Artificial Intelligence and Guideline-Augmented Prompting in Assessing the Need for Preoperative Cardiology Consultation

OBJECTIVE: With the growing elderly population worldwide, the number of annual surgical procedures has risen substantially, leading to an increase in the demand for preoperative cardiology consultations. In parallel, recent years have witnessed remarkable innovations in cardiology driven by advances in artificial intelligence (AI) and machine learning (ML). In this study, we aimed to evaluate the performance of three widely used AI models: ChatGPT-5, Deepseek-V3, and Gemini 2.0 Pro, in assessing the necessity of cardiology consultation in preoperative patients and to explore the potential contribution of guideline-augmented prompting in this context. METHOD: A council consisting of seven cardiologists and seven anesthesiologists was formed. Each physician evaluated 20 preoperative patient scenarios and provided recommendations on whether a separate cardiology consultation was necessary. For each case, the majority decision of the council was accepted as the reference standard. The same scenarios were presented to the three AI models, and their responses were recorded. Subsequently, the AI models with the highest concordance were integrated into the decision framework using guideline-augmented prompting, and the cases were re-evaluated. RESULTS: Although there was no statistically significant difference, ChatGPT-5 and Gemini 2.0 Pro showed higher concordance than Deepseek-V3 in preoperative consultation decisions (κ = 0.706 and κ = 0.681, respectively; 85% accuracy). Following the integration of guidelines into ChatGPT-5 and Gemini 2.0 Pro, the models were re-evaluated and demonstrated improved performance (κ =0.898, 95% accuracy). CONCLUSION: ChatGPT-5, Deepseek-V3, and Gemini 2.0 Pro demonstrated effectiveness in assessing the necessity of cardiology consultation in preoperatively evaluated patients. Moreover, the integration of guideline-augmented prompting was shown to improve the accuracy and reliability of AI model performance.