299 - Can AI Make the Right Call? Large Language Models in Pediatric Dental Trauma Management

Background: Large language models (LLMs) have shown potential as clinical decision assist tools, yet their performance in dental trauma assessment remains inadequately evaluated. This study compared five LLMs in answering dental trauma case scenarios based on International Association of Dental Traumatology (IADT) guidelines.

Methods: Thirty case scenarios including five trauma types (avulsion, intrusion, extrusion, lateral luxation, and root fracture/crown-root fracture) were developed from published case reports. Five LLMs (ChatGPT5, DeepSeek, Gemini, Claude, Grok) received standardized prompts with PDF case files containing patient history and imaging. Responses addressing diagnosis, treatment, and prognosis were scored (0-2 scale). Statistical analysis employed Kruskal-Wallis test and Dunn's post-hoc comparisons with Bonferroni correction.

Results: Models performed similarly in diagnosis (P=.314) and treatment planning (P =.158) but differed significantly in prognosis evaluation (P < .001), with ChatGPT5 superior to all others. Overall, ChatGPT5 scored highest (135/180, 75.0%), closely followed by DeepSeek (134/180, 74.4%), both significantly better than Grok (116/180, 64.4%), Claude (118/180, 65.6%), and Gemini (120/180, 66.7%) (H = 11.24, P =.024). By trauma type, only intrusion injuries showed significant differences (P= .012), with ChatGPT5 reaching 97.2% accuracy versus Claude (66.7%), Grok (72.2%), DeepSeek (83.3%), and Gemini (80.6%). Avulsion, extrusion, lateral luxation, and fractures showed no significant differences between models. With statistical significance level set as P <0.05.

Conclusions: DeepSeek extracts text from images, while multimodal models like Claude, Gemini, and GPT-5 can directly interpret them. Both GPT-5 and DeepSeek performed strongly compared to other models, with GPT-5 showing higher accuracy in complex prognostic reasoning and intrusion injury management.