Clinical Safety Thresholds and Decision Risks of Search-Augmented Large Language Models in Acute Pancreatitis: A Guideline-Anchored Evaluation for Surgical Practice
Burak AltunpakBackground
Patients and clinicians increasingly consult search-augmented Large Language Models (LLMs) for point-of-care decision support. However, their safety thresholds against newly revised guidelines remain unverified. This study established a safety baseline evaluating the clinical accuracy, completeness, quality, and readability of ChatGPT-5.3 and Gemini 3.0 Flash anchored to the 2025 IAP/APA acute pancreatitis guidelines.
Methods
Twenty-seven consensus-derived clinical scenarios spanning diagnosis, management, intervention, and discharge were queried using zero-shot prompts. Three blinded general surgeons evaluated outputs using 5-point Likert scales for accuracy, completeness, and Global Quality Score (GQS), alongside binary error metrics. Readability was quantified via FRES, FKGL, and SMOG indices. Inter-rater agreement was assessed using Intraclass Correlation Coefficients (ICC).
Results
Gemini demonstrated significantly higher completeness than ChatGPT (median [IQR]: 4.33 [4.33-4.67] vs 4.33 [4.00-4.33]; corrected
Conclusion
Both LLMs display strong baseline fidelity to the 2025 IAP/APA guidelines. However, persistent critical errors and high academic reading levels preclude autonomous deployment. Establishing this safety floor is an indispensable prerequisite for human-in-the-loop usability trials, positioning LLMs strictly as supervised clinical co-pilots.