Comparing Artificial Intelligence (ChatGPT, Gemini, DeepSeek) and Oral Surgeons in Detecting Clinically Relevant Drug–Drug Interactions in Dental Therapy
Subhi Tayeb, Carlo Barausse, Gerardo Pellegrino, Martina Sansavini, Roberto Pistilli, Pietro Felice
Abstract
Patients undergoing oral surgery are frequently polymedicated and preoperative prescriptions (analgesics, corticosteroids, antibiotics) can generate clinically significant drug–drug interactions (DDIs) associated with bleeding risk, serotonin toxicity, cardiovascular instability and other adverse events. This study prospectively evaluated whether large language models (LLMs) can assist in detecting clinically relevant DDIs at the point of care. Five LLMs (ChatGPT-5, DeepSeek-Chat, DeepSeek-Reasoner, Gemini-Flash, and Gemini-Pro) were compared with a panel of experienced oral surgeons in 500 standardized oral-surgery cases constructed from realistic chronic medication profiles and typical postoperative regimens. For each case, all chronic and procedure-related drugs were provided and the task was to identify DDIs and rate their severity using an ordinal Lexicomp-based scale (A–X), with D/X considered “action required”. Primary outcomes were exact agreement with surgeon consensus and ordinal concordance; secondary outcomes included sensitivity for actionable DDIs, specificity, error pattern and response latency. DeepSeek-Chat reached the highest exact agreement with surgeons (50.6%) and showed perfect specificity (100%) but low sensitivity (18%), missing 82% of actionable D/X alerts. ChatGPT-5 showed the highest sensitivity (98.0%) but lower specificity (56.7%) and generated more false-positive warnings. Median response time was 3.6 s for the fastest model versus 225 s for expert review. These findings indicate that current LLMs can deliver rapid, structured DDI screening in oral surgery but exhibit distinct safety trade-offs between missed critical interactions and alert overcalling. They should therefore be considered as decision-support tools rather than substitutes for clinical judgment and their integration should prioritize validated, supervised workflows.