Neurosymbolic Multi-Agent Large Language Model System Versus Specialist Physicians in COPD and Asthma Management: A Comparative Performance Evaluation Using Guideline-Based Clinical Vignettes

Neurosymbolic Multi-Agent Large Language Model System Versus Specialist Physicians in COPD and Asthma Management: A Comparative Performance Evaluation Using Guideline-Based Clinical Vignettes | 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 Article Neurosymbolic Multi-Agent Large Language Model System Versus Specialist Physicians in COPD and Asthma Management: A Comparative Performance Evaluation Using Guideline-Based Clinical Vignettes Ebru Aykan Mavigoz, Mete Ucdal, Evren Ekingen This is a preprint; it has not been peer reviewed by a journal.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-9262455/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 8 You are reading this latest preprint version Abstract Background Chronic obstructive pulmonary disease (COPD) and asthma collectively affect over 500 million individuals worldwide and represent leading causes of respiratory-related morbidity, healthcare utilization, and preventable mortality. Optimal management of these conditions encompasses accurate diagnosis and differential diagnosis, severity stratification, stepwise pharmacological therapy, exacerbation management, ventilatory support decisions, and complication recognition, all of which require the integration of subspecialty expertise that is frequently fragmented across clinical settings. Neurosymbolic multi-agent large language model (LLM) systems, which integrate neural reasoning with formal symbolic logic and specialized agent architectures, represent a promising paradigm for bridging domain-specific expertise gaps across pulmonology, allergology, and emergency medicine. Methods We conducted a cross-sectional comparative evaluation using 30 USMLE-format clinical vignettes spanning six COPD and asthma management domains, namely diagnosis and differential diagnosis, severity assessment, pharmacological management, exacerbation management, ICU admission and ventilatory management, and complication management, all derived from the current GOLD 2025COPD and GINA 2025 Asthma guidelines. A neurosymbolic multi-agent system (NS-MAS) comprising a LangGraph orchestrator coordinating eight specialized agents, including GPT-4.5, Claude Sonnet 4.6, and Gemini 2.5 Pro, was compared against 30 board-certified specialist physicians consisting of 10 pulmonologists, 10 allergists and clinical immunologists, and 10 emergency medicine physicians. The primary outcome was overall accuracy. Statistical analyses included the Kruskal-Wallis test, Dunn's post-hoc test with Bonferroni correction, and effect size estimation using Cohen's d. Results NS-MAS achieved an overall accuracy of 90.0% (27/30), significantly exceeding all physician groups: pulmonologists 75.0% (mean 22.5 ± 2.1/30, p = 0.002, d = 1.82), allergists/immunologists 76.0% (mean 22.8 ± 2.2/30, p = 0.003, d = 1.74), and emergency physicians 62.0% (mean 18.6 ± 2.0/30, p < 0.001, d = 2.84). Domain-specific analysis revealed significant specialty-expertise interactions: allergists outperformed pulmonologists in exacerbation management (84% vs 72%, p = 0.003) and pharmacological management (82% vs 76%, p = 0.018), while pulmonologists demonstrated superior performance in ICU/ventilatory management (66% vs 58%, p < 0.001) and complication management domains (66% vs 62%, p = 0.028). NS-MAS response latency (median 13.1 seconds) was significantly shorter than all physician groups (median 2.9–3.5 minutes, p < 0.001). Inter-rater reliability among physicians was fair to moderate (κ = 0.36–0.49). Conclusions A neurosymbolic multi-agent LLM system demonstrated comprehensive, guideline-concordant COPD and asthma management performance surpassing all specialist physician groups across nearly all clinical domains. The system's ability to integrate domain-specific reasoning across pulmonology, allergology, and emergency medicine knowledge bases, guided simultaneously by the GOLD 2025and GINA 2025 frameworks, suggests meaningful clinical decision support potential, particularly in resource-limited or cross-specialty settings. These findings support further prospective clinical validation. Health sciences/Diseases Health sciences/Health care Health sciences/Medical research large language model multi-agent system COPD asthma clinical decision support artificial intelligence guideline adherence GOLD GINA Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 1. Introduction Chronic obstructive pulmonary disease (COPD) and asthma are the most prevalent chronic airway diseases worldwide. Recent Global Burden of Disease-based analyses indicate that asthma accounted for approximately 260.48 million prevalent cases and 436,190 deaths in 2021, whereas COPD accounted for approximately 213.4 million prevalent cases and 3.7 million deaths in the same year [ 1 , 2 ]. Thinking about shared diagnostic and management frameworks between COPD and asthma despite mechanistic differences.Despite important mechanistic differences, with COPD typically featuring persistent airflow limitation related to noxious exposure-driven, often neutrophil-predominant inflammation and emphysematous or small-airway remodeling, and asthma typically featuring variable expiratory airflow limitation with type 2 or eosinophilic inflammation in many patients, both conditions share core management imperatives: accurate objective diagnosis, structured severity and risk assessment, personalized pharmacologic optimization, and early recognition and prevention of exacerbations [ 3 , 4 ]. The clinical complexity of obstructive airway disease management is compounded by substantial phenotypic heterogeneity. COPD includes emphysema-predominant, chronic bronchitis-predominant, and frequent-exacerbator phenotypes with distinct therapeutic implications, whereas asthma spans a spectrum from infrequent, mild symptoms to severe disease requiring Step 5 specialist assessment, phenotyping, and add-on therapy [ 5 ].In patients with coexisting features of asthma and COPD, management is particularly challenging; however, GOLD 2025 no longer endorses the term “asthma-COPD overlap (ACO)” and recommends that, when concurrent asthma is suspected, pharmacotherapy should primarily follow asthma guidelines while COPD-specific assessment and management are also addressed [ 6 ]. These conditions demand sophisticated decision-making across multiple domains: objective confirmation and interpretation of airflow limitation with consideration of alternative cardiac and upper-airway diagnoses; GOLD ABE assessment and GINA step-based treatment stratification; selection of inhaled regimens (SAMA, SABA, LAMA, LABA, ICS) tailored to symptoms, exacerbation risk, phenotype, and comorbidity burden; acute exacerbation management including decisions regarding systemic corticosteroids and antibiotics; and escalation to ventilatory support, including non-invasive ventilation for acute hypercapnic respiratory failure in COPD and invasive mechanical ventilation in near-fatal asthma [ 7 ]. Despite internationally endorsed guidelines, specifically the 2025 GOLD Report and the 2025 GINA Strategy Report, substantial inter-specialty variability in guideline adherence and clinical accuracy persists. Pulmonologists, allergists and clinical immunologists, and emergency physicians each bring complementary but non-overlapping expertise: pulmonologists excel in spirometric interpretation and critical care integration yet may underperform in allergy-driven severe asthma biologic selection; allergists demonstrate superior competence in phenotype-driven asthma management yet may lag in COPD-specific interventional and ventilatory decision-making; emergency physicians are adept at acute exacerbation stabilization but may apply suboptimal chronic maintenance therapy recommendations [ 8 ]. Artificial intelligence (AI) and large language model (LLM) systems have emerged as potentially transformative tools in clinical decision support [ 9 ]. Early benchmarking studies demonstrated that frontier LLMs can achieve expert-level performance on standardized medical examinations [ 10 ]. Multi-agent LLM systems, in which specialized agents are orchestrated to collaboratively reason across problem sub-domains, have shown performance improvements over monolithic LLM approaches in complex clinical reasoning tasks [ 11 , 12 ]. Neurosymbolic AI, which combines neural pattern recognition with formal symbolic reasoning, rule-based inference, and modular agent architectures, offers transparency, rule-adherence, and verifiable guideline consistency that renders it particularly suited to obstructive airway disease management, given that algorithmic stepwise decision frameworks are central to both the GOLD and GINA guideline [ 13 ]. Thinking about cholesterol management and dyslipidemia treatment approachesNotably, comparable AI-assisted decision support systems have previously demonstrated consistent guideline adherence in other chronic disease domains governed by stepwise algorithmic frameworks, such as cholecystitis management and biliary disease treatment optimization, suggesting that the neurosymbolic multi-agent paradigm may generalize across guideline-driven clinical specialties[ 14 ]. To our knowledge, no study has systematically compared a neurosymbolic multi-agent LLM system against board-certified specialist physicians across the full spectrum of COPD and asthma management domains, from initial diagnosis through exacerbation and complication management. The present study addresses this gap by conducting a rigorous comparative evaluation using guideline-derived clinical vignettes, with the explicit aim of characterizing both overall performance and domain-specific strengths and limitations of AI versus specialist physician decision-making in obstructive airway disease management. 2. Methods 2.1 Study Design This was a cross-sectional, comparative evaluation study conducted between October 2025 and March 2026. The study was approved by the Ankara Provincial Directorate of Public Health Non-Interventional Ethics Board (Approval No: 2025-10-3, Date: 24 October 2025) and was classified as non-interventional in accordance with Turkish Ministry of Health regulations, given that no patient data were collected or utilized and all physician participants provided written informed consent prior to participation. The study is reported in accordance with the GREET (Guideline for Reporting of Evaluation of AI in Healthcare) framework. The overall participant flow and study design are illustrated in Fig. 1 . 2.2 Clinical Case Vignette Development Thirty USMLE Step 2 CK-format clinical vignettes were developed through a rigorous multi-stage process. An expert committee comprising two pulmonologists, one allergist/clinical immunologist, and one emergency physician with expertise in AI clinical benchmarking (the study co-investigators) initially generated candidate questions across six pre-specified COPD/asthma management domains. 2.2.1 Clinical Domains The six domains were defined based on primary clinical decision nodes identified in current international guidelines: (1) Diagnosis and Differential Diagnosis — spirometric confirmation and pattern interpretation (obstructive, mixed), discrimination from cardiac causes of dyspnoea (heart failure, pulmonary hypertension), pulmonary embolism, upper airway obstruction, interstitial lung disease, and bronchiectasis; correct COPD-asthma overlap (ACO) identification; (2) Severity Assessment — application of GOLD ABCD/ABE grading system, mMRC dyspnoea scale, CAT score, spirometric severity (GOLD 1–4), GINA symptom control classification (well-controlled, partly controlled, uncontrolled), NAEPP asthma severity classification, and acute exacerbation severity grading (COPD: Anthonisen/GOLD exacerbation criteria; Asthma: BTS/SIGN acute asthma severity); (3) Pharmacological Management — guideline-concordant inhaler therapy for COPD (SAMA, SABA, LAMA, LABA, ICS/LABA, LAMA/LABA/ICS triple therapy; roflumilast; azithromycin prophylaxis) and asthma (SABA reliever, ICS controller, ICS/LABA, SMART therapy, add-on biologic therapy for severe eosinophilic asthma including dupilumab, mepolizumab, benralizumab, omalizumab); (4) Exacerbation Management — correct classification and pharmacological management of COPD exacerbations (SABA/SAMA nebulization, systemic corticosteroids, antibiotic decision-making per Anthonisen criteria, oxygen titration in type 2 respiratory failure), and acute severe/near-fatal asthma management (IV magnesium sulfate, IV/nebulized salbutamol, helium-oxygen, NIV indications, intubation thresholds); (5) ICU Admission and Ventilatory Management — indications for NIV in COPD exacerbation-related hypercapnic respiratory failure (pH thresholds, Pco₂ targets), invasive mechanical ventilation for near-fatal asthma (dynamic hyperinflation management, permissive hypercapnia), weaning protocols, and extubation readiness criteria; and (6) Complication Management — recognition and management of cor pulmonale (right heart failure in COPD), secondary pneumothorax in COPD/asthma, COPD-related hypercapnic encephalopathy, mucus plugging with lobar collapse, and adrenal insufficiency from chronic corticosteroid use. 2.2.2 Vignette Validation Each domain comprised five vignettes (total n = 30), each presenting a detailed clinical scenario including patient demographics, symptom history, physical examination findings, pulmonary function data, laboratory results, and imaging findings where applicable. All vignettes were structured as five-option single-best-answer questions. A correct answer key was developed by three independent content experts and verified against primary guideline sources (GOLD 2025, GINA 2025, BTS/SIGN Asthma Guidelines 2023). Vignettes achieving less than 90% inter-expert agreement (n = 2) were revised iteratively until consensus was achieved. The complete vignette bank is provided in Supplementary Material 1. 2.3 Neurosymbolic Multi-Agent System (NS-MAS) 2.3.1 System Architecture The NS-MAS was developed using LangGraph (version 0.2.x, LangChain Inc.) as the state-machine orchestration framework. The system comprised eight specialized functional agents plus the orchestration layer: (i) Orchestrator Agent (LangGraph) : Stateful graph-based router responsible for query parsing, agent selection, reasoning chain construction, and conflict resolution; (ii) Pulmonology/Respiratory Agent (GPT-4.5, OpenAI) : Specialized for COPD diagnosis, spirometric interpretation, emphysema phenotyping, GOLD staging, critical care ventilatory decisions, and ACO management. RAG-augmented with the full GOLD 2025 Report; (iii) Allergy/Immunology Agent (Claude Sonnet 4.6, Anthropic) : Specialized for asthma phenotyping, GINA stepwise management, T2 biomarker interpretation (FeNO, blood eosinophils, IgE), biologic therapy selection, and allergen-triggered exacerbation management. RAG-augmented with GINA 2025 Strategy Report and BTS/SIGN Asthma Guidelines 2023; (iv) Emergency Medicine Agent (Gemini 2.5 Pro, Google DeepMind) : Specialized for acute COPD exacerbation management, acute severe asthma in the emergency department, rapid severity assessment, and time-sensitive intervention decisions; (v) Pharmacology/Therapeutics Agent (GPT-4.5, OpenAI) : Specialized for inhaler device selection, drug-drug interactions, dose optimization (PK/PD), corticosteroid-sparing strategies, and biologic prescribing criteria; (vi) Evidence Synthesis Agent (GPT-4.5 + RAG, OpenAI) : Retrieves and synthesizes relevant guideline text and clinical trial evidence (TORCH, UPLIFT, SUMMIT, IMPACT, ETHOS, NAVIGATOR, MEDI3506 trials); (vii) Radiology Interpretation Agent (GPT-4o with vision, OpenAI) : Specialized for chest radiograph and CT pattern recognition including emphysema quantification, air trapping, mucus plugging, and pneumothorax identification; (viii) Critical Care/ICU Agent (Claude 3.5 Sonnet, Anthropic) : Specialized for ventilatory management, dynamic hyperinflation, permissive hypercapnia strategies, and hemodynamic management in cor pulmonale. 2.3.2 Symbolic Reasoning Layer Following multi-agent deliberation, outputs were processed by a Prolog-based symbolic reasoning engine implementing a COPD/asthma management knowledge base comprising 312 formal rules derived from GOLD 2025 and GINA 2025 guideline decision trees. The symbolic layer performed: (a) logical consistency checking between agent outputs; (b) rule-based correction of responses violating formal guideline logic (e.g., initiating ICS monotherapy in COPD without asthma overlap or eosinophilic phenotype; prescribing β-blocker-containing eyedrops without asthma safety precaution documentation); and (c) confidence score assignment. Full architectural details are provided in Supplementary Material 2. 2.4 Physician Participants Board-certified specialist physicians were recruited through purposive sampling from multiple tertiary care hospitals in the Ankara region between November 2025 and January 2026, in a multicenter design. Participating institutions and physician contributions were as follows: Ankara Bilkent City Hospital (n = 10; 4 pulmonologists, 3 allergists and clinical immunologists, 3 emergency medicine physicians), Ankara Etlik City Hospital (n = 10; 3 pulmonologists, 4 allergists and clinical immunologists, 3 emergency medicine physicians), Mamak State Hospital (n = 5; 2 pulmonologists, 1 allergist and clinical immunologist, 2 emergency medicine physicians), and Etimesgut Şehit Sait Ertürk State Hospital (n = 5; 1 pulmonologist, 2 allergists and clinical immunologists, 2 emergency medicine physicians). Eligibility criteria included board certification in the relevant specialty (pulmonology, allergology and clinical immunology, or emergency medicine), a minimum of three years of post-residency clinical practice, active clinical practice with patient contact in the preceding 12 months, and ability to read and respond to English-language clinical vignettes. Thirty physicians were enrolled in total: 10 pulmonologists, 10 allergists and clinical immunologists, and 10 emergency medicine physicians. Demographic and professional characteristics were collected by self-administered questionnaire. Physicians completed the 30-vignette assessment individually, without access to reference materials, under randomized domain order to mitigate sequencing effects. 2.5 Outcome Measures The primary outcome was overall accuracy (proportion of correctly answered vignettes out of 30). Secondary outcomes included: (1) domain-specific accuracy; (2) response latency; (3) inter-rater reliability (Cohen's kappa and ICC); and (4) logistic regression analysis of predictors of correct responses among physician participants. 2.6 Statistical Analysis All statistical analyses were conducted in R version 4.3.2. Non-parametric tests were used throughout given non-normal distribution of accuracy scores (Shapiro-Wilk, all p < 0.05). Group comparisons employed the Kruskal-Wallis H test for omnibus comparisons and Dunn's post-hoc test with Bonferroni correction for pairwise comparisons. Effect sizes were estimated using Cohen's d from ranked data. NS-MAS vs physician group comparisons used one-sample Wilcoxon signed-rank tests against the physician group's median. Inter-rater reliability was assessed using Cohen's kappa (κ) and ICC (two-way mixed, consistency). Statistical significance was α = 0.05 (two-tailed); all p-values are Bonferroni-adjusted for domain-level comparisons. 3. Results 3.1 Participant Characteristics Thirty specialist physicians completed the study: 10 pulmonologists, 10 allergists and clinical immunologists, and 10 emergency medicine physicians. Mean age ranged from 38.9 ± 5.7 years in the emergency medicine group to 43.8 ± 6.9 years in the pulmonology group. Mean post-residency experience was 13.6 ± 6.5 years for pulmonologists, 10.4 ± 5.1 years for allergists, and 9.4 ± 4.3 years for emergency physicians. Emergency medicine physicians managed the highest median annual obstructive airway disease caseload (median 220, IQR 170–300), reflecting acute exacerbation volume rather than longitudinal disease management. Self-reported familiarity with the GOLD 2025 guideline was highest among pulmonologists (mean 4.3 ± 0.5 on a 5-point scale), whereas familiarity with the GINA 2025 guideline was highest among allergists (mean 4.4 ± 0.4). Recent guideline-related continuing medical education attendance in the past two years was reported by 70% of pulmonologists, 70% of allergists, and 50% of emergency physicians. Full demographic and professional characteristics are summarized in Table 1. 3.2 Overall Accuracy The NS-MAS achieved an overall accuracy of 90.0% (27 correct out of 30 questions), significantly exceeding all three physician groups (Table 2, Figure 2). Among specialist physicians, allergists and clinical immunologists demonstrated marginally higher overall accuracy (mean 22.8 ± 2.2 out of 30, 76.0%) compared with pulmonologists (mean 22.5 ± 2.1 out of 30, 75.0%), while emergency physicians performed lowest across all human groups (mean 18.6 ± 2.0 out of 30, 62.0%). The omnibus Kruskal-Wallis test across all four groups revealed a highly significant difference (H = 29.1, df = 3, p < 0.001). Pairwise comparisons between NS-MAS and each physician group all reached statistical significance at Bonferroni-adjusted thresholds: NS-MAS versus pulmonologists (median difference +4.5 questions, p = 0.002, Cohen's d = 1.82, large effect), NS-MAS versus allergists (median difference +4.2 questions, p = 0.003, d = 1.74, large effect), and NS-MAS versus emergency physicians (median difference +8.4 questions, p < 0.001, d = 2.84, large effect). The difference in overall accuracy between pulmonologists and allergists did not reach statistical significance (p = 0.412, d = 0.14), whereas both groups significantly outperformed emergency physicians (pulmonology vs emergency: p = 0.009, d = 0.92; allergist vs emergency: p = 0.004, d = 1.08). Full pairwise comparisons and effect size estimates are presented in Table 3. 3.3 Domain-Specific Performance Domain-level accuracy results are presented in Table 2 and displayed as a radar chart in Figure 3. Several clinically important specialty-expertise interactions were identified across the six management domains. 3.3.1 Diagnosis and Differential Diagnosis The NS-MAS achieved perfect accuracy in the diagnostic domain (5/5, 100%). Pulmonologists (mean 4.4 ± 0.5 out of 5, 88%) and allergists (mean 4.2 ± 0.6 out of 5, 84%) demonstrated comparable performance, both significantly exceeding emergency physicians (mean 3.4 ± 0.5 out of 5, 68%; p < 0.001 for both comparisons). The NS-MAS advantage over all physician groups combined was large (mean difference +2.0 questions per 5, p < 0.001, d = 2.02). Common diagnostic errors among emergency physicians included misclassification of cardiac-origin dyspnoea as COPD exacerbation in elderly patients with preserved spirometry and failure to identify asthma-COPD overlap (ACO) in patients with features of both conditions. 3.3.2 Severity Assessment The NS-MAS answered all five severity assessment questions correctly (5/5, 100%). Allergists demonstrated marginally higher accuracy (mean 4.3 ± 0.5 out of 5, 86%) compared with pulmonologists (mean 4.1 ± 0.6 out of 5, 82%), with both groups significantly outperforming emergency physicians (mean 3.5 ± 0.7 out of 5, 70%; Kruskal-Wallis p = 0.008). The most frequent error pattern involved misapplication of the updated GOLD ABE classification — specifically, conflation of the abolished GOLD category C with the current mMRC/CAT-based GOLD B group — and underrecognition of severe or life-threatening acute asthma per BTS/SIGN criteria. 3.3.3 Pharmacological Management The pharmacological management domain was the only domain in which a physician group numerically exceeded NS-MAS accuracy: allergists achieved 82% (mean 4.1 ± 0.5 out of 5) compared with the NS-MAS score of 80% (4/5), though this difference did not reach statistical significance (p = 0.680). Pulmonologists scored 76% (mean 3.8 ± 0.6 out of 5) and emergency physicians 60% (mean 3.0 ± 0.7 out of 5). The overall domain Kruskal-Wallis test was significant (p = 0.004). Key discrimination errors included failure to initiate triple ICS/LABA/LAMA therapy in symptomatic GOLD Group E patients already receiving dual LAMA/LABA, prescription of ICS monotherapy in COPD without eosinophilic phenotype evidence, and incorrect biologic selection in severe uncontrolled asthma. 3.3.4 Exacerbation Management This domain revealed the most pronounced specialty-expertise interaction among physician groups. The NS-MAS achieved perfect accuracy (5/5, 100%). Allergists demonstrated significant superiority over pulmonologists (mean 4.2 ± 0.6 out of 5, 84% vs mean 3.6 ± 0.7 out of 5, 72%; p = 0.003, d = 0.93, large effect), particularly in acute severe asthma management scenarios involving intravenous magnesium sulfate dosing, heliox indications, and NIV thresholds in near-fatal asthma. Emergency physicians performed most poorly in this domain (mean 2.8 ± 0.6 out of 5, 56%; Kruskal-Wallis p < 0.001), consistent with suboptimal adherence to algorithmic antibiotic prescribing in COPD exacerbations per Anthonisen criteria and underuse of systemic corticosteroids at guideline-recommended doses and durations (Figure 3). 3.3.5 ICU Admission and Ventilatory Management In contrast to the exacerbation domain, pulmonologists outperformed all other physician groups in ICU and ventilatory management (mean 3.3 ± 0.8 out of 5, 66%), though the NS-MAS again achieved perfect accuracy (5/5, 100%). Allergists demonstrated the lowest performance among specialist groups in this domain (mean 2.9 ± 0.7 out of 5, 58%), significantly below pulmonologists (p < 0.001, d = 0.54, medium effect). Emergency physicians scored 62% (mean 3.1 ± 0.7 out of 5). Errors among allergists concentrated in NIV initiation thresholds for COPD-related hypercapnic respiratory failure, permissive hypercapnia strategy in mechanically ventilated near-fatal asthma, and dynamic hyperinflation recognition and management. The bidirectional expertise asymmetry between allergists and pulmonologists across the exacerbation and ICU/ventilatory domains is visually apparent in the radar chart (Figure 3) and statistically confirmed in Table 3. 3.3.6 Complication Management Complication management was the only domain in which the NS-MAS underperformed relative to at least one physician group. The NS-MAS scored 60% (3/5), being equaled by pulmonologists (mean 3.3 ± 0.7 out of 5, 66%). Allergists scored 62% (mean 3.1 ± 0.8 out of 5) and emergency physicians 56% (mean 2.8 ± 0.6 out of 5), the lowest of any group in this domain (Kruskal-Wallis p = 0.005). The NS-MAS errors occurred in two scenarios: one involving individualized dose-tapering decisions for corticosteroid-dependent asthma with adrenal insufficiency risk and one involving complex cor pulmonale exacerbation management requiring integration of right ventricular function assessment, diuretic titration, and long-term oxygen therapy prescription rationale. Notably, no physician group achieved high accuracy in this domain (range 56–66%), confirming the intrinsic difficulty of complication management reasoning across all groups. Per-question accuracy proportions across all 30 vignettes and all four groups are displayed in the heatmap in Figure 5. The visual analysis confirms domain-specific expertise clustering: questions 14 and 15 (exacerbation management, antimicrobial decision-making) and question 25 (ICU/ventilatory management, dynamic hyperinflation) showed the widest intergroup accuracy disparities, underscoring their discriminative value. 3.4 Response Latency NS-MAS generated responses with a median latency of 13.1 seconds per question (IQR 10.2–17.4 seconds), representing a 13-fold reduction compared with the fastest physician group (emergency physicians, median 2.91 minutes per question, IQR 2.2–3.5 minutes). Pulmonologists had the longest median response time (3.47 minutes, IQR 2.6–4.3 minutes), followed by allergists (3.12 minutes, IQR 2.3–4.0 minutes). The between-group latency difference was highly significant (Kruskal-Wallis H = 1762.4, p < 0.001), as illustrated in Figure 4. NS-MAS latency did not significantly differ across the six clinical domains (Kruskal-Wallis p = 0.19), indicating domain-agnostic processing efficiency irrespective of query complexity. 3.5 Inter-Rater Reliability Among Physicians Within-group pairwise Cohen's kappa values were in the fair-to-moderate range for all specialty groups: pulmonologists κ = 0.49 (95% CI 0.41–0.57), allergists κ = 0.46 (95% CI 0.38–0.54), and emergency physicians κ = 0.36 (95% CI 0.27–0.45). The overall intraclass correlation coefficient (ICC, two-way mixed, consistency) across all 30 physicians was 0.42 (95% CI 0.34–0.50), indicating moderate inter-rater agreement. These values quantify the substantial inter-individual variability in clinical obstructive airway disease decision-making that current practice tolerates even within specialty-defined peer groups. In contrast, the NS-MAS produced identical responses across three independent runs for 28 of 30 questions (93.3% consistency). The two inconsistent questions (Q7, severity assessment, GINA symptom control classification borderline; Q28, complication management, corticosteroid tapering) corresponded to scenarios in which the GINA 2025 guideline itself acknowledges boundary-condition ambiguity. 3.6 Predictors of Correct Response In multivariable mixed-effects logistic regression modeling correct response at the question level among physicians (n = 900 observations: 30 physicians × 30 questions; random intercept per physician), the following independent predictors of accuracy were identified: specialty group membership (allergist vs emergency medicine: OR 1.94, 95% CI 1.43–2.63, p < 0.001; pulmonology vs emergency medicine: OR 1.86, 95% CI 1.38–2.51, p < 0.001); post-residency years of experience (OR 1.05 per year, 95% CI 1.01–1.10, p = 0.024); self-reported GOLD 2025 guideline familiarity (OR 1.38 per unit, 95% CI 1.14–1.67, p = 0.001); self-reported GINA 2025 guideline familiarity (OR 1.42 per unit, 95% CI 1.17–1.72, p < 0.001); and recent guideline-related CME attendance (OR 1.58, 95% CI 1.12–2.23, p = 0.009). Clinical domain was also a significant predictor: physicians were most likely to answer correctly in the diagnosis domain (reference category) and least likely in ICU and ventilatory management (OR 0.48, 95% CI 0.36–0.64, p < 0.001) and exacerbation management for non-allergist physicians (OR 0.54, 95% CI 0.41–0.72, p < 0.001). Annual obstructive airway disease caseload was not an independent predictor after adjustment for specialty group and guideline familiarity (OR 1.00, 95% CI 0.99–1.01, p = 0.612). Full regression results are presented in Table 4. 4. Discussion This study demonstrates that a neurosymbolic multi-agent LLM system can achieve comprehensive, guideline-concordant performance across the full spectrum of COPD and asthma management clinical decision domains, surpassing all specialist physician groups in overall accuracy and exceeding each group within its respective area of primary expertise, with the singular exception of the complication management domain. The NS-MAS overall accuracy of 90.0% significantly exceeded pulmonologists (75.0%, p = 0.002, d = 1.82), allergists and clinical immunologists (76.0%, p = 0.003, d = 1.74), and emergency physicians (62.0%, p < 0.001, d = 2.84), with large effect sizes across all comparisons. These findings advance understanding of the conditions under which neurosymbolic AI clinical decision support systems may offer additive value relative to individual specialist judgment in obstructive airway disease management. The superiority of the NS-MAS across five of six clinical domains likely reflects its capacity to simultaneously integrate the GOLD 2025 ABE framework, GINA 2025 stepwise therapy algorithms, and asthma-COPD overlap (ACO) diagnostic criteria through parallel agent reasoning. This cognitive integration represents a genuine challenge even for clinicians primarily trained in one of the two guideline ecosystems. Recent real-world analyses continue to show suboptimal implementation of GOLD and GINA recommendations, with guideline-concordant care often clustering around one-half of cases in routine practice and varying substantially by care setting, specialty, and the treatment domain assessed [15]. The NS-MAS generated identical responses across three independent runs for 28 of 30 questions, yielding 93.3% consistency, while inter-rater reliability across physician groups ranged from κ = 0.36 to 0.49. These findings quantify the substantial degree of decision variability tolerated in routine clinical practice and underscore the operational relevance of reproducible algorithmic decision support. The most clinically significant finding is the bidirectional expertise asymmetry between pulmonologists and allergists across the exacerbation and ICU/ventilatory management domains. Allergists significantly outperformed pulmonologists in exacerbation management (84% vs 72%, p = 0.003, d = 0.93), particularly in acute severe asthma scenarios requiring intravenous magnesium sulfate dosing and near-fatal asthma NIV thresholds, competencies anchored in allergist-specific training frameworks. Conversely, pulmonologists outperformed allergists in ICU and ventilatory management (66% vs 58%, p < 0.001, d = 0.54), reflecting their greater exposure to permissive hypercapnia strategy, dynamic hyperinflation recognition, and COPD-specific NIV pH-guided protocols. This pattern is consistent with published evidence showing that physicians’ knowledge and care patterns remain strongly shaped by specialty-specific domains of expertise, while communication and interpretive differences between specialties create persistent knowledge boundaries even among highly trained clinicians[16]. The NS-MAS resolved this asymmetry by simultaneously applying both GOLD 2025 and GINA 2025 reasoning streams, achieving perfect accuracy (100%) in both domains. The pharmacological management domain was the only domain in which a physician group numerically exceeded NS-MAS performance, with allergists achieving 82% versus the NS-MAS score of 80%, though this difference did not reach statistical significance (p = 0.680). Allergists' superior biologic therapy selection, driven by daily clinical exposure to blood eosinophil thresholds, fractional exhaled nitric oxide (FeNO) interpretation, and IgE-guided omalizumab dosing, likely represents a form of tacit clinical expertise that current RAG-augmented agent architectures do not fully capture. The GINA 2025 add-on therapy algorithm involves nuanced biomarker integration requiring contextual patient-level judgment that extends beyond structured algorithmic logic, as has been observed with other precision medicine frameworks [17, 18]. This finding identifies a specific area for future NS-MAS architecture refinement, particularly regarding real-time biomarker integration. Complication management was the only domain in which the NS-MAS underperformed relative to at least one physician group, scoring 60% (3/5) compared with pulmonologists at 66%. NS-MAS errors occurred in two high-complexity scenarios: one involving individualized corticosteroid dose-tapering decisions in asthma with adrenal insufficiency risk, and one involving cor pulmonale exacerbation management requiring integration of right ventricular function, diuretic titration, and long-term oxygen therapy rationale. This aligns with a growing body of literature demonstrating that AI decision support performs less optimally in prognostically ambiguous, contextually nuanced scenarios compared with structured algorithmic decisions [19]. Critically, no physician group achieved high accuracy in this domain (range 56–66%), confirming that complication management in obstructive airway disease represents a genuine zone of clinical uncertainty for both AI systems and human specialists. Emergency physicians demonstrated the lowest overall accuracy (62.0%) and the largest performance gap relative to NS-MAS (d = 2.84). This finding is clinically relevant given that emergency departments represent the primary acute contact point for COPD and asthma exacerbations, with emergency physicians managing the highest annual caseload (median 220 cases per year, IQR 170–300) in this study. Suboptimal adherence to Anthonisen antibiotic prescribing criteria and underuse of systemic corticosteroids at guideline-recommended doses and durations were the predominant error patterns, consistent with previously published audits of emergency department COPD management [15]. The 13-fold response latency advantage of NS-MAS (median 13.1 seconds vs 2.91 minutes for emergency physicians, Kruskal-Wallis H = 1762.4, p < 0.001) carries direct clinical relevance: in acute severe asthma, escalation to intravenous magnesium sulfate is recommended when there is inadequate response to intensive initial therapy, and earlier administration has been associated with improved emergency department process measures and short-term clinical response; in COPD exacerbation, delayed NIV initiation is associated with higher in-hospital mortality [20]. The NS-MAS architecture evaluated in this study differs from prior single-agent LLM benchmarking studies in several important ways. First, the agent specialization structure mirrors multidisciplinary team reasoning, distributing domain expertise across dedicated agents rather than requiring a single model to carry all obstructive airway disease subspecialty knowledge. Second, the symbolic reasoning layer, a Prolog-based engine implementing 312 formal rules derived from GOLD 2025 and GINA 2025 decision trees, provides formal guideline-rule enforcement that single neural models cannot guarantee, addressing a core limitation of current LLMs in safety-critical medical applications [21, 22]. Third, the LangGraph orchestration enables traceable, auditable reasoning pathways essential for clinical deployment trustworthiness. Fourth, for ACO queries the dual co-activation of Pulmonology and Allergy/Immunology agents simultaneously resolved the otherwise systematic knowledge gap observed in each single-specialty physician group. Multivariable mixed-effects logistic regression identified GOLD 2025 guideline familiarity (OR 1.38 per unit, p = 0.001), GINA 2025 guideline familiarity (OR 1.42 per unit, p < 0.001), and recent CME attendance (OR 1.58, p = 0.009) as independent predictors of correct response among physicians, after adjustment for specialty group membership and post-residency experience. Annual caseload was not an independent predictor (OR 1.00, p = 0.612), suggesting that volume of clinical exposure does not compensate for guideline-specific knowledge currency. This finding supports targeted continuing medical education aligned with current GOLD and GINA updates, together with structured AI training and supervised AI-assisted decision support, as an integrated strategy for improving the quality, consistency, and safety of obstructive airway disease management[23]. Several limitations require acknowledgment. First, vignette-based evaluation does not capture the full complexity of real-world clinical decision-making, which incorporates dynamic patient trajectories, nonverbal clinical information, shared decision-making, and systemic healthcare context. Second, the NS-MAS was RAG-augmented with full GOLD 2025 and GINA 2025 guideline texts while physicians had no reference access, a condition that favors the AI system but reflects its intended clinical deployment context. Third, although the study employed a multicenter design across four institutions in the Ankara region (Ankara Bilkent City Hospital, Ankara Etlik City Hospital, Mamak State Hospital, and Etimesgut Şehit Sait Ertürk State Hospital), all participating centers were geographically confined to a single metropolitan area (Ankara, Turkey), and the physician sample of 10 per specialty may not represent the full distribution of specialist competence nationally or globally; future studies should recruit from geographically diverse institutions across multiple regions and healthcare systems. Fourth, LLM model versions and guideline content evolve rapidly, and performance benchmarks require periodic recalibration as both GOLD and GINA issue annual updates. Fifth, this study does not address implementation feasibility, clinician acceptance, alert fatigue, or patient-level outcome effects in deployed clinical settings, which should be the subject of prospective clinical trials. 5. Conclusions A neurosymbolic multi-agent LLM system orchestrating eight specialized agents demonstrated significantly superior guideline-concordant COPD and asthma management performance relative to board-certified pulmonologists, allergists and clinical immunologists, and emergency physicians across the majority of clinical decision domains. The system achieved particular advantages in domains where individual specialty expertise is most limited, namely ICU and ventilatory management for allergists, and exacerbation-specific pharmacological management for pulmonologists and emergency physicians. The complication management domain represented the one area where pulmonologist physicians equaled or exceeded NS-MAS performance, suggesting that contextually complex, longitudinally nuanced management scenarios remain a relative limitation of current multi-agent architectures. These findings provide foundational evidence for the clinical decision support potential of neurosymbolic multi-agent AI in obstructive airway disease management, warranting prospective validation in real-world acute and chronic care settings. Declarations Ethics Approval and Consent to Participate This study was approved by the Ankara İl Sağlık Müdürlüğü Müdahalesiz Etik Kurul (Non-Interventional Ethics Board; Approval No: 2025-10-3, Date: 24 October 2025). All physician participants provided written informed consent prior to participation. No patient data were collected or used in this study. Availability of Data and Materials The full clinical vignette bank (Supplementary Material 1) and NS-MAS prompt architectures (Supplementary Material 2) are provided with this manuscript. Anonymized physician response data are available from the corresponding author upon reasonable request. Competing Interests The authors declare no competing interests. No external funding was received from AI technology companies. The use of commercial LLM APIs (OpenAI, Anthropic, Google DeepMind) was self-funded by the research team. Funding This study received no dedicated external funding. Research was conducted as part of the doctoral research program of M.T.Ü. at Hacettepe University Faculty of Medicine. Authors' Contributions E.A.M. contributed to study design, clinical vignette development, data collection, and critical manuscript revision. M.T.Ü. led study design, NS-MAS development and implementation, statistical analysis, and manuscript writing. E.E. contributed to clinical vignette development (emergency medicine domain), data collection, and manuscript revision. All authors read and approved the final manuscript. References Cao, Z., et al., Burden of chronic obstructive pulmonary disease and its attributable risk factors in 204 countries and territories, 1990-2021: results from the Global Burden of Disease Study 2021. BMJ Public Health, 2026. 4 (1): p. e002489. Zhang, L., et al., Global, regional and national burden of asthma from 1990 to 2021: a systematic analysis for the Global Burden of Disease Study 2021. BMJ Open Respiratory Research, 2025. 12 (1): p. e003144. Fricker, M. and R. Lokwani, COPD: the role of neutrophils in inflammation, pathophysiology, and as drug targets. Clin Sci (Lond), 2025. 139 (20): p. 1199-1214. Matsunaga, K., et al., Guidance for type 2 inflammatory biomarkers. Respir Investig, 2025. 63 (3): p. 273-288. Chan, R., N.E. Horn, and S. Siddiqui, Precision Medicine for Asthma: Tailored to its Severity and Endotype/Phenotype. Allergy Asthma Immunol Res, 2026. 18 (1): p. 19-38. Xie, C., et al., Toward precision medicine in COPD: phenotypes, endotypes, biomarkers, and treatable traits. Respiratory Research, 2025. 26 (1): p. 274. Gayen, S., et al., Critical Care Management of Severe Asthma Exacerbations. Journal of Clinical Medicine, 2024. 13 (3): p. 859. Halpin, D.M.G., et al., Addressing the Global Challenges of COPD and Asthma: A Shared Vision From the Global Initiative for Chronic Obstructive Pulmonary Disease (GOLD) and the Global Initiative for Asthma (GINA). Respirology, 2026. 31 (2): p. 135-140. Chen, S.F., et al., LLM-assisted systematic review of large language models in clinical medicine. Nature Medicine, 2026. 32 (3): p. 1152-1159. Phan, L., et al., A benchmark of expert-level academic questions to assess AI capabilities. Nature, 2026. 649 (8099): p. 1139-1146. Ucdal, M., K. Yurtsever, and E. Ekingen, Multidisciplinary artificial intelligence systems versus single-model approaches for the diagnosis and management of ileus and volvulus. BMC Gastroenterol, 2026. 26 (1): p. 124. Sorka, M., et al., A multi-agent approach to neurological clinical reasoning. PLOS Digit Health, 2025. 4 (12): p. e0001106. Acharya, K. and H. Song, A Comprehensive Review of Neuro-symbolic AI for Robustness, Uncertainty Quantification, and Intervenability. Arabian Journal for Science and Engineering, 2026. 51 (1): p. 35-67. Ekingen, E. and M. Ucdal, Performance Comparison of a Neuro-Symbolic Large Language Model System Versus Human Experts in Acute Cholecystitis Management. J Clin Med, 2026. 15 (5). Cuperus, L.J.A., et al., Adherence to Inhaled Medication Guidelines in Patients with COPD: Evaluating the GOLD 2023 Strategy in a Real-World Secondary Care Setting. Int J Chron Obstruct Pulmon Dis, 2025. 20 : p. 3877-3891. Braam, A., et al., Collaboration Between Physicians from Different Medical Specialties in Hospital Settings: A Systematic Review. J Multidiscip Healthc, 2022. 15 : p. 2277-2300. Balasubramanyam, S., E.K. George, and E. Wang, Precision medicine and choosing a biologic in asthma: understanding the current state of knowledge for predictors of response and clinical remission. Curr Opin Allergy Clin Immunol, 2025. 25 (1): p. 66-74. Quek, E., N. Horn, and S. Siddiqui, Precision Medicine in Asthma: The Role of Biomarkers. Immunotargets Ther, 2025. 14 : p. 1479-1513. Bean, A.M., et al., Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine, 2026. 32 (2): p. 609-615. Alzain, M., et al., Impact of early vs. delayed intravenous magnesium sulfate on clinical outcomes in pediatric severe asthma: a retrospective cohort study. Frontiers in Disaster and Emergency Medicine, 2025. Volume 3 - 2025 . Liu, Q., et al., EvoMDT: a self-evolving multi-agent system for structured clinical decision-making in multi-cancer. npj Digital Medicine, 2026. 9 (1): p. 124. Tan, X., et al. Prolog-Driven Rule-Based Diagnostics with Large Language Models for Precise Clinical Decision Support . 2026. Cham: Springer Nature Switzerland. Qunaibi, E.A., et al., Effectiveness of Informed AI Use on Clinical Competence of General Practitioners and Internists: Pre-Post Intervention Study. JMIR Med Educ, 2026. 12 : p. e75534. Tables Table 1. Participant Demographic and Professional Characteristics Characteristic Pulmonology (n=10) Allergist/Immunology (n=10) Emergency Medicine (n=10) Demographics Age, years, mean ± SD 43.8 ± 6.9 40.2 ± 6.1 38.9 ± 5.7 Gender, n (%) Female 3 (30%) 5 (50%) 4 (40%) Post-residency experience, years 13.6 ± 6.5 10.4 ± 5.1 9.4 ± 4.3 Academic & Institutional Profile Attending/Specialist, n (%) 8 (80%) 8 (80%) 9 (90%) Associate Professor or above 3 (30%) 2 (20%) 1 (10%) University hospital, n (%) 6 (60%) 5 (50%) 4 (40%) Annual COPD/Asthma cases, median (IQR) 140 (100–200) 185 (140–250) 220 (170–300) Guideline Familiarity GOLD 2025 COPD Guideline familiarity (1–5) 4.3 ± 0.5 3.2 ± 0.8 2.9 ± 0.9 GINA 2025 Asthma Guideline familiarity (1–5) 3.9 ± 0.6 4.4 ± 0.4 2.7 ± 0.8 CME in past 2 years, n (%) 7 (70%) 7 (70%) 5 (50%) Values are mean ± SD, median (IQR), or n (%) as appropriate. CME, continuing medical education Table 2. Accuracy by Clinical Domain and Group Domain (5 questions each) NS-MAS (n/5) Pulmonology mean ± SD Allergist/Immunol. mean ± SD Emergency mean ± SD p-value* 1. Diagnosis & differential 5/5 (100%) 4.4 ± 0.5 (88%) 4.2 ± 0.6 (84%) 3.4 ± 0.5 (68%) < 0.001 2. Severity assessment 5/5 (100%) 4.1 ± 0.6 (82%) 4.3 ± 0.5 (86%) 3.5 ± 0.7 (70%) 0.008 3. Pharmacological mgmt. 4/5 (80%) 3.8 ± 0.6 (76%) 4.1 ± 0.5 (82%) 3.0 ± 0.7 (60%) 0.004 4. Exacerbation mgmt. 5/5 (100%) 3.6 ± 0.7 (72%) 4.2 ± 0.6 (84%) 2.8 ± 0.6 (56%) < 0.001 5. ICU admission & ventilation 5/5 (100%) 3.3 ± 0.8 (66%) 2.9 ± 0.7 (58%) 3.1 ± 0.7 (62%) < 0.001 6. Complication management 3/5 (60%) 3.3 ± 0.7 (66%) 3.1 ± 0.8 (62%) 2.8 ± 0.6 (56%) 0.005 OVERALL (out of 30) 27/30 (90.0%) 22.5 ± 2.1 (75.0%) 22.8 ± 2.2 (76.0%) 18.6 ± 2.0 (62.0%) < 0.001 Data for physician groups are mean ± SD (percentage). NS-MAS data are raw score/5 (percentage). *Kruskal-Wallis test across all four groups; p-values Bonferroni-adjusted for six domain comparisons (threshold 0.0083). ICU, intensive care unit; NS-MAS, neurosymbolic multi-agent system. Table 3. Pairwise Comparisons and Effect Sizes Comparison Median diff. (questions) U statistic p-value (adj.)* Effect size (Cohen's d) NS-MAS vs specialist groups NS-MAS vs pulmonology +4.5 (90% vs 75%) — 0.002 1.82 (large) NS-MAS vs allergist/immunol. +4.2 (90% vs 76%) — 0.003 1.74 (large) NS-MAS vs emergency med. +8.4 (90% vs 62%) — < 0.001 2.84 (large) Between specialist groups Pulmonology vs allergist −0.3 (75% vs 76%) 44.1 0.412 (NS) 0.14 (trivial) Pulmonology vs emergency +3.9 (75% vs 62%) 26.8 0.009 0.92 (large) Allergist vs emergency +4.2 (76% vs 62%) 23.9 0.004 1.08 (large) Domain-specific interactions Diagnosis: NS-MAS vs all phys. +2.0 per 5 questions — < 0.001 2.02 (large) Exacerbation: allergy vs pulm. +0.6 (4.2 vs 3.6) 30.4 0.003 0.93 (large) ICU/vent: pulm. vs allergy +0.4 (3.3 vs 2.9) 34.1 < 0.001 0.54 (medium) Pharmacology: allergy vs emerg. +1.1 (4.1 vs 3.0) 31.2 0.007 0.82 (large) Complication: pulm. vs emerg. +0.5 (3.3 vs 2.8) 35.8 0.028 0.68 (medium) Dunn post-hoc test with Bonferroni correction for pairwise comparisons. For NS-MAS vs physician groups, one-sample Wilcoxon signed-rank test against physician group median. Effect size interpretation: trivial (d 0.8). *For non-allergist physicians in the exacerbation domain. NS, not significant; OR, odds ratio. Table 4. Multivariable Mixed-Effects Logistic Regression: Predictors of Correct Response Among Physicians Predictor OR 95% CI p-value Interpretation Specialty: allergist vs emergency 1.94 1.43–2.63 < 0.001 Large advantage Specialty: pulmonology vs emergency 1.86 1.38–2.51 < 0.001 Large advantage Post-residency experience (per year) 1.05 1.01–1.10 0.024 Modest positive GOLD 2025 guideline familiarity (per unit) 1.38 1.14–1.67 0.001 Positive GINA 2025 guideline familiarity (per unit) 1.42 1.17–1.72 < 0.001 Positive Recent CME attendance (yes vs no) 1.58 1.12–2.23 0.009 Positive Domain: ICU/ventilation vs diagnosis 0.48 0.36–0.64 < 0.001 Harder domain Domain: exacerbation vs diagnosis* 0.54 0.41–0.72 < 0.001 Harder domain Annual caseload (per case) 1.00 0.99–1.01 0.612 Not significant n = 900 observations (30 physicians × 30 questions); random intercept per physician. Reference categories: specialty = emergency medicine; domain = diagnosis. CME, continuing medical education; GOLD, Global Initiative for Chronic Obstructive Lung Disease; GINA, Global Initiative for Asthma; OR, odds ratio; CI, confidence interval. Additional Declarations No competing interests reported. Supplementary Files supplementary2architecturefinal21.03.docx supplementary1vignettes22.03.docx GraphicalAbstract.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviews received at journal 18 May, 2026 Reviewers agreed at journal 20 Apr, 2026 Reviewers agreed at journal 18 Apr, 2026 Reviewers invited by journal 18 Apr, 2026 Editor invited by journal 06 Apr, 2026 Editor assigned by journal 31 Mar, 2026 Submission checks completed at journal 31 Mar, 2026 First submitted to journal 30 Mar, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9262455","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":629423776,"identity":"afa19974-9636-455f-a2f3-41257808598b","order_by":0,"name":"Ebru Aykan Mavigoz","email":"","orcid":"","institution":"Etimesgut Asker Hastanesi","correspondingAuthor":false,"prefix":"","firstName":"Ebru","middleName":"Aykan","lastName":"Mavigoz","suffix":""},{"id":629423777,"identity":"f71edfc7-be93-4a68-a96f-00104871627e","order_by":1,"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":629423778,"identity":"8a0afe0d-d9c1-43c3-9f0c-aff59fbb1dd7","order_by":2,"name":"Evren Ekingen","email":"","orcid":"","institution":"Etimesgut Asker Hastanesi","correspondingAuthor":false,"prefix":"","firstName":"Evren","middleName":"","lastName":"Ekingen","suffix":""}],"badges":[],"createdAt":"2026-03-30 04:53:51","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9262455/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9262455/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":107929616,"identity":"8c5d6971-a459-4322-af36-150951e0c4c2","added_by":"auto","created_at":"2026-04-27 16:19:28","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":255171,"visible":true,"origin":"","legend":"\u003cp\u003eStudy Design and Participant Flow Diagram. Participant flow from eligibility assessment through completion of the 30-item COPD/asthma management vignette battery. Three specialist cohorts (pulmonology, allergist/immunology, emergency medicine; n = 10 per group) and the NS-MAS underwent standardized evaluation.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/3a1b5c19e03de72ee5ef96bf.png"},{"id":107929618,"identity":"30141f62-f68f-4129-ae19-7ba8b980e859","added_by":"auto","created_at":"2026-04-27 16:19:28","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":124176,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAccuracy performance of NS-MAS vs specialist physician groups. \u003c/strong\u003e(A) Overall accuracy across all 30 questions for each group. Error bars represent ±1 SD for physician groups; NS-MAS accuracy is shown as a single value (majority vote across 3 independent runs, no SD applicable). Significance brackets indicate Wilcoxon signed-rank test results (Bonferroni-adjusted): **p \u0026lt; 0.01; ***p \u0026lt; 0.001. NS-MAS significantly exceeded all physician groups: vs pulmonology p = 0.002 (d = 1.82); vs allergist/immunology p = 0.003 (d = 1.74); vs emergency medicine p \u0026lt; 0.001 (d = 2.84). (B) Domain-specific accuracy for each group across the six COPD and asthma management domains. The bidirectional expertise asymmetry between allergists (superior in exacerbation management, 84%) and pulmonologists (superior in ICU/ventilatory management, 66% vs 58%) is visually apparent in panel B.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations: NS-MAS, neurosymbolic multi-agent system; SD, standard deviation; d, Cohen's d effect size.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/5be6f6a57e2d1f14a9385114.png"},{"id":107929620,"identity":"ac70ea25-a354-4150-9433-3d9e21144c8d","added_by":"auto","created_at":"2026-04-27 16:19:28","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3287155,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eRadar chart of domain-specific performance profiles. \u003c/strong\u003eSpider/radar plot showing the proportional accuracy (%) of each group across the six clinical domains on a scale of 0–100%. The NS-MAS profile (navy, solid line) encompasses all other groups in five of six domains. The one exception is the complication management domain (lower-left vertex), where pulmonologists (purple dashed line) achieved equivalent accuracy (66% vs 60%). The bidirectional expertise asymmetry between allergists (teal dash-dot) and pulmonologists is clearly visible: allergists demonstrate superior performance in exacerbation management (84% vs 72%) while pulmonologists outperform allergists in ICU/ventilatory management (66% vs 58%). Emergency physicians (amber dotted line) show the most restricted profile, with particularly low accuracy in exacerbation management (56%) and complication management (56%).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations: NS-MAS, neurosymbolic multi-agent system; ICU, intensive care unit; COPD, chronic obstructive pulmonary disease.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/d6990f93f229b6dfe589ef6b.png"},{"id":107929622,"identity":"69c3f71f-1d92-481b-a1d0-572fcb26b538","added_by":"auto","created_at":"2026-04-27 16:19:28","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":216694,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eResponse latency distribution: NS-MAS vs specialist physicians. \u003c/strong\u003e(A) Box plot of NS-MAS response time per question in seconds (n = 90 queries: 30 vignettes × 3 independent trial runs). Median 13.1 seconds (IQR 10.2–17.4 seconds). (B) Box plots of physician response time per question in minutes for each specialist group (n = 300 per box: 10 physicians × 30 questions). Median response times: pulmonologists 3.47 min (IQR 2.6–4.3 min), allergists/immunologists 3.12 min (IQR 2.3–4.0 min), emergency physicians 2.91 min (IQR 2.2–3.5 min). Boxes show median and IQR; whiskers extend to 1.5×IQR; individual outliers shown as circles. The between-group latency difference was highly significant (Kruskal-Wallis H = 1762.4, p \u0026lt; 0.001), representing a 13-fold speed advantage for the NS-MAS over the fastest physician group. NS-MAS latency did not significantly differ across the six clinical domains (Kruskal-Wallis p = 0.19).\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations: NS-MAS, neurosymbolic multi-agent system; IQR, interquartile range; Md, median.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/7998cbcc09849961d200048f.png"},{"id":107929623,"identity":"94f29354-53f6-420b-8f4d-b3009847d1b2","added_by":"auto","created_at":"2026-04-27 16:19:28","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":2774325,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePer-question accuracy heatmap. \u003c/strong\u003eMatrix visualization of accuracy proportions for each of the 30 clinical vignettes (columns) across the four evaluation groups (rows). Color scale ranges from red (0% accuracy) through yellow (50%) to dark green (100%). Questions are organized by clinical domain (six domains of five questions each, separated by white vertical lines). Domain labels are shown above the matrix. Arrows above the heatmap indicate questions with the widest intergroup accuracy disparities: Questions 14 and 15 (exacerbation management domain, antimicrobial decision-making per Anthonisen criteria and GOLD 2025) and Question 25 (ICU/ventilatory management domain, dynamic hyperinflation recognition and permissive hypercapnia strategy in near-fatal asthma), which were identified a priori as high-discriminative-value questions. The NS-MAS row (top) achieves near-uniform green across five of six domain blocks, with the complication management domain (rightmost block) the only exception. The emergency medicine row (bottom) shows the most red-dominant profile, particularly in exacerbation and complication management domains.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eAbbreviations: NS-MAS, neurosymbolic multi-agent system; Q, question; D1, Diagnosis; D2, Severity Assessment; D3, Pharmacological Management; D4, Exacerbation Management; D5, ICU \u0026amp; Ventilation; D6, Complication Management.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"5.png","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/4b51299562fe509656c4a589.png"},{"id":108008782,"identity":"d65fa953-5bbd-4fa3-91cb-c106ffbfe6e7","added_by":"auto","created_at":"2026-04-28 13:08:24","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":7039211,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/05790b65-0193-478d-8924-d7535ae3e3e8.pdf"},{"id":108007472,"identity":"ef247d80-52f8-4df3-bd88-5f23327d9697","added_by":"auto","created_at":"2026-04-28 13:00:11","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":173007,"visible":true,"origin":"","legend":"","description":"","filename":"supplementary2architecturefinal21.03.docx","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/bb61a7f2a5369612d35372ea.docx"},{"id":108007449,"identity":"2efa321e-e3b4-4a5d-8373-41df5ab6c42c","added_by":"auto","created_at":"2026-04-28 13:00:06","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":28486,"visible":true,"origin":"","legend":"","description":"","filename":"supplementary1vignettes22.03.docx","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/ee05d125b8cad64d8ae532c9.docx"},{"id":108006181,"identity":"8388e355-af57-42dd-b50f-928133a32735","added_by":"auto","created_at":"2026-04-28 12:54:21","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":433376,"visible":true,"origin":"","legend":"","description":"","filename":"GraphicalAbstract.docx","url":"https://assets-eu.researchsquare.com/files/rs-9262455/v1/718e93cebd8d2e573e45a2fa.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Neurosymbolic Multi-Agent Large Language Model System Versus Specialist Physicians in COPD and Asthma Management: A Comparative Performance Evaluation Using Guideline-Based Clinical Vignettes","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003eChronic obstructive pulmonary disease (COPD) and asthma are the most prevalent chronic airway diseases worldwide. Recent Global Burden of Disease-based analyses indicate that asthma accounted for approximately 260.48\u0026nbsp;million prevalent cases and 436,190 deaths in 2021, whereas COPD accounted for approximately 213.4\u0026nbsp;million prevalent cases and 3.7\u0026nbsp;million deaths in the same year [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Thinking about shared diagnostic and management frameworks between COPD and asthma despite mechanistic differences.Despite important mechanistic differences, with COPD typically featuring persistent airflow limitation related to noxious exposure-driven, often neutrophil-predominant inflammation and emphysematous or small-airway remodeling, and asthma typically featuring variable expiratory airflow limitation with type 2 or eosinophilic inflammation in many patients, both conditions share core management imperatives: accurate objective diagnosis, structured severity and risk assessment, personalized pharmacologic optimization, and early recognition and prevention of exacerbations [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e, \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eThe clinical complexity of obstructive airway disease management is compounded by substantial phenotypic heterogeneity. COPD includes emphysema-predominant, chronic bronchitis-predominant, and frequent-exacerbator phenotypes with distinct therapeutic implications, whereas asthma spans a spectrum from infrequent, mild symptoms to severe disease requiring Step 5 specialist assessment, phenotyping, and add-on therapy [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e].In patients with coexisting features of asthma and COPD, management is particularly challenging; however, GOLD 2025 no longer endorses the term \u0026ldquo;asthma-COPD overlap (ACO)\u0026rdquo; and recommends that, when concurrent asthma is suspected, pharmacotherapy should primarily follow asthma guidelines while COPD-specific assessment and management are also addressed [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e]. These conditions demand sophisticated decision-making across multiple domains: objective confirmation and interpretation of airflow limitation with consideration of alternative cardiac and upper-airway diagnoses; GOLD ABE assessment and GINA step-based treatment stratification; selection of inhaled regimens (SAMA, SABA, LAMA, LABA, ICS) tailored to symptoms, exacerbation risk, phenotype, and comorbidity burden; acute exacerbation management including decisions regarding systemic corticosteroids and antibiotics; and escalation to ventilatory support, including non-invasive ventilation for acute hypercapnic respiratory failure in COPD and invasive mechanical ventilation in near-fatal asthma [\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eDespite internationally endorsed guidelines, specifically the 2025 GOLD Report and the 2025 GINA Strategy Report, substantial inter-specialty variability in guideline adherence and clinical accuracy persists. Pulmonologists, allergists and clinical immunologists, and emergency physicians each bring complementary but non-overlapping expertise: pulmonologists excel in spirometric interpretation and critical care integration yet may underperform in allergy-driven severe asthma biologic selection; allergists demonstrate superior competence in phenotype-driven asthma management yet may lag in COPD-specific interventional and ventilatory decision-making; emergency physicians are adept at acute exacerbation stabilization but may apply suboptimal chronic maintenance therapy recommendations [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eArtificial intelligence (AI) and large language model (LLM) systems have emerged as potentially transformative tools in clinical decision support [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Early benchmarking studies demonstrated that frontier LLMs can achieve expert-level performance on standardized medical examinations [\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e]. Multi-agent LLM systems, in which specialized agents are orchestrated to collaboratively reason across problem sub-domains, have shown performance improvements over monolithic LLM approaches in complex clinical reasoning tasks [\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e]. Neurosymbolic AI, which combines neural pattern recognition with formal symbolic reasoning, rule-based inference, and modular agent architectures, offers transparency, rule-adherence, and verifiable guideline consistency that renders it particularly suited to obstructive airway disease management, given that algorithmic stepwise decision frameworks are central to both the GOLD and GINA guideline [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e]. Thinking about cholesterol management and dyslipidemia treatment approachesNotably, comparable AI-assisted decision support systems have previously demonstrated consistent guideline adherence in other chronic disease domains governed by stepwise algorithmic frameworks, such as cholecystitis management and biliary disease treatment optimization, suggesting that the neurosymbolic multi-agent paradigm may generalize across guideline-driven clinical specialties[\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eTo our knowledge, no study has systematically compared a neurosymbolic multi-agent LLM system against board-certified specialist physicians across the full spectrum of COPD and asthma management domains, from initial diagnosis through exacerbation and complication management. The present study addresses this gap by conducting a rigorous comparative evaluation using guideline-derived clinical vignettes, with the explicit aim of characterizing both overall performance and domain-specific strengths and limitations of AI versus specialist physician decision-making in obstructive airway disease management.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study Design\u003c/h2\u003e \u003cp\u003eThis was a cross-sectional, comparative evaluation study conducted between October 2025 and March 2026. The study was approved by the Ankara Provincial Directorate of Public Health Non-Interventional Ethics Board (Approval No: 2025-10-3, Date: 24 October 2025) and was classified as non-interventional in accordance with Turkish Ministry of Health regulations, given that no patient data were collected or utilized and all physician participants provided written informed consent prior to participation. The study is reported in accordance with the GREET (Guideline for Reporting of Evaluation of AI in Healthcare) framework. The overall participant flow and study design are illustrated in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Clinical Case Vignette Development\u003c/h2\u003e \u003cp\u003eThirty USMLE Step 2 CK-format clinical vignettes were developed through a rigorous multi-stage process. An expert committee comprising two pulmonologists, one allergist/clinical immunologist, and one emergency physician with expertise in AI clinical benchmarking (the study co-investigators) initially generated candidate questions across six pre-specified COPD/asthma management domains.\u003c/p\u003e \u003cdiv id=\"Sec5\" class=\"Section3\"\u003e \u003ch2\u003e2.2.1 Clinical Domains\u003c/h2\u003e \u003cp\u003eThe six domains were defined based on primary clinical decision nodes identified in current international guidelines: (1) \u003cb\u003eDiagnosis and Differential Diagnosis\u003c/b\u003e \u0026mdash; spirometric confirmation and pattern interpretation (obstructive, mixed), discrimination from cardiac causes of dyspnoea (heart failure, pulmonary hypertension), pulmonary embolism, upper airway obstruction, interstitial lung disease, and bronchiectasis; correct COPD-asthma overlap (ACO) identification; (2) \u003cb\u003eSeverity Assessment\u003c/b\u003e \u0026mdash; application of GOLD ABCD/ABE grading system, mMRC dyspnoea scale, CAT score, spirometric severity (GOLD 1\u0026ndash;4), GINA symptom control classification (well-controlled, partly controlled, uncontrolled), NAEPP asthma severity classification, and acute exacerbation severity grading (COPD: Anthonisen/GOLD exacerbation criteria; Asthma: BTS/SIGN acute asthma severity); (3) \u003cb\u003ePharmacological Management\u003c/b\u003e \u0026mdash; guideline-concordant inhaler therapy for COPD (SAMA, SABA, LAMA, LABA, ICS/LABA, LAMA/LABA/ICS triple therapy; roflumilast; azithromycin prophylaxis) and asthma (SABA reliever, ICS controller, ICS/LABA, SMART therapy, add-on biologic therapy for severe eosinophilic asthma including dupilumab, mepolizumab, benralizumab, omalizumab); (4) \u003cb\u003eExacerbation Management\u003c/b\u003e \u0026mdash; correct classification and pharmacological management of COPD exacerbations (SABA/SAMA nebulization, systemic corticosteroids, antibiotic decision-making per Anthonisen criteria, oxygen titration in type 2 respiratory failure), and acute severe/near-fatal asthma management (IV magnesium sulfate, IV/nebulized salbutamol, helium-oxygen, NIV indications, intubation thresholds); (5) \u003cb\u003eICU Admission and Ventilatory Management\u003c/b\u003e \u0026mdash; indications for NIV in COPD exacerbation-related hypercapnic respiratory failure (pH thresholds, Pco₂ targets), invasive mechanical ventilation for near-fatal asthma (dynamic hyperinflation management, permissive hypercapnia), weaning protocols, and extubation readiness criteria; and (6) \u003cb\u003eComplication Management\u003c/b\u003e \u0026mdash; recognition and management of cor pulmonale (right heart failure in COPD), secondary pneumothorax in COPD/asthma, COPD-related hypercapnic encephalopathy, mucus plugging with lobar collapse, and adrenal insufficiency from chronic corticosteroid use.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section3\"\u003e \u003ch2\u003e2.2.2 Vignette Validation\u003c/h2\u003e \u003cp\u003eEach domain comprised five vignettes (total n\u0026thinsp;=\u0026thinsp;30), each presenting a detailed clinical scenario including patient demographics, symptom history, physical examination findings, pulmonary function data, laboratory results, and imaging findings where applicable. All vignettes were structured as five-option single-best-answer questions. A correct answer key was developed by three independent content experts and verified against primary guideline sources (GOLD 2025, GINA 2025, BTS/SIGN Asthma Guidelines 2023). Vignettes achieving less than 90% inter-expert agreement (n\u0026thinsp;=\u0026thinsp;2) were revised iteratively until consensus was achieved. The complete vignette bank is provided in Supplementary Material 1.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Neurosymbolic Multi-Agent System (NS-MAS)\u003c/h2\u003e \u003cdiv id=\"Sec8\" class=\"Section3\"\u003e \u003ch2\u003e2.3.1 System Architecture\u003c/h2\u003e \u003cp\u003eThe NS-MAS was developed using LangGraph (version 0.2.x, LangChain Inc.) as the state-machine orchestration framework. The system comprised eight specialized functional agents plus the orchestration layer: (i) \u003cb\u003eOrchestrator Agent (LangGraph)\u003c/b\u003e: Stateful graph-based router responsible for query parsing, agent selection, reasoning chain construction, and conflict resolution; (ii) \u003cb\u003ePulmonology/Respiratory Agent (GPT-4.5, OpenAI)\u003c/b\u003e: Specialized for COPD diagnosis, spirometric interpretation, emphysema phenotyping, GOLD staging, critical care ventilatory decisions, and ACO management. RAG-augmented with the full GOLD 2025 Report; (iii) \u003cb\u003eAllergy/Immunology Agent (Claude Sonnet 4.6, Anthropic)\u003c/b\u003e: Specialized for asthma phenotyping, GINA stepwise management, T2 biomarker interpretation (FeNO, blood eosinophils, IgE), biologic therapy selection, and allergen-triggered exacerbation management. RAG-augmented with GINA 2025 Strategy Report and BTS/SIGN Asthma Guidelines 2023; (iv) \u003cb\u003eEmergency Medicine Agent (Gemini 2.5 Pro, Google DeepMind)\u003c/b\u003e: Specialized for acute COPD exacerbation management, acute severe asthma in the emergency department, rapid severity assessment, and time-sensitive intervention decisions; (v) \u003cb\u003ePharmacology/Therapeutics Agent (GPT-4.5, OpenAI)\u003c/b\u003e: Specialized for inhaler device selection, drug-drug interactions, dose optimization (PK/PD), corticosteroid-sparing strategies, and biologic prescribing criteria; (vi) \u003cb\u003eEvidence Synthesis Agent (GPT-4.5\u0026thinsp;+\u0026thinsp;RAG, OpenAI)\u003c/b\u003e: Retrieves and synthesizes relevant guideline text and clinical trial evidence (TORCH, UPLIFT, SUMMIT, IMPACT, ETHOS, NAVIGATOR, MEDI3506 trials); (vii) \u003cb\u003eRadiology Interpretation Agent (GPT-4o with vision, OpenAI)\u003c/b\u003e: Specialized for chest radiograph and CT pattern recognition including emphysema quantification, air trapping, mucus plugging, and pneumothorax identification; (viii) \u003cb\u003eCritical Care/ICU Agent (Claude 3.5 Sonnet, Anthropic)\u003c/b\u003e: Specialized for ventilatory management, dynamic hyperinflation, permissive hypercapnia strategies, and hemodynamic management in cor pulmonale.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003ch2\u003e2.3.2 Symbolic Reasoning Layer\u003c/h2\u003e \u003cp\u003eFollowing multi-agent deliberation, outputs were processed by a Prolog-based symbolic reasoning engine implementing a COPD/asthma management knowledge base comprising 312 formal rules derived from GOLD 2025 and GINA 2025 guideline decision trees. The symbolic layer performed: (a) logical consistency checking between agent outputs; (b) rule-based correction of responses violating formal guideline logic (e.g., initiating ICS monotherapy in COPD without asthma overlap or eosinophilic phenotype; prescribing β-blocker-containing eyedrops without asthma safety precaution documentation); and (c) confidence score assignment. Full architectural details are provided in Supplementary Material 2.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Physician Participants\u003c/h2\u003e \u003cp\u003eBoard-certified specialist physicians were recruited through purposive sampling from multiple tertiary care hospitals in the Ankara region between November 2025 and January 2026, in a multicenter design. Participating institutions and physician contributions were as follows: Ankara Bilkent City Hospital (n\u0026thinsp;=\u0026thinsp;10; 4 pulmonologists, 3 allergists and clinical immunologists, 3 emergency medicine physicians), Ankara Etlik City Hospital (n\u0026thinsp;=\u0026thinsp;10; 3 pulmonologists, 4 allergists and clinical immunologists, 3 emergency medicine physicians), Mamak State Hospital (n\u0026thinsp;=\u0026thinsp;5; 2 pulmonologists, 1 allergist and clinical immunologist, 2 emergency medicine physicians), and Etimesgut Şehit Sait Ert\u0026uuml;rk State Hospital (n\u0026thinsp;=\u0026thinsp;5; 1 pulmonologist, 2 allergists and clinical immunologists, 2 emergency medicine physicians). Eligibility criteria included board certification in the relevant specialty (pulmonology, allergology and clinical immunology, or emergency medicine), a minimum of three years of post-residency clinical practice, active clinical practice with patient contact in the preceding 12 months, and ability to read and respond to English-language clinical vignettes. Thirty physicians were enrolled in total: 10 pulmonologists, 10 allergists and clinical immunologists, and 10 emergency medicine physicians. Demographic and professional characteristics were collected by self-administered questionnaire. Physicians completed the 30-vignette assessment individually, without access to reference materials, under randomized domain order to mitigate sequencing effects.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Outcome Measures\u003c/h2\u003e \u003cp\u003eThe primary outcome was overall accuracy (proportion of correctly answered vignettes out of 30). Secondary outcomes included: (1) domain-specific accuracy; (2) response latency; (3) inter-rater reliability (Cohen's kappa and ICC); and (4) logistic regression analysis of predictors of correct responses among physician participants.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Statistical Analysis\u003c/h2\u003e \u003cp\u003eAll statistical analyses were conducted in R version 4.3.2. Non-parametric tests were used throughout given non-normal distribution of accuracy scores (Shapiro-Wilk, all p\u0026thinsp;\u0026lt;\u0026thinsp;0.05). Group comparisons employed the Kruskal-Wallis H test for omnibus comparisons and Dunn's post-hoc test with Bonferroni correction for pairwise comparisons. Effect sizes were estimated using Cohen's d from ranked data. NS-MAS vs physician group comparisons used one-sample Wilcoxon signed-rank tests against the physician group's median. Inter-rater reliability was assessed using Cohen's kappa (κ) and ICC (two-way mixed, consistency). Statistical significance was α\u0026thinsp;=\u0026thinsp;0.05 (two-tailed); all p-values are Bonferroni-adjusted for domain-level comparisons.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003ch2\u003e3.1 Participant Characteristics\u003c/h2\u003e\n\u003cp\u003eThirty specialist physicians completed the study: 10 pulmonologists, 10 allergists and clinical immunologists, and 10 emergency medicine physicians. Mean age ranged from 38.9 \u0026plusmn; 5.7 years in the emergency medicine group to 43.8 \u0026plusmn; 6.9 years in the pulmonology group. Mean post-residency experience was 13.6 \u0026plusmn; 6.5 years for pulmonologists, 10.4 \u0026plusmn; 5.1 years for allergists, and 9.4 \u0026plusmn; 4.3 years for emergency physicians. Emergency medicine physicians managed the highest median annual obstructive airway disease caseload (median 220, IQR 170\u0026ndash;300), reflecting acute exacerbation volume rather than longitudinal disease management. Self-reported familiarity with the GOLD 2025 guideline was highest among pulmonologists (mean 4.3 \u0026plusmn; 0.5 on a 5-point scale), whereas familiarity with the GINA 2025 guideline was highest among allergists (mean 4.4 \u0026plusmn; 0.4). Recent guideline-related continuing medical education attendance in the past two years was reported by 70% of pulmonologists, 70% of allergists, and 50% of emergency physicians. Full demographic and professional characteristics are summarized in Table 1.\u003c/p\u003e\n\u003ch2\u003e3.2 Overall Accuracy\u003c/h2\u003e\n\u003cp\u003eThe NS-MAS achieved an overall accuracy of 90.0% (27 correct out of 30 questions), significantly exceeding all three physician groups (Table 2, Figure 2). Among specialist physicians, allergists and clinical immunologists demonstrated marginally higher overall accuracy (mean 22.8 \u0026plusmn; 2.2 out of 30, 76.0%) compared with pulmonologists (mean 22.5 \u0026plusmn; 2.1 out of 30, 75.0%), while emergency physicians performed lowest across all human groups (mean 18.6 \u0026plusmn; 2.0 out of 30, 62.0%). The omnibus Kruskal-Wallis test across all four groups revealed a highly significant difference (H = 29.1, df = 3, p \u0026lt; 0.001). Pairwise comparisons between NS-MAS and each physician group all reached statistical significance at Bonferroni-adjusted thresholds: NS-MAS versus pulmonologists (median difference +4.5 questions, p = 0.002, Cohen\u0026apos;s d = 1.82, large effect), NS-MAS versus allergists (median difference +4.2 questions, p = 0.003, d = 1.74, large effect), and NS-MAS versus emergency physicians (median difference +8.4 questions, p \u0026lt; 0.001, d = 2.84, large effect). The difference in overall accuracy between pulmonologists and allergists did not reach statistical significance (p = 0.412, d = 0.14), whereas both groups significantly outperformed emergency physicians (pulmonology vs emergency: p = 0.009, d = 0.92; allergist vs emergency: p = 0.004, d = 1.08). Full pairwise comparisons and effect size estimates are presented in Table 3.\u003c/p\u003e\n\u003ch2\u003e3.3 Domain-Specific Performance\u003c/h2\u003e\n\u003cp\u003eDomain-level accuracy results are presented in Table 2 and displayed as a radar chart in Figure 3. Several clinically important specialty-expertise interactions were identified across the six management domains.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3.1 Diagnosis and Differential Diagnosis\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe NS-MAS achieved perfect accuracy in the diagnostic domain (5/5, 100%). Pulmonologists (mean 4.4 \u0026plusmn; 0.5 out of 5, 88%) and allergists (mean 4.2 \u0026plusmn; 0.6 out of 5, 84%) demonstrated comparable performance, both significantly exceeding emergency physicians (mean 3.4 \u0026plusmn; 0.5 out of 5, 68%; p \u0026lt; 0.001 for both comparisons). The NS-MAS advantage over all physician groups combined was large (mean difference +2.0 questions per 5, p \u0026lt; 0.001, d = 2.02). Common diagnostic errors among emergency physicians included misclassification of cardiac-origin dyspnoea as COPD exacerbation in elderly patients with preserved spirometry and failure to identify asthma-COPD overlap (ACO) in patients with features of both conditions.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3.2 Severity Assessment\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe NS-MAS answered all five severity assessment questions correctly (5/5, 100%). Allergists demonstrated marginally higher accuracy (mean 4.3 \u0026plusmn; 0.5 out of 5, 86%) compared with pulmonologists (mean 4.1 \u0026plusmn; 0.6 out of 5, 82%), with both groups significantly outperforming emergency physicians (mean 3.5 \u0026plusmn; 0.7 out of 5, 70%; Kruskal-Wallis p = 0.008). The most frequent error pattern involved misapplication of the updated GOLD ABE classification \u0026mdash; specifically, conflation of the abolished GOLD category C with the current mMRC/CAT-based GOLD B group \u0026mdash; and underrecognition of severe or life-threatening acute asthma per BTS/SIGN criteria.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3.3 Pharmacological Management\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe pharmacological management domain was the only domain in which a physician group numerically exceeded NS-MAS accuracy: allergists achieved 82% (mean 4.1 \u0026plusmn; 0.5 out of 5) compared with the NS-MAS score of 80% (4/5), though this difference did not reach statistical significance (p = 0.680). Pulmonologists scored 76% (mean 3.8 \u0026plusmn; 0.6 out of 5) and emergency physicians 60% (mean 3.0 \u0026plusmn; 0.7 out of 5). The overall domain Kruskal-Wallis test was significant (p = 0.004). Key discrimination errors included failure to initiate triple ICS/LABA/LAMA therapy in symptomatic GOLD Group E patients already receiving dual LAMA/LABA, prescription of ICS monotherapy in COPD without eosinophilic phenotype evidence, and incorrect biologic selection in severe uncontrolled asthma.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3.4 Exacerbation Management\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis domain revealed the most pronounced specialty-expertise interaction among physician groups. The NS-MAS achieved perfect accuracy (5/5, 100%). Allergists demonstrated significant superiority over pulmonologists (mean 4.2 \u0026plusmn; 0.6 out of 5, 84% vs mean 3.6 \u0026plusmn; 0.7 out of 5, 72%; p = 0.003, d = 0.93, large effect), particularly in acute severe asthma management scenarios involving intravenous magnesium sulfate dosing, heliox indications, and NIV thresholds in near-fatal asthma. Emergency physicians performed most poorly in this domain (mean 2.8 \u0026plusmn; 0.6 out of 5, 56%; Kruskal-Wallis p \u0026lt; 0.001), consistent with suboptimal adherence to algorithmic antibiotic prescribing in COPD exacerbations per Anthonisen criteria and underuse of systemic corticosteroids at guideline-recommended doses and durations (Figure 3).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3.5 ICU Admission and Ventilatory Management\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eIn contrast to the exacerbation domain, pulmonologists outperformed all other physician groups in ICU and ventilatory management (mean 3.3 \u0026plusmn; 0.8 out of 5, 66%), though the NS-MAS again achieved perfect accuracy (5/5, 100%). Allergists demonstrated the lowest performance among specialist groups in this domain (mean 2.9 \u0026plusmn; 0.7 out of 5, 58%), significantly below pulmonologists (p \u0026lt; 0.001, d = 0.54, medium effect). Emergency physicians scored 62% (mean 3.1 \u0026plusmn; 0.7 out of 5). Errors among allergists concentrated in NIV initiation thresholds for COPD-related hypercapnic respiratory failure, permissive hypercapnia strategy in mechanically ventilated near-fatal asthma, and dynamic hyperinflation recognition and management. The bidirectional expertise asymmetry between allergists and pulmonologists across the exacerbation and ICU/ventilatory domains is visually apparent in the radar chart (Figure 3) and statistically confirmed in Table 3.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e\u003cem\u003e3.3.6 Complication Management\u003c/em\u003e\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eComplication management was the only domain in which the NS-MAS underperformed relative to at least one physician group. The NS-MAS scored 60% (3/5), being equaled by pulmonologists (mean 3.3 \u0026plusmn; 0.7 out of 5, 66%). Allergists scored 62% (mean 3.1 \u0026plusmn; 0.8 out of 5) and emergency physicians 56% (mean 2.8 \u0026plusmn; 0.6 out of 5), the lowest of any group in this domain (Kruskal-Wallis p = 0.005). The NS-MAS errors occurred in two scenarios: one involving individualized dose-tapering decisions for corticosteroid-dependent asthma with adrenal insufficiency risk and one involving complex cor pulmonale exacerbation management requiring integration of right ventricular function assessment, diuretic titration, and long-term oxygen therapy prescription rationale. Notably, no physician group achieved high accuracy in this domain (range 56\u0026ndash;66%), confirming the intrinsic difficulty of complication management reasoning across all groups.\u003c/p\u003e\n\u003cp\u003ePer-question accuracy proportions across all 30 vignettes and all four groups are displayed in the heatmap in Figure 5. The visual analysis confirms domain-specific expertise clustering: questions 14 and 15 (exacerbation management, antimicrobial decision-making) and question 25 (ICU/ventilatory management, dynamic hyperinflation) showed the widest intergroup accuracy disparities, underscoring their discriminative value.\u003c/p\u003e\n\u003ch2\u003e3.4 Response Latency\u003c/h2\u003e\n\u003cp\u003eNS-MAS generated responses with a median latency of 13.1 seconds per question (IQR 10.2\u0026ndash;17.4 seconds), representing a 13-fold reduction compared with the fastest physician group (emergency physicians, median 2.91 minutes per question, IQR 2.2\u0026ndash;3.5 minutes). Pulmonologists had the longest median response time (3.47 minutes, IQR 2.6\u0026ndash;4.3 minutes), followed by allergists (3.12 minutes, IQR 2.3\u0026ndash;4.0 minutes). The between-group latency difference was highly significant (Kruskal-Wallis H = 1762.4, p \u0026lt; 0.001), as illustrated in Figure 4. NS-MAS latency did not significantly differ across the six clinical domains (Kruskal-Wallis p = 0.19), indicating domain-agnostic processing efficiency irrespective of query complexity.\u003c/p\u003e\n\u003ch2\u003e3.5 Inter-Rater Reliability Among Physicians\u003c/h2\u003e\n\u003cp\u003eWithin-group pairwise Cohen\u0026apos;s kappa values were in the fair-to-moderate range for all specialty groups: pulmonologists \u0026kappa; = 0.49 (95% CI 0.41\u0026ndash;0.57), allergists \u0026kappa; = 0.46 (95% CI 0.38\u0026ndash;0.54), and emergency physicians \u0026kappa; = 0.36 (95% CI 0.27\u0026ndash;0.45). The overall intraclass correlation coefficient (ICC, two-way mixed, consistency) across all 30 physicians was 0.42 (95% CI 0.34\u0026ndash;0.50), indicating moderate inter-rater agreement. These values quantify the substantial inter-individual variability in clinical obstructive airway disease decision-making that current practice tolerates even within specialty-defined peer groups. In contrast, the NS-MAS produced identical responses across three independent runs for 28 of 30 questions (93.3% consistency). The two inconsistent questions (Q7, severity assessment, GINA symptom control classification borderline; Q28, complication management, corticosteroid tapering) corresponded to scenarios in which the GINA 2025 guideline itself acknowledges boundary-condition ambiguity.\u003c/p\u003e\n\u003ch2\u003e3.6 Predictors of Correct Response\u003c/h2\u003e\n\u003cp\u003eIn multivariable mixed-effects logistic regression modeling correct response at the question level among physicians (n = 900 observations: 30 physicians \u0026times; 30 questions; random intercept per physician), the following independent predictors of accuracy were identified: specialty group membership (allergist vs emergency medicine: OR 1.94, 95% CI 1.43\u0026ndash;2.63, p \u0026lt; 0.001; pulmonology vs emergency medicine: OR 1.86, 95% CI 1.38\u0026ndash;2.51, p \u0026lt; 0.001); post-residency years of experience (OR 1.05 per year, 95% CI 1.01\u0026ndash;1.10, p = 0.024); self-reported GOLD 2025 guideline familiarity (OR 1.38 per unit, 95% CI 1.14\u0026ndash;1.67, p = 0.001); self-reported GINA 2025 guideline familiarity (OR 1.42 per unit, 95% CI 1.17\u0026ndash;1.72, p \u0026lt; 0.001); and recent guideline-related CME attendance (OR 1.58, 95% CI 1.12\u0026ndash;2.23, p = 0.009). Clinical domain was also a significant predictor: physicians were most likely to answer correctly in the diagnosis domain (reference category) and least likely in ICU and ventilatory management (OR 0.48, 95% CI 0.36\u0026ndash;0.64, p \u0026lt; 0.001) and exacerbation management for non-allergist physicians (OR 0.54, 95% CI 0.41\u0026ndash;0.72, p \u0026lt; 0.001). Annual obstructive airway disease caseload was not an independent predictor after adjustment for specialty group and guideline familiarity (OR 1.00, 95% CI 0.99\u0026ndash;1.01, p = 0.612). Full regression results are presented in Table 4.\u003c/p\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThis study demonstrates that a neurosymbolic multi-agent LLM system can achieve comprehensive, guideline-concordant performance across the full spectrum of COPD and asthma management clinical decision domains, surpassing all specialist physician groups in overall accuracy and exceeding each group within its respective area of primary expertise, with the singular exception of the complication management domain. The NS-MAS overall accuracy of 90.0% significantly exceeded pulmonologists (75.0%, p = 0.002, d = 1.82), allergists and clinical immunologists (76.0%, p = 0.003, d = 1.74), and emergency physicians (62.0%, p \u0026lt; 0.001, d = 2.84), with large effect sizes across all comparisons. These findings advance understanding of the conditions under which neurosymbolic AI clinical decision support systems may offer additive value relative to individual specialist judgment in obstructive airway disease management.\u003c/p\u003e\n\u003cp\u003eThe superiority of the NS-MAS across five of six clinical domains likely reflects its capacity to simultaneously integrate the GOLD 2025 ABE framework, GINA 2025 stepwise therapy algorithms, and asthma-COPD overlap (ACO) diagnostic criteria through parallel agent reasoning. This cognitive integration represents a genuine challenge even for clinicians primarily trained in one of the two guideline ecosystems. Recent real-world analyses continue to show suboptimal implementation of GOLD and GINA recommendations, with guideline-concordant care often clustering around one-half of cases in routine practice and varying substantially by care setting, specialty, and the treatment domain assessed\u0026nbsp;[15]. The NS-MAS generated identical responses across three independent runs for 28 of 30 questions, yielding 93.3% consistency, while inter-rater reliability across physician groups ranged from \u0026kappa; = 0.36 to 0.49. These findings quantify the substantial degree of decision variability tolerated in routine clinical practice and underscore the operational relevance of reproducible algorithmic decision support.\u003c/p\u003e\n\u003cp\u003eThe most clinically significant finding is the bidirectional expertise asymmetry between pulmonologists and allergists across the exacerbation and ICU/ventilatory management domains. Allergists significantly outperformed pulmonologists in exacerbation management (84% vs 72%, p = 0.003, d = 0.93), particularly in acute severe asthma scenarios requiring intravenous magnesium sulfate dosing and near-fatal asthma NIV thresholds, competencies anchored in allergist-specific training frameworks. Conversely, pulmonologists outperformed allergists in ICU and ventilatory management (66% vs 58%, p \u0026lt; 0.001, d = 0.54), reflecting their greater exposure to permissive hypercapnia strategy, dynamic hyperinflation recognition, and COPD-specific NIV pH-guided protocols. This pattern is consistent with published evidence showing that physicians\u0026rsquo; knowledge and care patterns remain strongly shaped by specialty-specific domains of expertise, while communication and interpretive differences between specialties create persistent knowledge boundaries even among highly trained clinicians[16].\u0026nbsp;The NS-MAS resolved this asymmetry by simultaneously applying both GOLD 2025\u0026nbsp;and GINA 2025 reasoning streams, achieving perfect accuracy (100%) in both domains.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe pharmacological management domain was the only domain in which a physician group numerically exceeded NS-MAS performance, with allergists achieving 82% versus the NS-MAS score of 80%, though this difference did not reach statistical significance (p = 0.680). Allergists\u0026apos; superior biologic therapy selection, driven by daily clinical exposure to blood eosinophil thresholds, fractional exhaled nitric oxide (FeNO) interpretation, and IgE-guided omalizumab dosing, likely represents a form of tacit clinical expertise that current RAG-augmented agent architectures do not fully capture. The GINA 2025 add-on therapy algorithm involves nuanced biomarker integration requiring contextual patient-level judgment that extends beyond structured algorithmic logic, as has been observed with other precision medicine frameworks [17, 18]. This finding identifies a specific area for future NS-MAS architecture refinement, particularly regarding real-time biomarker integration.\u003c/p\u003e\n\u003cp\u003eComplication management was the only domain in which the NS-MAS underperformed relative to at least one physician group, scoring 60% (3/5) compared with pulmonologists at 66%. NS-MAS errors occurred in two high-complexity scenarios: one involving individualized corticosteroid dose-tapering decisions in asthma with adrenal insufficiency risk, and one involving cor pulmonale exacerbation management requiring integration of right ventricular function, diuretic titration, and long-term oxygen therapy rationale. This aligns with a growing body of literature demonstrating that AI decision support performs less optimally in prognostically ambiguous, contextually nuanced scenarios compared with structured algorithmic decisions\u0026nbsp;[19]. Critically, no physician group achieved high accuracy in this domain (range 56\u0026ndash;66%), confirming that complication management in obstructive airway disease represents a genuine zone of clinical uncertainty for both AI systems and human specialists.\u003c/p\u003e\n\u003cp\u003eEmergency physicians demonstrated the lowest overall accuracy (62.0%) and the largest performance gap relative to NS-MAS (d = 2.84). This finding is clinically relevant given that emergency departments represent the primary acute contact point for COPD and asthma exacerbations, with emergency physicians managing the highest annual caseload (median 220 cases per year, IQR 170\u0026ndash;300) in this study. Suboptimal adherence to Anthonisen antibiotic prescribing criteria and underuse of systemic corticosteroids at guideline-recommended doses and durations were the predominant error patterns, consistent with previously published audits of emergency department COPD management [15]. The 13-fold response latency advantage of NS-MAS (median 13.1 seconds vs 2.91 minutes for emergency physicians, Kruskal-Wallis H = 1762.4, p \u0026lt; 0.001) carries direct clinical relevance: in acute severe asthma, escalation to intravenous magnesium sulfate is recommended when there is inadequate response to intensive initial therapy, and earlier administration has been associated with improved emergency department process measures and short-term clinical response; in COPD exacerbation, delayed NIV initiation is associated with higher in-hospital mortality\u0026nbsp;[20].\u003c/p\u003e\n\u003cp\u003eThe NS-MAS architecture evaluated in this study differs from prior single-agent LLM benchmarking studies in several important ways. First, the agent specialization structure mirrors multidisciplinary team reasoning, distributing domain expertise across dedicated agents rather than requiring a single model to carry all obstructive airway disease subspecialty knowledge. Second, the symbolic reasoning layer, a Prolog-based engine implementing 312 formal rules derived from GOLD 2025 and GINA 2025 decision trees, provides formal guideline-rule enforcement that single neural models cannot guarantee, addressing a core limitation of current LLMs in safety-critical medical applications [21, 22]. Third, the LangGraph orchestration enables traceable, auditable reasoning pathways essential for clinical deployment trustworthiness. Fourth, for ACO queries the dual co-activation of Pulmonology and Allergy/Immunology agents simultaneously resolved the otherwise systematic knowledge gap observed in each single-specialty physician group.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eMultivariable mixed-effects logistic regression identified GOLD 2025 guideline familiarity (OR 1.38 per unit, p = 0.001), GINA 2025 guideline familiarity (OR 1.42 per unit, p \u0026lt; 0.001), and recent CME attendance (OR 1.58, p = 0.009) as independent predictors of correct response among physicians, after adjustment for specialty group membership and post-residency experience. Annual caseload was not an independent predictor (OR 1.00, p = 0.612), suggesting that volume of clinical exposure does not compensate for guideline-specific knowledge currency. This finding supports targeted continuing medical education aligned with current GOLD and GINA updates, together with structured AI training and supervised AI-assisted decision support, as an integrated strategy for improving the quality, consistency, and safety of obstructive airway disease management[23].\u003c/p\u003e\n\u003cp\u003eSeveral limitations require acknowledgment. First, vignette-based evaluation does not capture the full complexity of real-world clinical decision-making, which incorporates dynamic patient trajectories, nonverbal clinical information, shared decision-making, and systemic healthcare context. Second, the NS-MAS was RAG-augmented with full GOLD 2025 and GINA 2025 guideline texts while physicians had no reference access, a condition that favors the AI system but reflects its intended clinical deployment context. Third, although the study employed a multicenter design across four institutions in the Ankara region (Ankara Bilkent City Hospital, Ankara Etlik City Hospital, Mamak State Hospital, and Etimesgut Şehit Sait Ert\u0026uuml;rk State Hospital), all participating centers were geographically confined to a single metropolitan area (Ankara, Turkey), and the physician sample of 10 per specialty may not represent the full distribution of specialist competence nationally or globally; future studies should recruit from geographically diverse institutions across multiple regions and healthcare systems. Fourth, LLM model versions and guideline content evolve rapidly, and performance benchmarks require periodic recalibration as both GOLD and GINA issue annual updates. Fifth, this study does not address implementation feasibility, clinician acceptance, alert fatigue, or patient-level outcome effects in deployed clinical settings, which should be the subject of prospective clinical trials.\u003c/p\u003e"},{"header":"5. Conclusions","content":"\u003cp\u003eA neurosymbolic multi-agent LLM system orchestrating eight specialized agents demonstrated significantly superior guideline-concordant COPD and asthma management performance relative to board-certified pulmonologists, allergists and clinical immunologists, and emergency physicians across the majority of clinical decision domains. The system achieved particular advantages in domains where individual specialty expertise is most limited, namely ICU and ventilatory management for allergists, and exacerbation-specific pharmacological management for pulmonologists and emergency physicians. The complication management domain represented the one area where pulmonologist physicians equaled or exceeded NS-MAS performance, suggesting that contextually complex, longitudinally nuanced management scenarios remain a relative limitation of current multi-agent architectures. These findings provide foundational evidence for the clinical decision support potential of neurosymbolic multi-agent AI in obstructive airway disease management, warranting prospective validation in real-world acute and chronic care settings.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eEthics Approval and Consent to Participate\u003c/h2\u003e\n\u003cp\u003eThis study was approved by the Ankara İl Sağlık M\u0026uuml;d\u0026uuml;rl\u0026uuml;ğ\u0026uuml; M\u0026uuml;dahalesiz Etik Kurul (Non-Interventional Ethics Board; Approval No: 2025-10-3, Date: 24 October 2025). All physician participants provided written informed consent prior to participation. No patient data were collected or used in this study.\u003c/p\u003e\n\u003ch2\u003eAvailability of Data and Materials\u003c/h2\u003e\n\u003cp\u003eThe full clinical vignette bank (Supplementary Material 1) and NS-MAS prompt architectures (Supplementary Material 2) are provided with this manuscript. Anonymized physician response data are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003ch2\u003eCompeting Interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare no competing interests. No external funding was received from AI technology companies. The use of commercial LLM APIs (OpenAI, Anthropic, Google DeepMind) was self-funded by the research team.\u003c/p\u003e\n\u003ch2\u003eFunding\u003c/h2\u003e\n\u003cp\u003eThis study received no dedicated external funding. Research was conducted as part of the doctoral research program of M.T.\u0026Uuml;. at Hacettepe University Faculty of Medicine.\u003c/p\u003e\n\u003ch2\u003eAuthors\u0026apos; Contributions\u003c/h2\u003e\n\u003cp\u003eE.A.M. contributed to study design, clinical vignette development, data collection, and critical manuscript revision. M.T.\u0026Uuml;. led study design, NS-MAS development and implementation, statistical analysis, and manuscript writing. E.E. contributed to clinical vignette development (emergency medicine domain), data collection, and manuscript revision. All authors read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eCao, Z., et al., \u003cem\u003eBurden of chronic obstructive pulmonary disease and its attributable risk factors in 204 countries and territories, 1990-2021: results from the Global Burden of Disease Study 2021.\u003c/em\u003e BMJ Public Health, 2026. \u003cstrong\u003e4\u003c/strong\u003e(1): p. e002489.\u003c/li\u003e\n\u003cli\u003eZhang, L., et al., \u003cem\u003eGlobal, regional and national burden of asthma from 1990 to 2021: a systematic analysis for the Global Burden of Disease Study 2021.\u003c/em\u003e BMJ Open Respiratory Research, 2025. \u003cstrong\u003e12\u003c/strong\u003e(1): p. e003144.\u003c/li\u003e\n\u003cli\u003eFricker, M. and R. Lokwani, \u003cem\u003eCOPD: the role of neutrophils in inflammation, pathophysiology, and as drug targets.\u003c/em\u003e Clin Sci (Lond), 2025. \u003cstrong\u003e139\u003c/strong\u003e(20): p. 1199-1214.\u003c/li\u003e\n\u003cli\u003eMatsunaga, K., et al., \u003cem\u003eGuidance for type 2 inflammatory biomarkers.\u003c/em\u003e Respir Investig, 2025. \u003cstrong\u003e63\u003c/strong\u003e(3): p. 273-288.\u003c/li\u003e\n\u003cli\u003eChan, R., N.E. Horn, and S. Siddiqui, \u003cem\u003ePrecision Medicine for Asthma: Tailored to its Severity and Endotype/Phenotype.\u003c/em\u003e Allergy Asthma Immunol Res, 2026. \u003cstrong\u003e18\u003c/strong\u003e(1): p. 19-38.\u003c/li\u003e\n\u003cli\u003eXie, C., et al., \u003cem\u003eToward precision medicine in COPD: phenotypes, endotypes, biomarkers, and treatable traits.\u003c/em\u003e Respiratory Research, 2025. \u003cstrong\u003e26\u003c/strong\u003e(1): p. 274.\u003c/li\u003e\n\u003cli\u003eGayen, S., et al., \u003cem\u003eCritical Care Management of Severe Asthma Exacerbations.\u003c/em\u003e Journal of Clinical Medicine, 2024. \u003cstrong\u003e13\u003c/strong\u003e(3): p. 859.\u003c/li\u003e\n\u003cli\u003eHalpin, D.M.G., et al., \u003cem\u003eAddressing the Global Challenges of COPD and Asthma: A Shared Vision From the Global Initiative for Chronic Obstructive Pulmonary Disease (GOLD) and the Global Initiative for Asthma (GINA).\u003c/em\u003e Respirology, 2026. \u003cstrong\u003e31\u003c/strong\u003e(2): p. 135-140.\u003c/li\u003e\n\u003cli\u003eChen, S.F., et al., \u003cem\u003eLLM-assisted systematic review of large language models in clinical medicine.\u003c/em\u003e Nature Medicine, 2026. \u003cstrong\u003e32\u003c/strong\u003e(3): p. 1152-1159.\u003c/li\u003e\n\u003cli\u003ePhan, L., et al., \u003cem\u003eA benchmark of expert-level academic questions to assess AI capabilities.\u003c/em\u003e Nature, 2026. \u003cstrong\u003e649\u003c/strong\u003e(8099): p. 1139-1146.\u003c/li\u003e\n\u003cli\u003eUcdal, M., K. Yurtsever, and E. Ekingen, \u003cem\u003eMultidisciplinary artificial intelligence systems versus single-model approaches for the diagnosis and management of ileus and volvulus.\u003c/em\u003e BMC Gastroenterol, 2026. \u003cstrong\u003e26\u003c/strong\u003e(1): p. 124.\u003c/li\u003e\n\u003cli\u003eSorka, M., et al., \u003cem\u003eA multi-agent approach to neurological clinical reasoning.\u003c/em\u003e PLOS Digit Health, 2025. \u003cstrong\u003e4\u003c/strong\u003e(12): p. e0001106.\u003c/li\u003e\n\u003cli\u003eAcharya, K. and H. Song, \u003cem\u003eA Comprehensive Review of Neuro-symbolic AI for Robustness, Uncertainty Quantification, and Intervenability.\u003c/em\u003e Arabian Journal for Science and Engineering, 2026. \u003cstrong\u003e51\u003c/strong\u003e(1): p. 35-67.\u003c/li\u003e\n\u003cli\u003eEkingen, E. and M. Ucdal, \u003cem\u003ePerformance Comparison of a Neuro-Symbolic Large Language Model System Versus Human Experts in Acute Cholecystitis Management.\u003c/em\u003e J Clin Med, 2026. \u003cstrong\u003e15\u003c/strong\u003e(5).\u003c/li\u003e\n\u003cli\u003eCuperus, L.J.A., et al., \u003cem\u003eAdherence to Inhaled Medication Guidelines in Patients with COPD: Evaluating the GOLD 2023 Strategy in a Real-World Secondary Care Setting.\u003c/em\u003e Int J Chron Obstruct Pulmon Dis, 2025. \u003cstrong\u003e20\u003c/strong\u003e: p. 3877-3891.\u003c/li\u003e\n\u003cli\u003eBraam, A., et al., \u003cem\u003eCollaboration Between Physicians from Different Medical Specialties in Hospital Settings: A Systematic Review.\u003c/em\u003e J Multidiscip Healthc, 2022. \u003cstrong\u003e15\u003c/strong\u003e: p. 2277-2300.\u003c/li\u003e\n\u003cli\u003eBalasubramanyam, S., E.K. George, and E. Wang, \u003cem\u003ePrecision medicine and choosing a biologic in asthma: understanding the current state of knowledge for predictors of response and clinical remission.\u003c/em\u003e Curr Opin Allergy Clin Immunol, 2025. \u003cstrong\u003e25\u003c/strong\u003e(1): p. 66-74.\u003c/li\u003e\n\u003cli\u003eQuek, E., N. Horn, and S. Siddiqui, \u003cem\u003ePrecision Medicine in Asthma: The Role of Biomarkers.\u003c/em\u003e Immunotargets Ther, 2025. \u003cstrong\u003e14\u003c/strong\u003e: p. 1479-1513.\u003c/li\u003e\n\u003cli\u003eBean, A.M., et al., \u003cem\u003eReliability of LLMs as medical assistants for the general public: a randomized preregistered study.\u003c/em\u003e Nature Medicine, 2026. \u003cstrong\u003e32\u003c/strong\u003e(2): p. 609-615.\u003c/li\u003e\n\u003cli\u003eAlzain, M., et al., \u003cem\u003eImpact of early vs. delayed intravenous magnesium sulfate on clinical outcomes in pediatric severe asthma: a retrospective cohort study.\u003c/em\u003e Frontiers in Disaster and Emergency Medicine, 2025. \u003cstrong\u003eVolume 3 - 2025\u003c/strong\u003e.\u003c/li\u003e\n\u003cli\u003eLiu, Q., et al., \u003cem\u003eEvoMDT: a self-evolving multi-agent system for structured clinical decision-making in multi-cancer.\u003c/em\u003e npj Digital Medicine, 2026. \u003cstrong\u003e9\u003c/strong\u003e(1): p. 124.\u003c/li\u003e\n\u003cli\u003eTan, X., et al. \u003cem\u003eProlog-Driven Rule-Based Diagnostics with Large Language Models for Precise Clinical Decision Support\u003c/em\u003e. 2026. Cham: Springer Nature Switzerland.\u003c/li\u003e\n\u003cli\u003eQunaibi, E.A., et al., \u003cem\u003eEffectiveness of Informed AI Use on Clinical Competence of General Practitioners and Internists: Pre-Post Intervention Study.\u003c/em\u003e JMIR Med Educ, 2026. \u003cstrong\u003e12\u003c/strong\u003e: p. e75534.\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003e\u003cstrong\u003eTable 1. Participant Demographic and Professional Characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"633\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003e\u003cstrong\u003eCharacteristic\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u003cstrong\u003ePulmonology (n=10)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u003cstrong\u003eAllergist/Immunology (n=10)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u003cstrong\u003eEmergency Medicine (n=10)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003e\u003cstrong\u003e\u003cem\u003eDemographics\u003c/em\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eAge, years, mean \u0026plusmn; SD\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e43.8 \u0026plusmn; 6.9\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e40.2 \u0026plusmn; 6.1\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e38.9 \u0026plusmn; 5.7\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eGender, n (%) Female\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e3 (30%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e5 (50%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e4 (40%)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003ePost-residency experience, years\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e13.6 \u0026plusmn; 6.5\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e10.4 \u0026plusmn; 5.1\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e9.4 \u0026plusmn; 4.3\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003e\u003cstrong\u003e\u003cem\u003eAcademic \u0026amp; Institutional Profile\u003c/em\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eAttending/Specialist, n (%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e8 (80%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e8 (80%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e9 (90%)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eAssociate Professor or above\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e3 (30%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e2 (20%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e1 (10%)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eUniversity hospital, n (%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e6 (60%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e5 (50%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e4 (40%)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eAnnual COPD/Asthma cases, median (IQR)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e140 (100\u0026ndash;200)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e185 (140\u0026ndash;250)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e220 (170\u0026ndash;300)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003e\u003cstrong\u003e\u003cem\u003eGuideline Familiarity\u003c/em\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eGOLD 2025 COPD Guideline familiarity (1\u0026ndash;5)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e4.3 \u0026plusmn; 0.5\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e3.2 \u0026plusmn; 0.8\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e2.9 \u0026plusmn; 0.9\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eGINA 2025 Asthma Guideline familiarity (1\u0026ndash;5)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e3.9 \u0026plusmn; 0.6\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e4.4 \u0026plusmn; 0.4\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e2.7 \u0026plusmn; 0.8\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eCME in past 2 years, n (%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e7 (70%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e7 (70%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 140px;\"\u003e5 (50%)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eValues are mean \u0026plusmn; SD, median (IQR), or n (%) as appropriate. CME, continuing medical education\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Accuracy by Clinical Domain and Group\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"633\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e\u003cstrong\u003eDomain (5 questions each)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u003cstrong\u003eNS-MAS (n/5)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\u003cstrong\u003ePulmonology mean \u0026plusmn; SD\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e\u003cstrong\u003eAllergist/Immunol. mean \u0026plusmn; SD\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\u003cstrong\u003eEmergency mean \u0026plusmn; SD\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\u003cstrong\u003ep-value*\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e1. Diagnosis \u0026amp; differential\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e5/5 (100%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e4.4 \u0026plusmn; 0.5 (88%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e4.2 \u0026plusmn; 0.6 (84%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.4 \u0026plusmn; 0.5 (68%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e2. Severity assessment\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e5/5 (100%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e4.1 \u0026plusmn; 0.6 (82%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e4.3 \u0026plusmn; 0.5 (86%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.5 \u0026plusmn; 0.7 (70%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e0.008\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e3. Pharmacological mgmt.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e4/5 (80%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.8 \u0026plusmn; 0.6 (76%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e4.1 \u0026plusmn; 0.5 (82%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.0 \u0026plusmn; 0.7 (60%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e0.004\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e4. Exacerbation mgmt.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e5/5 (100%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.6 \u0026plusmn; 0.7 (72%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e4.2 \u0026plusmn; 0.6 (84%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e2.8 \u0026plusmn; 0.6 (56%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e5. ICU admission \u0026amp; ventilation\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e5/5 (100%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.3 \u0026plusmn; 0.8 (66%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e2.9 \u0026plusmn; 0.7 (58%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.1 \u0026plusmn; 0.7 (62%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e6. Complication management\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e3/5 (60%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e3.3 \u0026plusmn; 0.7 (66%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e3.1 \u0026plusmn; 0.8 (62%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e2.8 \u0026plusmn; 0.6 (56%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e0.005\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 167px;\"\u003e\u003cstrong\u003eOVERALL (out of 30)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e27/30 (90.0%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e22.5 \u0026plusmn; 2.1 (75.0%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 127px;\"\u003e22.8 \u0026plusmn; 2.2 (76.0%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e18.6 \u0026plusmn; 2.0 (62.0%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 87px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eData for physician groups are mean \u0026plusmn; SD (percentage). NS-MAS data are raw score/5 (percentage). *Kruskal-Wallis test across all four groups; p-values Bonferroni-adjusted for six domain comparisons (threshold 0.0083). ICU, intensive care unit; NS-MAS, neurosymbolic multi-agent system.\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 3. Pairwise Comparisons and Effect Sizes\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"633\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003e\u003cstrong\u003eComparison\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\u003cstrong\u003eMedian diff. (questions)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u003cstrong\u003eU statistic\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\u003cstrong\u003ep-value (adj.)*\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e\u003cstrong\u003eEffect size (Cohen\u0026apos;s d)\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003e\u003cstrong\u003e\u003cem\u003eNS-MAS vs specialist groups\u003c/em\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eNS-MAS vs pulmonology\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+4.5 (90% vs 75%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u0026mdash;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.002\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e1.82 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eNS-MAS vs allergist/immunol.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+4.2 (90% vs 76%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u0026mdash;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.003\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e1.74 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eNS-MAS vs emergency med.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+8.4 (90% vs 62%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u0026mdash;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e2.84 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003e\u003cstrong\u003e\u003cem\u003eBetween specialist groups\u003c/em\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003ePulmonology vs allergist\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\u0026minus;0.3 (75% vs 76%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e44.1\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.412 (NS)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e0.14 (trivial)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003ePulmonology vs emergency\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+3.9 (75% vs 62%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e26.8\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.009\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e0.92 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eAllergist vs emergency\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+4.2 (76% vs 62%)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e23.9\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.004\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e1.08 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003e\u003cstrong\u003e\u003cem\u003eDomain-specific interactions\u003c/em\u003e\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e\u0026nbsp;\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eDiagnosis: NS-MAS vs all phys.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+2.0 per 5 questions\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e\u0026mdash;\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e2.02 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eExacerbation: allergy vs pulm.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+0.6 (4.2 vs 3.6)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e30.4\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.003\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e0.93 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eICU/vent: pulm. vs allergy\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+0.4 (3.3 vs 2.9)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e34.1\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e0.54 (medium)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003ePharmacology: allergy vs emerg.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+1.1 (4.1 vs 3.0)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e31.2\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.007\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e0.82 (large)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 180px;\"\u003eComplication: pulm. vs emerg.\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 113px;\"\u003e+0.5 (3.3 vs 2.8)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 80px;\"\u003e35.8\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 107px;\"\u003e0.028\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 153px;\"\u003e0.68 (medium)\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003eDunn post-hoc test with Bonferroni correction for pairwise comparisons. For NS-MAS vs physician groups, one-sample Wilcoxon signed-rank test against physician group median. Effect size interpretation: trivial (d \u0026lt; 0.2), small (d 0.2\u0026ndash;0.5), medium (d 0.5\u0026ndash;0.8), large (d \u0026gt; 0.8). *For non-allergist physicians in the exacerbation domain. NS, not significant; OR, odds ratio.\u003c/em\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 4. Multivariable Mixed-Effects Logistic Regression: Predictors of Correct Response Among Physicians\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"633\" class=\"fr-table-selection-hover\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003e\u003cstrong\u003ePredictor\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u003cstrong\u003eOR\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u003cstrong\u003e95% CI\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u003cstrong\u003ep-value\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003e\u003cstrong\u003eInterpretation\u003c/strong\u003e\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eSpecialty: allergist vs emergency\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.94\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.43\u0026ndash;2.63\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003eLarge advantage\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eSpecialty: pulmonology vs emergency\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.86\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.38\u0026ndash;2.51\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003eLarge advantage\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003ePost-residency experience (per year)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.05\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.01\u0026ndash;1.10\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.024\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003eModest positive\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eGOLD 2025 guideline familiarity (per unit)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.38\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.14\u0026ndash;1.67\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003ePositive\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eGINA 2025 guideline familiarity (per unit)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.42\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.17\u0026ndash;1.72\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003ePositive\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eRecent CME attendance (yes vs no)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.58\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.12\u0026ndash;2.23\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.009\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003ePositive\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eDomain: ICU/ventilation vs diagnosis\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.48\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.36\u0026ndash;0.64\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003eHarder domain\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eDomain: exacerbation vs diagnosis*\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.54\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.41\u0026ndash;0.72\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e\u0026lt; 0.001\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003eHarder domain\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 213px;\"\u003eAnnual caseload (per case)\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e1.00\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.99\u0026ndash;1.01\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 100px;\"\u003e0.612\u003cbr\u003e\u003c/td\u003e\n \u003ctd style=\"width: 120px;\"\u003eNot significant\u003cbr\u003e\u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u003cem\u003en = 900 observations (30 physicians \u0026times; 30 questions); random intercept per physician. Reference categories: specialty = emergency medicine; domain = diagnosis. CME, continuing medical education; GOLD, Global Initiative for Chronic Obstructive Lung Disease; GINA, Global Initiative for Asthma; OR, odds ratio; CI, confidence interval.\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\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":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"large language model, multi-agent system, COPD, asthma, clinical decision support, artificial intelligence, guideline adherence, GOLD, GINA","lastPublishedDoi":"10.21203/rs.3.rs-9262455/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9262455/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\u003eChronic obstructive pulmonary disease (COPD) and asthma collectively affect over 500 million individuals worldwide and represent leading causes of respiratory-related morbidity, healthcare utilization, and preventable mortality. Optimal management of these conditions encompasses accurate diagnosis and differential diagnosis, severity stratification, stepwise pharmacological therapy, exacerbation management, ventilatory support decisions, and complication recognition, all of which require the integration of subspecialty expertise that is frequently fragmented across clinical settings. Neurosymbolic multi-agent large language model (LLM) systems, which integrate neural reasoning with formal symbolic logic and specialized agent architectures, represent a promising paradigm for bridging domain-specific expertise gaps across pulmonology, allergology, and emergency medicine.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eMethods\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe conducted a cross-sectional comparative evaluation using 30 USMLE-format clinical vignettes spanning six COPD and asthma management domains, namely diagnosis and differential diagnosis, severity assessment, pharmacological management, exacerbation management, ICU admission and ventilatory management, and complication management, all derived from the current GOLD 2025COPD and GINA 2025 Asthma guidelines. A neurosymbolic multi-agent system (NS-MAS) comprising a LangGraph orchestrator coordinating eight specialized agents, including GPT-4.5, Claude Sonnet 4.6, and Gemini 2.5 Pro, was compared against 30 board-certified specialist physicians consisting of 10 pulmonologists, 10 allergists and clinical immunologists, and 10 emergency medicine physicians. The primary outcome was overall accuracy. Statistical analyses included the Kruskal-Wallis test, Dunn's post-hoc test with Bonferroni correction, and effect size estimation using Cohen's d.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eResults\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNS-MAS achieved an overall accuracy of 90.0% (27/30), significantly exceeding all physician groups: pulmonologists 75.0% (mean 22.5 ± 2.1/30, p = 0.002, d = 1.82), allergists/immunologists 76.0% (mean 22.8 ± 2.2/30, p = 0.003, d = 1.74), and emergency physicians 62.0% (mean 18.6 ± 2.0/30, p \u0026lt; 0.001, d = 2.84). Domain-specific analysis revealed significant specialty-expertise interactions: allergists outperformed pulmonologists in exacerbation management (84% vs 72%, p = 0.003) and pharmacological management (82% vs 76%, p = 0.018), while pulmonologists demonstrated superior performance in ICU/ventilatory management (66% vs 58%, p \u0026lt; 0.001) and complication management domains (66% vs 62%, p = 0.028). NS-MAS response latency (median 13.1 seconds) was significantly shorter than all physician groups (median 2.9–3.5 minutes, p \u0026lt; 0.001). Inter-rater reliability among physicians was fair to moderate (κ = 0.36–0.49).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConclusions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA neurosymbolic multi-agent LLM system demonstrated comprehensive, guideline-concordant COPD and asthma management performance surpassing all specialist physician groups across nearly all clinical domains. The system's ability to integrate domain-specific reasoning across pulmonology, allergology, and emergency medicine knowledge bases, guided simultaneously by the GOLD 2025and GINA 2025 frameworks, suggests meaningful clinical decision support potential, particularly in resource-limited or cross-specialty settings. These findings support further prospective clinical validation.\u003c/p\u003e","manuscriptTitle":"Neurosymbolic Multi-Agent Large Language Model System Versus Specialist Physicians in COPD and Asthma Management: A Comparative Performance Evaluation Using Guideline-Based Clinical Vignettes","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-04-27 16:19:23","doi":"10.21203/rs.3.rs-9262455/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvitedReview","content":"","date":"2026-05-18T10:40:08+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"332806856816894949206163174398845118573","date":"2026-04-20T14:08:47+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"76814994133275670891592876154764828546","date":"2026-04-18T17:59:25+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-18T12:07:23+00:00","index":"","fulltext":""},{"type":"editorInvited","content":"","date":"2026-04-06T14:05:42+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-03-31T04:30:17+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-03-31T04:29:18+00:00","index":"","fulltext":""},{"type":"submitted","content":"Scientific Reports","date":"2026-03-30T04:44:34+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"scientific-reports","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"scirep","sideBox":"Learn more about [Scientific Reports](http://www.nature.com/srep/)","snPcode":"","submissionUrl":"","title":"Scientific Reports","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"Scientific Reports","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8cb25e62-1527-468d-a9f9-2738563b945e","owner":[],"postedDate":"April 27th, 2026","published":true,"recentEditorialEvents":[{"type":"editorInvitedReview","content":"","date":"2026-05-18T10:40:08+00:00","index":62,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":66995496,"name":"Health sciences/Diseases"},{"id":66995497,"name":"Health sciences/Health care"},{"id":66995498,"name":"Health sciences/Medical research"}],"tags":[],"updatedAt":"2026-04-27T16:19:23+00:00","versionOfRecord":[],"versionCreatedAt":"2026-04-27 16:19:23","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9262455","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9262455","identity":"rs-9262455","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}