Comparative Performance of GPT-5 and Gemini Models in Decision Support for Bariatric Surgery: A Simulation-Based Study


Çalışkan Y. K., BAŞAK F.

Surgical Laparoscopy, Endoscopy and Percutaneous Techniques, cilt.36, sa.5, 2026 (SCI-Expanded, Scopus)

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 36 Sayı: 5
  • Basım Tarihi: 2026
  • Doi Numarası: 10.1097/sle.0000000000001478
  • Dergi Adı: Surgical Laparoscopy, Endoscopy and Percutaneous Techniques
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, CINAHL, EMBASE, MEDLINE
  • Anahtar Kelimeler: artificial intelligence, bariatric surgery, clinical decision support, Gemini, GPT-5, large language models, patient safety, simulation
  • Sağlık Bilimleri Üniversitesi Adresli: Evet

Özet

Background: – Artificial intelligence (AI) models are increasingly proposed for surgical decision support, yet their reliability in high-risk domains such as bariatric surgery remains uncertain. Prior bariatric AI studies have mainly evaluated patient education, isolated question-answering, or predictive modeling rather than paired safety-focused decision simulation across the perioperative pathway. Materials and Methods: – This simulation-based paired-comparison study evaluated GPT-5 and Gemini using 60 expert-validated bariatric surgery vignettes spanning preoperative, intraoperative, and postoperative decision-making. Both models were queried under standardized low-temperature conditions. Three bariatric surgeons independently scored outputs for accuracy, safety, guideline adherence, and reasoning quality. Reporting was aligned with STROBE, with AI diagnostic-evaluation elements informed by STARD-AI and analytic transparency principles informed by TRIPOD+AI. Results: – GPT-5 outperformed Gemini across all domains. GPT-5 achieved higher accuracy (87.3% vs. 72.4%), stronger guideline adherence (78% vs. 46%), and fewer unsafe recommendations (2.1% vs. 8.4%). Composite scores were significantly higher for GPT-5 (6.92±0.61) than for Gemini (5.43±0.88; P<0.001). Gemini was 4 times more likely to produce an unsafe action than GPT-5 (risk ratio, 4.0; 95% CI, 2.2-7.3). The largest accuracy gap occurred in intraoperative scenarios. Expert concordance was substantial for GPT-5 (κ=0.71) and moderate for Gemini (κ=0.54). Conclusion: – GPT-5 demonstrated superior accuracy, guideline alignment, and safety compared with Gemini in bariatric surgery decision simulations. Although these findings support the potential of advanced LLMs as supervised decision-support tools, unsafe outputs persisted, reinforcing the need for model version transparency, prospective validation, and clinician oversight before clinical deployment.