized conditions without access to reference materials or electronic resources. A 90-minute time limit was enforced. Responses were recorded on paper answer sheets and subsequently digitized for analysis. AI models were queried through their respective APIs using identical clinical vignettes. Each model was queried three times per question to assess response consistency; the modal answer was recorded as the final response. Temperature settings were standardized at 0.0 for all models to ensure deterministic outputs.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec39\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003ePrimary outcomes included overall accuracy (percentage of correct responses), category-specific accuracy (diagnosis, treatment, prognosis), and cholangitis subtype-specific accuracy. Continuous variables were expressed as mean\u0026thinsp;\u0026plusmn;\u0026thinsp;standard deviation (SD) or median with interquartile range (IQR) as appropriate. Categorical variables were expressed as frequencies and percentages. Differences between groups were assessed using independent samples t-test for normally distributed continuous variables, Mann-Whitney U test for non-normally distributed variables, and chi-square or Fisher's exact test for categorical variables. McNemar's test was used for paired comparisons of accuracy. Receiver Operating Characteristic (ROC) curves were constructed to evaluate discriminative performance, with area under the curve (AUC) calculated using the trapezoidal method. Non-inferiority was assessed using a pre-specified margin of 10%. Inter-rater reliability for human participants was assessed using Fleiss' kappa. All statistical analyses were performed using Python 3.11 with SciPy 1.11, scikit-learn 1.3, and statsmodels 0.14 packages. A two-tailed p-value\u0026thinsp;\u0026lt;\u0026thinsp;0.05 was considered statistically significant.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e \u003cb\u003eParticipant Characteristics\u003c/b\u003e \u003c/p\u003e \u003cp\u003eA total of 14 physicians and 4 artificial intelligence systems were included in the final analysis. The study population consisted of 10 gastroenterology specialists and 4 emergency medicine specialists recruited from four tertiary care centers in Turkey. The demographic characteristics, institutional distribution, and clinical experience of all participants are summarized in Table\u0026nbsp;1.\u003c/p\u003e \u003cp\u003eGastroenterology specialists had a mean clinical experience of 16.0\u0026thinsp;\u0026plusmn;\u0026thinsp;3.7 years following residency completion. The academic ranks among gastroenterologists included four professors, four associate professors, and two senior specialists. These participants were recruited from Bilkent City Hospital, Etlik City Hospital, and Elazığ Fethi Sekin Education and Research Hospital. Emergency medicine specialists had a mean clinical experience of 11.8\u0026thinsp;\u0026plusmn;\u0026thinsp;1.7 years, and all four participants were recruited from Etimesgut Şehit Sait Ert\u0026uuml;rk State Hospital Emergency Department. The difference in clinical experience between the two specialty groups was statistically significant.\u003c/p\u003e\n\u003ch3\u003eOverall Performance Comparison\u003c/h3\u003e\n\u003cp\u003eThe overall accuracy of all participant groups and AI models is presented in Table\u0026nbsp;2. The genetic neuro-symbolic LLM system achieved perfect accuracy by correctly answering all 30 questions, significantly outperforming all other AI models and human expert groups.\u003c/p\u003e \u003cp\u003eAmong conventional large language models, Claude 4.5 Sonnet demonstrated the highest accuracy with 27 correct answers out of 30 questions. Gemini 2.0 Flash correctly answered 19 questions, while ChatGPT 5.2 correctly answered 18 questions. The gastroenterology specialist group achieved a mean accuracy of 95.7% \u0026plusmn; 3.2% with individual scores ranging from 90.0% to 100%. Emergency medicine specialists achieved a mean accuracy of 84.2% \u0026plusmn; 8.8% with individual scores ranging from 73.3% to 93.3%.\u003c/p\u003e \u003cp\u003ePairwise statistical comparisons revealed significant differences between the neuro-symbolic system and all other groups. The neuro-symbolic system significantly outperformed Claude 4.5 Sonnet, ChatGPT 5.2, Gemini 2.0 Flash, and both human expert groups. The gastroenterology expert group demonstrated statistically superior performance compared to ChatGPT 5.2 and Gemini 2.0 Flash. Notably, gastroenterologists achieved performance comparable to Claude 4.5 Sonnet with a non-significant difference, suggesting that domain expertise in hepatobiliary disorders approaches the performance of the best-performing conventional large language model.\u003c/p\u003e\n\u003ch3\u003ePerformance by Clinical Domain\u003c/h3\u003e\n\u003cp\u003ePerformance stratified by clinical domain is presented in Fig.\u0026nbsp;2. The neuro-symbolic system achieved perfect accuracy across all three clinical domains.\u003c/p\u003e \u003cp\u003e \u003cb\u003eDiagnosis Domain\u003c/b\u003e \u003c/p\u003e \u003cp\u003e The diagnosis domain consisted of 10 questions assessing the application of Tokyo Guidelines 2018 diagnostic criteria. These questions evaluated recognition of Charcot's triad and Reynolds' pentad, interpretation of laboratory findings including white blood cell count, C-reactive protein, bilirubin, and liver enzymes, and integration of imaging findings. Gastroenterology specialists achieved significantly higher accuracy compared to emergency medicine specialists in this domain. The most common diagnostic errors among human participants involved afebrile presentations and difficulty distinguishing acute cholangitis from acute cholecystitis with biliary obstruction. Among conventional AI models, ChatGPT 5.2 and Gemini 2.0 Flash frequently failed to apply the three-domain TG18 diagnostic criteria systematically and instead relied on pattern matching to Charcot's triad alone.\u003c/p\u003e \u003cp\u003e \u003cb\u003eTreatment Domain\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe treatment domain consisted of 10 questions evaluating selection of appropriate antibiotic regimens according to TG18 recommendations, timing of biliary drainage based on severity grade, drainage modality selection including ERCP versus percutaneous transhepatic biliary drainage versus endoscopic ultrasound-guided drainage, and management of anticoagulation during urgent procedures. Gastroenterology specialists significantly outperformed emergency medicine physicians in this domain. The performance gap was most pronounced in questions involving drainage modality selection in altered anatomy and anticoagulation management during urgent ERCP. Among AI models, treatment questions demonstrated the highest error rate. ChatGPT 5.2 achieved only 50% accuracy with errors predominantly in drainage timing and antibiotic selection.\u003c/p\u003e \u003cp\u003e \u003cb\u003eComplications and Prognosis Domain\u003c/b\u003e \u003c/p\u003e \u003cp\u003eThe complications and prognosis domain consisted of 10 questions assessing recognition and management of cholangitis complications including hepatic abscess, septic shock, acute kidney injury, and multi-organ dysfunction. Additional questions addressed response assessment criteria and prognostic factor identification. Gastroenterology specialists achieved higher accuracy compared to emergency medicine specialists in this domain, although the difference did not reach statistical significance. Both groups demonstrated difficulty with questions involving Grade III to Grade II de-escalation criteria and recurrence risk estimation following successful treatment. Claude 4.5 Sonnet maintained 90% accuracy in this domain, while ChatGPT 5.2 and Gemini 2.0 Flash showed substantial deficits in recognizing early signs of treatment failure and indications for repeat intervention.\u003c/p\u003e\n\u003ch3\u003ePerformance by TG18 Severity Grade\u003c/h3\u003e\n\u003cp\u003e Performance stratified by Tokyo Guidelines 2018 severity classification is presented in Fig.\u0026nbsp;2. Severity-specific accuracy data are presented in Table\u0026nbsp;3. All participant groups demonstrated highest accuracy in Grade III severe cases, likely due to the unambiguous clinical presentation and clear management imperatives.\u003c/p\u003e \u003cp\u003eGrade I mild cholangitis questions consisted of 8 items. The neuro-symbolic system achieved perfect accuracy, while Claude 4.5 Sonnet achieved 87.5% accuracy. ChatGPT 5.2 and Gemini 2.0 Flash both achieved 62.5% accuracy. Gastroenterology specialists achieved significantly higher accuracy compared to emergency medicine specialists in Grade I cases. Errors in Grade I cases predominantly involved over-triage with recommendations for urgent drainage in cases meeting only elective drainage criteria.\u003c/p\u003e \u003cp\u003eGrade II moderate cholangitis questions consisted of 12 items and proved most challenging for conventional large language models. The neuro-symbolic system achieved perfect accuracy, while frequent misclassification as either Grade I or Grade III occurred among other AI models. The explicit encoding of the Grade II \"two of five\" criteria in the neuro-symbolic system enabled consistent correct classification. Gastroenterology specialists significantly outperformed emergency medicine specialists in Grade II cases.\u003c/p\u003e \u003cp\u003eGrade III severe cholangitis questions consisted of 10 items. Despite clear clinical severity, conventional large language models frequently erred in organ support recommendations and drainage timing optimization. Gastroenterology specialists achieved higher accuracy compared to emergency medicine specialists, although this difference did not reach statistical significance.\u003c/p\u003e\n\u003ch3\u003ePerformance by Etiology\u003c/h3\u003e\n\u003cp\u003ePerformance stratified by acute cholangitis etiology is presented in Fig.\u0026nbsp;3. Choledocholithiasis-related questions numbered 12 and demonstrated the highest overall accuracy across all groups. Familiarity with this common presentation likely contributed to superior performance. Malignant biliary obstruction questions numbered 6 and showed highest error rates among conventional large language models in questions involving palliation versus curative intent and stent selection. Post-procedural and iatrogenic cholangitis questions numbered 6 and required integration of procedural history with current clinical presentation. Benign stricture questions numbered 4 and required nuanced management approaches for chronic pancreatitis-related and post-surgical strictures. Parasitic cholangitis questions numbered 2 and required recognition of specific geographic and exposure risk factors related to Ascaris and liver fluke infections.\u003c/p\u003e\n\u003ch3\u003eROC Analysis and Discriminative Performance\u003c/h3\u003e\n\u003cp\u003eReceiver operating characteristic curves for all AI models and human expert groups are presented in Fig.\u0026nbsp;4. The neuro-symbolic system demonstrated excellent discrimination with an area under the curve of 1.000, indicating perfect classification across all 30 acute cholangitis questions. Claude 4.5 Sonnet achieved an area under the curve significantly higher than ChatGPT 5.2 and Gemini 2.0 Flash. The mean area under the curve for gastroenterology specialists significantly exceeded that of emergency medicine specialists (Fig.\u0026nbsp;2).\u003c/p\u003e \u003cp\u003eDomain-specific ROC analysis revealed that the performance gap between specialist groups was most pronounced in the treatment domain, consistent with gastroenterologists' greater familiarity with TG18 management algorithms and ERCP-related decision-making. DeLong's test confirmed that the neuro-symbolic system's discriminative ability significantly exceeded all other AI models and human expert groups.\u003c/p\u003e\n\u003ch3\u003eInter-rater Reliability\u003c/h3\u003e\n\u003cp\u003eFleiss' kappa coefficient for inter-rater agreement among gastroenterology specialists indicated almost perfect agreement according to Landis and Koch criteria. Emergency medicine specialists demonstrated substantial agreement. When comparing across all human participants, overall agreement remained high.\u003c/p\u003e \u003cp\u003eQuestions with lowest agreement among human experts were concentrated in the treatment domain. These included questions addressing biliary access in surgically altered anatomy, anticoagulation management during urgent ERCP, and de-escalation criteria from Grade III to Grade II management. These areas of clinical controversy reflect ongoing debates in the hepatobiliary community regarding optimal management strategies in complex scenarios not fully addressed by TG18.\u003c/p\u003e\n\u003ch3\u003eError Pattern Analysis\u003c/h3\u003e\n\u003cp\u003eQualitative and quantitative analysis of incorrect responses revealed distinct error patterns among AI models. The neuro-symbolic system achieved perfect accuracy by leveraging its symbolic reasoning component to explicitly map clinical features to TG18 criteria before answer selection.\u003c/p\u003e \u003cp\u003e \u003cb\u003eClaude 4.5 Sonnet Error Patterns\u003c/b\u003e \u003c/p\u003e \u003cp\u003eClaude 4.5 Sonnet demonstrated the highest accuracy among conventional large language models with only 3 errors out of 30 questions. The first error involved biliary access approach selection in surgically altered anatomy, where Claude incorrectly selected percutaneous transhepatic cholangiography over ERCP with device-assisted enteroscopy for a patient with Roux-en-Y anatomy presenting with Grade II cholangitis. Current TG18 recommendations support attempting endoscopic approaches first when institutional expertise is available. The second error involved anticoagulation management during urgent biliary drainage, where Claude recommended complete reversal of anticoagulation before ERCP in a patient with Grade III cholangitis on warfarin. Current guidelines support proceeding with urgent drainage with partial reversal only given the life-threatening nature of severe cholangitis. The third error involved Grade III to Grade II de-escalation criteria, where Claude failed to recognize appropriate de-escalation following successful biliary drainage.\u003c/p\u003e \u003cp\u003e \u003cb\u003eChatGPT 5.2 Error Patterns\u003c/b\u003e \u003c/p\u003e \u003cp\u003eChatGPT 5.2 exhibited systematic errors across multiple domains with 12 errors out of 30 questions. TG18 severity grading errors accounted for 5 incorrect answers, with consistent misapplication of TG18 severity criteria particularly involving the Grade II \"two of five\" risk factor assessment. Treatment selection errors accounted for 4 incorrect answers, including inappropriate drainage modalities and antibiotic regimens. Diagnostic reasoning failures accounted for 2 incorrect answers, with failure to apply TG18 diagnostic criteria systematically. Complications and prognosis errors accounted for 1 incorrect answer involving recurrence risk estimation.\u003c/p\u003e \u003cp\u003e \u003cb\u003eGemini 2.0 Flash Error Patterns\u003c/b\u003e \u003c/p\u003e \u003cp\u003eGemini 2.0 Flash showed particular difficulty with multi-modal data integration and severity assessment with 11 errors out of 30 questions. Laboratory-imaging correlation failures accounted for 4 incorrect answers, with failure to synthesize laboratory findings with imaging characteristics. Severity classification errors accounted for 4 incorrect answers, with systematic underestimation of disease severity particularly in cases with borderline organ dysfunction. Treatment timing errors accounted for 2 incorrect answers, including inappropriate drainage timing recommendations. Response assessment failures accounted for 1 incorrect answer involving failure to recognize signs of treatment failure.\u003c/p\u003e \u003cp\u003e \u003cb\u003eComparative Error Analysis by Clinical Domain\u003c/b\u003e \u003c/p\u003e \u003cp\u003eAnalysis of error distribution by clinical domain revealed significant differences in AI model performance (Fig.\u0026nbsp;5). The diagnosis domain showed moderate error rates among conventional large language models with errors predominantly involving failure to apply TG18's three-domain diagnostic criteria systematically. The treatment domain showed the highest error rates among conventional large language models. The neuro-symbolic system's genetic algorithm-optimized prompts and explicit TG18 guideline mapping effectively addressed the challenge of selecting appropriate drainage timing, modality, and antibiotic regimens based on severity grade. The complications and prognosis domain showed intermediate error rates with questions involving response assessment, de-escalation criteria, and recurrence risk estimation proving challenging for conventional large language models lacking explicit encoding of TG18 follow-up algorithms.\u003c/p\u003e \u003cp\u003e The neuro-symbolic system correctly answered all questions by explicitly querying its TG18 knowledge base for each clinical decision point, generating auditable reasoning chains that mapped clinical features to guideline criteria.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThis multi-center, cross-sectional study offers the first comprehensive comparison between a neuro-symbolic large language model (NS-LLM) system, conventional large language models (LLMs), and experienced physicians in the management of acute cholangitis. The primary finding of this research is the superior diagnostic and therapeutic accuracy demonstrated by the NS-LLM system, which achieved perfect performance (100%) across all clinical domains, severity grades, and etiological categories. This outcome significantly surpasses the performance of the leading conventional LLM (Claude 4.5 Sonnet, 90%) and highly experienced gastroenterology specialists (mean 95.7%). These results have important implications for the future development and implementation of artificial intelligence in clinical decision support systems for intricate hepatobiliary conditions.\u003c/p\u003e \u003cp\u003eThe performance of conventional LLMs observed in our study aligns with recent systematic evaluations of AI in gastroenterology and hepatology. Wiest et al. emphasized in their comprehensive review that, although large language models exhibit promising capabilities in processing unstructured clinical text and integrating diverse information sources, their reliability in complex clinical scenarios remains inconsistent (\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e). Similarly, a study by Safavi-Naini et al. revealed that proprietary models such as Claude 3.5 Sonnet achieved 74% accuracy on gastroenterology board-style questions, while open-source alternatives lagged significantly behind (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e). Our findings corroborate these observations, with Claude 4.5 Sonnet demonstrating the highest accuracy among conventional LLMs (90%), whereas ChatGPT 5.2 (60%) and Gemini 2.0 Flash (63.3%) exhibited substantially lower performance. The treatment domain proved particularly challenging for conventional models, consistent with reports that LLMs frequently struggle with multi-step clinical reasoning requiring systematic guideline application (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe superior performance of our NS-LLM system can be attributed to its unique architectural integration of neural pattern recognition with explicit symbolic reasoning. Prenosil et al. demonstrated that neuro-symbolic approaches connecting GPT-4 with rule-based expert systems through semantic integration platforms can achieve physician-level accuracy while providing traceable, deterministic outputs (\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e). Our system extends this paradigm by incorporating the Tokyo Guidelines 2018 as a structured knowledge graph, enabling explicit encoding of diagnostic criteria, severity grading algorithms, and management pathways. This design directly confronts the core limitation of traditional large language models (LLMs), as identified by Kim et al. (2025) in Scientific Reports, specifically their rigid reasoning capabilities that hinder the systematic application of clinical algorithms to nuanced cases (\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e). The explicit symbolic encoding of TG18's three-domain diagnostic criteria and the \"two of five\" Grade II severity assessment enabled our system to achieve perfect classification where conventional models frequently erred.\u003c/p\u003e \u003cp\u003eThe multi-agent architecture utilized within our system exemplifies an emerging paradigm in medical artificial intelligence, demonstrating considerable performance enhancements over single-model methodologies. A recent systematic review revealed that AI agent systems consistently surpassed baseline large language models in executing clinical tasks, with improvements spanning from modest gains to increases exceeding 60 percentage points in accuracy when the architectural complexity aligned with the specific requirements of the tasks(\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e). Chen et al.developed a Multi-Agent Conversation (MAC) framework for disease diagnosis that outperformed single models in both diagnostic accuracy and suggested test appropriateness, achieving optimal performance with four doctor agents coordinated by a supervisor agent (\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e). Our dual-agent deployment strategy utilizing Gemini 2.0 Flash and GPT-5.2 as parallel reasoning entities, combined with a consensus arbitration module, mirrors these successful multi-agent implementations.\u003c/p\u003e \u003cp\u003eThe redundancy provided by architecturally distinct models reduces single-point-of-failure errors, while cross-validation of reasoning pathways increases confidence in generated responses, as supported by recent theoretical frameworks for agentic AI in healthcare (\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAn advantage of our neuro-symbolic approach is its potential to mitigate hallucinations, a phenomenon that represents the most significant barrier to clinical deployment of conventional LLMs. Studies have shown that large language models exhibit adversarial hallucination rates ranging from approximately 50% to 82% when presented with clinical vignettes containing fabricated details, and even the best-performing model reduced but did not eliminate these errors under mitigation strategies (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e). Asgari et al. (2025) reported a 1.47% hallucination rate and 3.45% omission rate in clinical note summarization tasks, emphasizing the need for robust safety frameworks (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e). Recent systematic analyses indicate that medical LLMs exhibit hallucination rates of 15% to 40% on clinical tasks, raising concerns about deployment readiness (\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e). Our symbolic knowledge base serves as a constraint-checking mechanism that validates candidate answers against hard-coded TG18 criteria, rejecting responses that violate established diagnostic and therapeutic guidelines. This architectural safeguard, combined with Monte Carlo dropout-based uncertainty quantification, provides multiple layers of protection against clinically consequential errors.\u003c/p\u003e \u003cp\u003eThe comparison with human expert physicians elucidated significant insights regarding the potential function of AI-assisted clinical decision support. Gastroenterology specialists attained exemplary performance, with a mean accuracy of 95.7%, approaching but not equaling the NS-LLM system, whereas emergency medicine physicians exhibited somewhat lower accuracy, with a mean of 84.2%, particularly in the domains of treatment decisions and severity grading. This performance differential between specialties underscores the significance of domain-specific expertise in managing complex hepatobiliary conditions and indicates that AI decision support systems could be especially beneficial in non-specialist environments where such expertise is less accessible. Notably, questions involving biliary access in surgically altered anatomy, anticoagulation management during urgent ERCP, and de-escalation criteria demonstrated the lowest inter-rater agreement among human experts, reflecting areas of ongoing clinical controversy not fully addressed by current guidelines.\u003c/p\u003e \u003cp\u003eThese findings align with Berry et al. (2025), who proposed a structured framework for integrating LLMs into gastroenterology practice, emphasizing the importance of multidisciplinary collaboration and continuous validation in real-world settings (\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe clinical implications of our findings extend beyond performance metrics to the broader question of how AI can be responsibly integrated into hepatobiliary clinical practice. Yuan et al. highlighted that agentic LLMs can access research findings, clinical case reports, and updated guidelines without additional training, enabling them to tackle complex tasks requiring iterative reasoning that align more closely with golden clinical procedures (\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e). Our system's ability to generate explainable reasoning chains that trace activated rules and supporting evidence addresses the interpretability requirements emphasized by Vidal et al. as essential for trustworthy deployment in medicine (\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e). The perfect accuracy achieved across different cholangitis etiologies, including relatively rare presentations such as parasitic cholangitis and post-procedural causes, suggests potential utility in educational settings and as a second-opinion tool for complex cases. However, as Soroush et al. cautioned in Gastroenterology, responsible deployment requires addressing challenges including output reliability, human-AI teaming, and infrastructure demands through comprehensive risk mitigation frameworks (\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOur results contribute to a rapidly evolving literature on AI performance in specialized medical domains. Gaber et al. evaluated LLM workflows in clinical decision support for triage and diagnosis, demonstrating that sophisticated prompting and retrieval strategies can substantially improve performance (\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e). The deficits observed in the treatment domain within our conventional LLM evaluation mirror the findings of Ong et al., who demonstrated that LLMs used as clinical decision support systems for medication safety exhibited variable performance across 16 clinical specialties (\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e). Ferber et al. developed an autonomous AI agent for oncology treatment planning that achieved significant improvements through tool-augmented reasoning, conceptually similar to our genetic algorithm optimization and symbolic constraint checking (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e). The convergence of these findings suggests that hybrid architectures combining neural flexibility with structured reasoning represent a promising direction for clinical AI development, particularly in domains requiring strict adherence to established guidelines.\u003c/p\u003e \u003cp\u003eSeveral limitations of this study merit acknowledgment. Firstly, the evaluation employed board-style multiple-choice questions rather than real-world clinical encounters, which may not comprehensively represent the intricacies of actual patient management, including scenarios with incomplete information, communication challenges, and time constraints. Secondly, the relatively small sample size of human participants, especially within the emergency medicine group (n\u0026thinsp;=\u0026thinsp;4), restricts the extent to which comparative conclusions can be generalized. Thirdly, the NS-LLM system's knowledge base was constructed solely on TG18, and its performance on cases outside these guidelines or requiring the integration of updated evidence remains unassessed. Fourthly, the study was conducted within a single country (Turkey), and the results may vary across different healthcare systems and patient populations. Fifthly, although the system attained perfect accuracy on the test set, the possibility of overfitting to the question format cannot be disregarded, and validation using novel cases in prospective studies is imperative. Lastly, the computational demands and latency of the multi-agent system were not systematically examined, which has significant implications for its application in real-time clinical settings implementation.\u003c/p\u003e \u003cp\u003eFuture research directions should address these limitations while exploring several promising avenues. Extension of the symbolic knowledge base to incorporate additional guidelines, emerging evidence, and local antimicrobial resistance patterns would enhance applicability. Integration with electronic health record systems for real-time clinical decision support warrants investigation, with attention to workflow integration and user acceptance. Multi-institutional prospective validation across diverse healthcare settings is essential to establish generalizability. Development of uncertainty communication interfaces that appropriately convey confidence levels to clinicians represents an important human factors consideration. Finally, comparative cost-effectiveness analyses examining the resource implications of deploying neuro-symbolic systems versus expanding specialist access would inform implementation decisions.\u003c/p\u003e \u003cp\u003e In conclusion, this study demonstrates that a genetic neuro-symbolic LLM system integrating multi-agent orchestration with explicit clinical guideline encoding achieves superior performance in acute cholangitis management compared to both conventional LLMs and human expert physicians. The architectural innovations of parallel neural reasoning, symbolic constraint validation, and consensus arbitration address fundamental limitations of purely neural approaches, including hallucination risk, guideline misapplication, and inconsistent multi-step reasoning. These findings establish a new benchmark for AI-assisted clinical decision support in complex hepatobiliary disease and provide a template for developing similar systems across other guideline-driven medical domains. As the healthcare community increasingly explores AI integration, neuro-symbolic architectures offer a promising pathway toward trustworthy, explainable, and clinically reliable decision support systems.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eDiagnostic Conclusion\u003c/h2\u003e \u003cp\u003eDefinite Acute Cholangitis (all three domains positive)\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eSeverity Conclusion\u003c/h2\u003e \u003cp\u003eGrade III (Severe) - Renal dysfunction criterion met (Cr\u0026thinsp;\u0026gt;\u0026thinsp;2.0 mg/dL)\u003c/p\u003e \u003c/p\u003e\u003cp\u003e \u003ch2\u003eConflict of Interest\u003c/h2\u003e \u003cp\u003eThe authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eM.U: Conceptualization, Methodology, Formal analysis, Data curation, Writing \u0026ndash; Original Draft, Visualization.E.E: Supervision, Project administration, Writing \u0026ndash; Review \u0026amp;amp; Editing. All authors have read and approved the final version of the manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe datasets generated and analyzed during the current study, including the 30 case-based questions, AI model responses, and human expert responses, are available in the Supplementary Materials.\u003c/p\u003e\n\u003cp\u003e \u003cb\u003eEthics Statement\u003c/b\u003e \u003c/p\u003e \u003cp\u003e This study was conducted in accordance with the Declaration of Helsinki. The Non-Interventional Ethics Committee of the Ankara Provincial Health Directorate approved the study protocol (Approval No: 2025-11-15, dated October 15, 2025). All participating physicians provided written informed consent prior to study enrollment. As this study utilized hypothetical clinical vignettes and did not involve any real patient data or direct patient care, patient consent was not applicable. The ethics committee confirmed that patient consent was not required given the simulation-based nature of the study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGaber F, Shaik M, Allega F, Bilecz AJ, Busch F, Goon K, et al. Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis. npj Digit Med. 2025;8(1):263.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWiest IC, Bhat M, Clusmann J, Schneider CV, Jiang X, Kather JN. Large language models for clinical decision support in gastroenterology and hepatology. Nat Rev Gastroenterol Hepatol. 2025;22(11):773\u0026ndash;87.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim J, Podlasek A, Shidara K, Liu F, Alaa A, Bernardo D. Limitations of large language models in clinical problem-solving arising from inflexible reasoning. Sci Rep. 2025;15(1):39426.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi H, Fu J-F, Python A. Implementing Large Language Models in Health Care: Clinician-Focused Review With Interactive Guideline. J Med Internet Res. 2025;27:e71916.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCozma M-A, Găman M-A, Srichawla BS, Dhali A, Manan MR, Nahian A, et al. Acute cholangitis: a state-of-the-art review. Annals Med Surg. 2024;86(8):4560\u0026ndash;74.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKiriyama S, Kozaka K, Takada T, Strasberg SM, Pitt HA, Gabata T, et al. Tokyo Guidelines 2018: diagnostic criteria and severity grading of acute cholangitis (with videos). J Hepato-Biliary-Pancreat Sci. 2018;25(1):17\u0026ndash;30.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYu E, Chu X, Zhang W, Meng X, Yang Y, Ji X, et al. Large Language Models in Medicine: Applications, Challenges, and Future Directions. Int J Med Sci. 2025;22(11):2792\u0026ndash;801.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMeziane L, Abbaoui W, Abdellaoui S, El Bhiri B, Ziti S. Narrative Review on Symbolic Approaches for Explainable Artificial Intelligence: Foundations, Challenges, and Perspectives. Engineering Proceedings [Internet]. 2025; 112(1):[39 p.].\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVidal M-E, Chudasama Y, Huang H, Purohit D, Torrente M. Integrating Knowledge Graphs with Symbolic AI: The Path to Interpretable Hybrid AI Systems in Medicine. J Web Semant. 2025;84:100856.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrenosil GA, Weitzel TK, Bello SC, Mingels C, Manzini G, Meier LP, et al. Neuro-symbolic AI for auditable cognitive information extraction from medical reports. Commun Med (Lond). 2025;5(1):491.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNawaz U, Anees-ur-Rahaman M, Saeed Z. A review of neuro-symbolic AI integrating reasoning and learning for advanced cognitive systems. Intell Syst Appl. 2025;26:200541.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ehttps://\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003edocs.langchain.com/oss/python/langgraph/graph-api?utm_source=chatgpt.com\u003c/span\u003e\u003cspan address=\"http://docs.langchain.com/oss/python/langgraph/graph-api?utm_source=chatgpt.com\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e [.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSafavi-Naini SAA, Ali S, Shahab O, Shahhoseini Z, Savage T, Rafiee S, et al. Benchmarking proprietary and open-source language and vision-language models for gastroenterology clinical reasoning. NPJ Digit Med. 2025;8(1):797.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIrfan B, Sirvent R. Large language models and the future of gastroenterology: dissecting the biopolitics of data in a global health ecosystem. Front Med. 2025;Volume 12\u0026ndash;2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrenosil GA, Weitzel TK, Bello SC, Mingels C, Manzini G, Meier LP, et al. Neuro-symbolic AI for auditable cognitive information extraction from medical reports. Commun Med. 2025;5(1):491.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim J, Podlasek A, Shidara K, Liu F, Alaa A, Bernardo D. Limitations of large language models in clinical problem-solving arising from inflexible reasoning. Sci Rep. 2025;15(1):39426.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGorenshtein A, Omar M, Glicksberg BS, Nadkarni GN, Klang E. AI Agents in Clinical Medicine: A Systematic Review. medRxiv. 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen X, Yi H, You M, Liu W, Wang L, Li H, et al. Enhancing diagnostic capability with multi-agents conversational large language models. npj Digit Med. 2025;8(1):159.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHinostroza Fuentes VG, Karim HA, Tan MJT, AlDahoul N. AI with agency: a vision for adaptive, efficient, and ethical healthcare. Front Digit Health. 2025;Volume 7\u0026ndash;2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBorkowski AA, Ben-Ari A, Multiagent. AI Systems in Health Care: Envisioning Next-Generation Intelligence. Fed Pract. 2025;42(5):188\u0026ndash;94.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOmar M, Sorin V, Collins JD, Reich D, Freeman R, Gavin N et al. Large Language Models Are Highly Vulnerable to Adversarial Hallucination Attacks in Clinical Decision Support: A Multi-Model Assurance Analysis. medRxiv. 2025:2025.03.18.25324184.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBerry P, Dhanakshirur RR, Khanna S. Utilizing large language models for gastroenterology research: a conceptual framework. Th Adv Gastroenterol. 2025;18:17562848251328577.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXu X, Sankar R. Large Language Model Agents for Biomedicine: A Comprehensive Review of Methods, Evaluations, Challenges, and Future Directions. Information. 2025;16(10):894.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChudasama Y, Huang H, Purohit D, Vidal ME. Toward Interpretable Hybrid AI: Integrating Knowledge Graphs and Symbolic Reasoning in Medicine. IEEE Access. 2025;13:39489\u0026ndash;509.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoroush A, Giuffr\u0026egrave; M, Chung S, Shung DL. Generative Artificial Intelligence in Clinical Medicine and Impact on Gastroenterology. Gastroenterology. 2025;169(3):502\u0026ndash;e171.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGaber F, Shaik M, Allega F, Bilecz AJ, Busch F, Goon K, et al. Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis. NPJ Digit Med. 2025;8(1):263.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOng JCL, Jin L, Elangovan K, Lim GYS, Lim DYZ, Sng GGR, et al. Large language model as clinical decision support system augments medication safety in 16 clinical specialties. Cell Rep Med. 2025;6(10):102323.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFerber D, El Nahhas OSM, W\u0026ouml;lflein G, Wiest IC, Clusmann J, Le\u0026szlig;mann ME, et al. Development and validation of an autonomous artificial intelligence agent for clinical decision-making in oncology. Nat Cancer. 2025;6(8):1337\u0026ndash;49.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1.\u0026nbsp;\u003c/strong\u003eDemographic and Professional Characteristics of Human Expert Participants\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"100%\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCharacteristic\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eGastroenterology (n=10)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eEmergency Medicine (n=4)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eTotal (n=14)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u003cstrong\u003ep-value\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eClinical experience, years (mean \u0026plusmn; SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e16.0 \u0026plusmn; 3.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e11.8 \u0026plusmn; 1.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e14.8 \u0026plusmn; 3.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e0.048*\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAge, years (mean \u0026plusmn; SD)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e45.2 \u0026plusmn; 4.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e40.5 \u0026plusmn; 2.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e43.9 \u0026plusmn; 4.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e0.067\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eSex, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e0.530\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Male\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e7 (70.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e3 (75.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e10 (71.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Female\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e3 (30.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e1 (25.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e4 (28.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAcademic title, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e0.089\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Professor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e4 (40.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e0 (0.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e4 (28.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Associate Professor\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e4 (40.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e0 (0.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e4 (28.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Specialist Physician\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e2 (20.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e4 (100.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e6 (42.9)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u003cstrong\u003eInstitution, n (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026mdash;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Ankara Bilkent City Hospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e3 (30.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e1 (25.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e4 (28.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Ankara Etlik City Hospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e3 (30.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e1 (25.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e4 (28.6)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Elazığ Fethi Sekin City Hospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e2 (20.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e1 (25.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e3 (21.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 233px;\"\u003e\n \u003cp\u003e\u0026nbsp; \u0026nbsp; Etimesgut Şehit Sait Ert\u0026uuml;rk State Hospital\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e2 (20.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 133px;\"\u003e\n \u003cp\u003e1 (25.0)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\n \u003cp\u003e3 (21.4)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 67px;\"\u003e\n \u003cp\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*Statistically significant (p\u0026lt;0.05).\u003c/p\u003e\n\u003cp\u003eContinuous variables are presented as mean \u0026plusmn; standard deviation and compared using independent samples t-test. Categorical variables are presented as n (%) and compared using Fisher\u0026apos;s exact test.\u003c/p\u003e\n\u003cp\u003eAbbreviations: SD, standard deviation.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2.\u0026nbsp;\u003c/strong\u003eOverall Diagnostic Accuracy of AI Models and Human Expert Groups in Acute Cholangitis Management Based on Tokyo Guidelines 2018\u003c/p\u003e\n\u003ctable style=\"width:100.0%;border-collapse:collapse;border:none;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:140.0pt;border-top:solid black 1.5pt;border-left: none white 1.0pt;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003eParticipant / Model\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003eCorrect Answers (n)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003eAccuracy (%)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:80.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003e95% CI\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003ep-value vs NS-LLM\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003ep-value vs Gastro\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:500.0pt;border-top:none;border-left:none white 1.0pt;border-bottom:solid #CCCCCC 1.0pt;border-right:none white 1.0pt;background:#E8E8E8;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003eArtificial Intelligence Models\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"border:none;border-bottom:none white 1.0pt;\" colspan=\"5\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:140.0pt;border:none white 1.0pt;border-top:none;background: #E8F5E9;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003eNeuro-Symbolic LLM\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#E8F5E9;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003e30\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#E8F5E9;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:80.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;background:#E8F5E9;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003e88.4\u0026ndash;100.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#E8F5E9;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003eReference\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#E8F5E9;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003e0.018*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:140.0pt;border:none white 1.0pt;border-top:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:13px;\"\u003eClaude 4.5 Sonnet\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e27\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e90.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:80.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e73.5\u0026ndash;97.9\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e0.042*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e0.089\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:140.0pt;border:none white 1.0pt;border-top:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:13px;\"\u003eGemini 2.0 Flash\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e19\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e63.3\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:80.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e43.9\u0026ndash;80.1\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e\u0026lt;0.001*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e\u0026lt;0.001*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:140.0pt;border:none white 1.0pt;border-top:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:13px;\"\u003eChatGPT 5.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e18\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e60.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:80.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e40.6\u0026ndash;77.3\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e\u0026lt;0.001*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e\u0026lt;0.001*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:500.0pt;border-top:none;border-left:none white 1.0pt;border-bottom:solid #CCCCCC 1.0pt;border-right:none white 1.0pt;background:#E8E8E8;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;color:black;\"\u003eHuman Expert Groups\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"border:none;border-bottom:none white 1.0pt;\" colspan=\"5\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:140.0pt;border:none white 1.0pt;border-top:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003eGastroenterology Specialists (n=10)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e28.7 \u0026plusmn; 0.9\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003e95.7 \u0026plusmn; 3.2\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:80.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e92.8\u0026ndash;98.6\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e0.018*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003eReference\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:140.0pt;border-top:none;border-left:none white 1.0pt;border-bottom:solid black 1.5pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003eEmergency Medicine Specialists (n=4)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e25.3 \u0026plusmn; 2.6\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:13px;\"\u003e84.2 \u0026plusmn; 8.8\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:80.0pt;border-top:none;border-left:none;border-bottom:solid black 1.5pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e70.2\u0026ndash;98.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e0.003*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:70.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;padding:2.5pt 4.0pt 2.5pt 4.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:13px;\"\u003e0.012*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*Statistically significant (p\u0026lt;0.05).\u003c/p\u003e\n\u003cp\u003eAI model accuracy is based on single assessment (30 questions). Human expert data are presented as mean \u0026plusmn; standard deviation. Pairwise comparisons between AI models were performed using McNemar\u0026apos;s test. Comparisons involving human expert groups were performed using chi-square test.\u003c/p\u003e\n\u003cp\u003eAbbreviations: NS-LLM, neuro-symbolic large language model; CI, confidence interval; Gastro, gastroenterology specialists.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3.\u0026nbsp;\u003c/strong\u003ePerformance Stratified by Clinical Domain, TG18 Severity Grade, and Etiology (Accuracy %)\u003c/p\u003e\n\u003ctable style=\"width:100.0%;border-collapse:collapse;border:none;\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border-top:solid black 1.5pt;border-left: none white 1.0pt;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003eCategory\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003eNS-LLM\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003eClaude 4.5\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003eChatGPT 5.2\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003eGemini 2.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003eGastroenterology (mean \u0026plusmn; SD)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003eEmergency Med (mean \u0026plusmn; SD)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-top:solid black 1.5pt;border-left: none;border-bottom:solid black 1.0pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;\"\u003ep-value (GE vs EM)\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:590.0pt;border-top:none;border-left:none white 1.0pt;border-bottom:solid #CCCCCC 1.0pt;border-right:none white 1.0pt;background:#E3F2FD;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003eClinical Domain\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"border:none;border-bottom:none white 1.0pt;\" colspan=\"7\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Diagnosis (n=10)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e90.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e70.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e70.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e96.0 \u0026plusmn; 5.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e82.5 \u0026plusmn; 9.6\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.018*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Treatment (n=10)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e90.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e50.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e50.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e96.0 \u0026plusmn; 5.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e80.0 \u0026plusmn; 8.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.003*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Complications/Prognosis (n=10)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e90.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFE0B2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e60.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e70.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e95.0 \u0026plusmn; 7.1\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e90.0 \u0026plusmn; 8.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.156\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:590.0pt;border-top:none;border-left:none white 1.0pt;border-bottom:solid #CCCCCC 1.0pt;border-right:none white 1.0pt;background:#E3F2FD;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003eTG18 Severity Grade\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"border:none;border-bottom:none white 1.0pt;\" colspan=\"7\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Grade I \u0026ndash; Mild (n=8)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e87.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e62.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e62.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e93.8 \u0026plusmn; 6.3\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e81.3 \u0026plusmn; 12.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.042*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Grade II \u0026ndash; Moderate (n=12)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e91.7\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e58.3\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e66.7\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e95.8 \u0026plusmn; 4.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e83.3 \u0026plusmn; 9.1\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.008*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Grade III \u0026ndash; Severe (n=10)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e90.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFE0B2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e60.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFE0B2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e60.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e97.0 \u0026plusmn; 4.8\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e87.5 \u0026plusmn; 9.6\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.067\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:590.0pt;border-top:none;border-left:none white 1.0pt;border-bottom:solid #CCCCCC 1.0pt;border-right:none white 1.0pt;background:#E3F2FD;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003eEtiology\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"border:none;border-bottom:none white 1.0pt;\" colspan=\"7\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u0026nbsp;\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Choledocholithiasis (n=12)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e91.7\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e66.7\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e66.7\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e97.5 \u0026plusmn; 3.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e87.5 \u0026plusmn; 7.2\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.012*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Malignant Biliary Obstruction (n=6)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e83.3\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e50.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e66.7\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e95.0 \u0026plusmn; 5.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e83.3 \u0026plusmn; 10.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.034*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Post-procedural/Iatrogenic (n=6)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e66.7\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e50.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e95.0 \u0026plusmn; 5.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e79.2 \u0026plusmn; 12.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.021*\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border:none white 1.0pt;border-top:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Benign Strictures (n=4)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e75.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e50.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;background:#FFF9C4;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e75.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e92.5 \u0026plusmn; 9.6\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-bottom:white 1.0pt;border-right: white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e81.3 \u0026plusmn; 12.5\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-bottom:white 1.0pt;border-right:white 1.0pt;border-style:none;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.089\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width:130.0pt;border-top:none;border-left:none white 1.0pt;border-bottom:solid black 1.5pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;'\u003e\u003cspan style=\"font-size:12px;\"\u003e\u0026nbsp; \u0026nbsp; Parasitic Cholangitis (n=2)\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cstrong\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;background:#C8E6C9;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e100.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e50.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:55.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;background:#FFCDD2;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;color:black;\"\u003e50.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-top:none;border-left:none;border-bottom:solid black 1.5pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e95.0 \u0026plusmn; 10.0\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:90.0pt;border-top:none;border-left:none;border-bottom:solid black 1.5pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e75.0 \u0026plusmn; 20.4\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width:60.0pt;border-top:none;border-left:none;border-bottom: solid black 1.5pt;border-right:none white 1.0pt;padding:2.0pt 3.0pt 2.0pt 3.0pt;\"\u003e\n \u003cp style='margin:0cm;font-size:16px;font-family:\"Times New Roman\",serif;text-align:center;'\u003e\u003cspan style=\"font-size:12px;\"\u003e0.156\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e*Statistically significant (p\u0026lt;0.05).\u003c/p\u003e\n\u003cp\u003eAI model data are presented as accuracy percentages. Human expert data are presented as mean \u0026plusmn; standard deviation. Statistical comparisons between gastroenterology (GE) and emergency medicine (EM) specialist groups were performed using independent samples t-test.\u003c/p\u003e\n\u003cp\u003eColor coding: Green (\u0026gt;95%), Yellow (65\u0026ndash;85%), Red (\u0026lt;65%).\u003c/p\u003e\n\u003cp\u003eAbbreviations: NS-LLM, neuro-symbolic large language model; TG18, Tokyo Guidelines 2018; SD, standard deviation.\u003c/p\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":"","lastPublishedDoi":"10.21203/rs.3.rs-8714103/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8714103/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\u003eLarge language models (LLMs) have shown promising results in medical decision support; Background: Large language models (LLMs) have demonstrated promising outcomes in medical decision support; however, their efficacy in managing complex hepatobiliary conditions remains insufficiently examined. We have developed a genetic neuro-symbolic LLM system that integrates multiple AI agents with neural-symbolic reasoning for the management of cholangitis, and we have compared its performance to that of conventional LLMs and human experts.genetic neuro-symbolic LLM system integrating multiple AI agents with neural-symbolic reasoning for cholangitis management and compared its performance against conventional LLMs and human experts.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis multi-center cross-sectional study included 30 case-based questions from American Board of Internal Medicine (ABIM) gastroenterology subspecialty examinations covering acute cholangitis. Questions were categorized into diagnosis (n = 10), treatment (n = 10), and complications/prognosis (n = 10). Performance of a genetic neuro-symbolic LLM system orchestrated via LangGraph was compared against Claude 4.5 Sonnet, ChatGPT 5.2, Gemini 3 Pro, 10 gastroenterology specialists, and 4 emergency medicine physicians from four tertiary centers in Turkey.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe genetic neuro-symbolic system achieved the highest overall accuracy (100%, 30/30), significantly outperforming Claude 4.5 Sonnet (90.0%), ChatGPT 5.2 (53.3%), Gemini 3 Pro (56.7%), gastroenterology experts (mean 96.3% ± 2.1%), and emergency medicine physicians (mean 89.2% ± 4.8%). The neuro-symbolic system demonstrated superior performance across all categories and cholangitis subtypes. Among human participants, gastroenterologists outperformed emergency physicians in treatment decisions (p = 0.012) and showed non-inferior performance to Gemini 3 Pro overall (p = 0.034). ROC analysis revealed excellent discrimination for the neuro-symbolic system (AUC = 1.000) compared to Claude (AUC = 0.924), ChatGPT (AUC = 0.687), and Gemini (AUC = 0.712).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe genetic neuro-symbolic LLM system demonstrated superior accuracy in cholangitis management compared to all conventional AI models and human experts. This multi-agent architecture with neural-symbolic reasoning represents a significant advancement in AI-assisted clinical decision support for complex hepatobiliary conditions.\u003c/p\u003e","manuscriptTitle":"Performance Comparison of a Neuro-Symbolic Large Language Model System Versus Conventional AI Models and Human Experts in Cholangitis Management","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-16 17:13:50","doi":"10.21203/rs.3.rs-8714103/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-27T09:49:48+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-15T08:48:11+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"68647961974482625327066978208054787587","date":"2026-03-05T04:49:14+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-03T18:40:10+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"276098569161781314538619437143311737023","date":"2026-02-10T15:57:32+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-02-10T15:39:22+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-02-03T06:19:11+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-29T14:45:46+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-29T14:42:42+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Medical Informatics and Decision Making","date":"2026-01-27T20:16:39+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":"February 16th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-15T09:53:46+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-16 17:13:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8714103","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8714103","identity":"rs-8714103","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}