The Performance of Large Language Models on Antibiotic Prophylaxis for Endodontic Treatments

This study compared the accuracy and repeatability of responses generated by three advanced language models (ChatGPT-4, Gemini, and DeepSeek) in determining the necessity and appropriate regimen of antibiotic prophylaxis in endodontic treatment. 15 questions on the basis of guidance published by the European Society of Endodontics in 2018 were developed and presented to each model by three different users, three times per day over 10 consecutive days. Accuracy was evaluated for individual questions and overall performance, and repeatability across days and times was statistically analysed. All models consistently provided correct responses for patients with penicillin allergy and those with a history of infective endocarditis. Significant differences were observed in responses regarding joint prosthesis, high-dose radiation to the jaws, bypass surgery history, and controlled diabetes, although overall accuracy did not differ significantly among models. These findings highlight the potential clinical value of the models, but their responses must be verified before clinical use.