Large language models (LLMs) are increasingly used in clinical and educational settings. However, there is a paucity of data on LLMs’ performance in specialized dental domains. This study assessed the performance of four LLMs, including two general-purpose models, ChatGPT-4o and DeepSeek-R1, and two research-focused models, Consensus and Perplexity, using a validated set of 50 multiple-choice questions in periodontology. Each LLM completed five independent trials encompassing the full question set under both primed and non-primed conditions. A validated answer key served as the benchmark. Performance was analyzed using one-way and two-way analysis of variance and independent-samples t-tests, with additional item-level analyses to identify questions that were consistently difficult. Overall accuracy across all models and trials was 65.0% (95% confidence interval: 63.4-66.6%) with a standard deviation of 3.0. There were no significant differences between models (p = 0.336). Role-specific priming, in which models were instructed to respond as board-certified periodontists, did not improve performance (p = 0.836). At the item level, four questions were never answered correctly, and several others were answered correctly in fewer than 13% of trials. These difficult items generally required detailed procedural knowledge, rare factual recall, or application of classification frameworks. Overall, these findings suggest that current LLMs demonstrate moderate domain knowledge in periodontology but fall short of the reliability required for unsupervised clinical decision support.
Rusa et al. (Thu,) studied this question.