Training ChatGPT for surgical decisions: Bariatric surgery analysis using algorithms and evidence.

BACKGROUND

Selecting the most appropriate bariatric surgery (BS) technique is a complex, individualized process. Artificial intelligence (AI) tools like ChatGPT may assist, but their clinical utility is unclear. This study evaluates whether ChatGPT's recommendations for BS improve after exposure to scientific literature and how they align with real-world clinical decisions.

METHODS

A retrospective single-center study included 283 patients who underwent primary BS between 2023 and 2025. No exclusion criteria were applied. Clinical variables (age, sex, BMI, comorbidities, and preoperative data) were collected. ChatGPT was asked to recommend the most suitable BS technique for each patient profile, first without context and then after being exposed to 412 open-access scientific articles. Recommendations were compared with actual clinical decisions using percentage agreement and Cohen's Kappa.

RESULTS

Initially, ChatGPT favored sleeve gastrectomy (SG, 56.8 %), followed by Roux-en-Y gastric bypass (RYGB, 26.8 %) and one-anastomosis gastric bypass (OAGB, 16.4 %); SADI-S was never suggested. Concordance with clinical practice was 20.0 % (Kappa = 0.003; p = 0.96). After training, SG recommendations decreased (35.7 %), RYGB increased (30.3 %), SADI-S emerged (17.1 %), and dual options appeared in 4 %. Concordance improved modestly to 25.8 % (Kappa = 0.068; p = 0.29), with a significant shift in global distribution (p < 0.00001).

CONCLUSIONS

ChatGPT adapts its recommendations after contextual training, but concordance with clinical judgment remains low. While potentially useful as an educational tool, ChatGPT is not yet reliable for autonomous surgical decision-making.