Banu Çiçek TEZ YAŞAR, Elif Gökçe ERKAN ACAR
Journal of Medicine and Palliative Care - 2026;7(5):956-964
Aims: This study evaluated the performance of three multimodal large language models (LLMs) ChatGPT, Gemini, and Claude in Demirjian-based dental age (DA) estimation from panoramic radiographs compared with specialist dentists. Methods: This retrospective cross-sectional study included 135 panoramic radiographs of children aged 6-13 years. Two specialist dentists confirmed inter-examiner reliability (ICC=0.90) on a pilot subset; the pediatric dentist then assessed all radiographs using the Demirjian method. The radiographs were analyzed once per model using standardized prompts in ChatGPT-5.5 Instant, Gemini 3.5 Flash, and Claude Sonnet 4.6; thus, the reported metrics reflect single-session model performance. Demirjian stages were converted into gender-specific maturity scores and DAs. Performance was assessed using bias, mean absolute error (MAE), and intraclass correlation coefficient (ICC). Statistical analyses included paired-samples t-tests, Wilcoxon, Friedman, Mann-Whitney U, and independent-samples t-tests (p<0.05). Results: Specialists overestimated chronological age (CA) by 1.07+/-1.07 years. Gemini and Claude significantly overestimated age (bias: 2.50+/-1.98 and 2.38+/-1.89 years, respectively), whereas ChatGPT underestimated it (-0.54+/-1.50 years). ChatGPT had the lowest MAE relative to CA (1.26+/-0.98 years), while Gemini showed the highest agreement with specialists (ICC=0.521). Claude and ChatGPT showed poor agreement (ICC=0.439 and 0.390, respectively). No significant gender-related differences were observed in bias or MAE (p>0.05). Conclusion: ChatGPT yielded the lowest MAE relative to CA; Gemini showed the highest specialist agreement, which should not be equated with accuracy. Current multimodal LLMs remain insufficiently accurate and consistent to replace specialist dentists and should be considered adjunctive decision-support tools.