Employing multimodal large language models for the evaluation of palatal wound healing

Visual assessment of wound healing is essential in clinical practice but can show inter-observer variability. This study evaluated whether publicly accessible multimodal large language models (LLMs) could provide reliable wound assessment with minimal calibration, comparing their performance to periodontology specialists. Seventy palatal wound images from a rat experimental model were evaluated by 24 periodontology specialists and two multimodal LLMs, GPT-5 (OpenAI; accessed via the ChatGPT web interface) and Gemini 2.5 Pro (Google; accessed via the Gemini web interface), using a 1–10 scoring scale. Both LLMs were exposed to six reference images with detailed scoring criteria as calibration input. Each model completed five independent evaluation rounds. Agreement with the specialist mean score and within-model reliability were assessed using intraclass correlation coefficients (ICC), Pearson and Spearman correlations, and Bland–Altman analysis. Inter-specialist reliability was excellent (ICC(2,24) = 0.985). Gemini 2.5 Pro demonstrated good agreement with specialists (ICC = 0.810; 95% CI: 0.711–0.877; r = 0.81), a mean bias of + 0.13 points, and a mean absolute error of 1.10 points. Its repeatability across single scoring runs was moderate (ICC(2,1) = 0.699), while the reliability of the five-run mean was excellent (ICC(2,5) = 0.921). In contrast, GPT-5 showed poor agreement (ICC = 0.296; 95% CI: 0.068–0.494; r = 0.52), poor reliability at both levels (ICC(2,1) = 0.071; ICC(2,5) = 0.277), and a mean absolute error of 1.53 points. After brief calibration with six reference images, Gemini 2.5 Pro achieved good agreement with the mean specialist score and high reliability of the five-run mean score, although the repeatability of a single scoring run was only moderate. In contrast, GPT-5 showed poor agreement and poor reliability at both levels. Performance in rubric-based visual wound scoring was therefore strongly model-dependent, and results obtained with one multimodal LLM should not be generalized to others.