From Guidelines to Clicklists: GPT-5-Generated ERAS Checklists Improve Guideline Coverage for Bariatric and...

Area of Science:

• Surgical Pathway Optimization
• Artificial Intelligence in Healthcare
• Clinical Informatics

Background:

• Enhanced Recovery After Surgery (ERAS) pathways are crucial for bariatric and gastrointestinal (GI) cancer surgery outcomes.
• Inconsistent real-world adherence to ERAS pathways necessitates improved implementation tools.
• Digital tools for ERAS face limitations in maintenance and completeness, prompting exploration of AI solutions.

Purpose of the Study:

• To evaluate the coverage, clarity, and potential bias of ERAS checklists generated by large language models (LLMs) compared to traditional methods.
• To assess the feasibility of using AI, specifically GPT-5, for rapidly creating structured ERAS checklists for bariatric and GI cancer surgery.
• To address concerns of
• bundle inflation
• and its impact on ERAS pathway adherence.

Main Methods:

• A cross-sectional observational study (March-June 2025) compared AI-generated (GPT-5) ERAS checklists with guideline-derived traditional checklists.
• Twelve ERAS checklists were generated by AI (6 bariatric; 6 GI cancer) and 12 were curated from ERAS Society guidelines.
• Three blinded raters assessed item coverage, clarity (5-point Likert scale), and bias using a predefined rubric; interrater reliability was evaluated using Cohen's kappa.

Main Results:

• AI-generated checklists exhibited significantly higher guideline-item coverage (97.0% ± 2.1%) and clarity (4.8 ± 0.2) compared to traditional checklists (89.0% ± 3.2% and 4.2 ± 0.3, respectively; p=0.021).
• Excellent interrater agreement (κ=0.92) was observed, with no systematic demographic bias identified.
• Limitations included reduced context-specific tailoring in AI checklists, emphasizing the need for implementability review to prevent overly long checklists.

Conclusions:

• GPT-5-generated ERAS checklists demonstrate superior guideline coverage and clarity for bariatric and GI cancer surgery compared to traditional checklists.
• AI-generated checklists should serve as draft "master lists" requiring local curation to balance core versus conditional elements for optimal implementation.
• Prospective validation studies are essential to confirm real-world effectiveness and address context-specific omissions in AI-generated ERAS pathways.