Trajectory-aware risk stratification of oral lichen planus using a multimodal large language model: a longitudinal diagnostic accuracy study.

To evaluate the performance of a multimodal large language model (LLM) for longitudinal trajectory classification and risk stratification of oral lichen planus (OLP), compared with expert panel consensus. This retrospective diagnostic accuracy study included 300 patients with histopathologically confirmed OLP and at least 24 months of follow-up. Multimodal longitudinal case profiles (serial clinical records, intraoral photographs, and histopathology reports) were independently assessed by (ChatGPT, OpenAI) and an expert panel (the reference standard), both blinded to the results. The primary outcome was sensitivity for the detecting expert-defined high-risk cases (one-versus-rest). Secondary outcomes included overall trajectory classification and three-level risk stratification. Expert consensus classified 156 cases (52.0%) as stable benign, 92 (30.7%) as inflammatory progression, and 52 (17.3%) as suspicious malignant evolution. Risk stratification was 73 low (24.3%), 161 moderate (53.7%), and 66 high (22.0%). For high-risk detection, sensitivity was 78.8% (95% CI 67.2-87.5), and specificity was 99.6% (95% CI 97.6-100.0). Overall trajectory classification accuracy was 94.7%. Three-level risk stratification accuracy was 76.3%, with most errors representing downward shifts (98.6%). The multimodal LLM showed high concordance with expert consensus for longitudinal OLP surveillance, with high accuracy of trajectory classification and high specificity for high-risk identification. These findings suggest that multimodal LLMs may support trajectory-based assessment through integration of longitudinal clinical information, although prospective external validation remains necessary.