Mehmet Gümüş KANMAZ, Burcu KANMAZ
Journal of Basic and Clinical Health Sciences - 2026;10(3):413-423
Purpose: This study aimed to evaluate the photograph-based visual classification performance of contemporary multimodal large language models for gingival recession defects according to Miller's classification. Materials and Methods: A dataset of 100 standardized intraoral photographs showing gingival recession was used. Five periodontists independently classified each photograph according to Miller's classification using only the visual information available in the image. In cases of disagreement, a panel discussion was performed to establish a final periodontist visual consensus classification, which served as the reference comparator. The photographs were then submitted once to ChatGPT 5, Gemini 2.5 Pro, Grok 3, and Copilot GPT-5. Model outputs were compared with the periodontist panel consensus classification using overall and balanced accuracy, class-wise precision, sensitivity/recall, F1-score, macro-F1, weighted-F1, Cohen's kappa, and weighted kappa. Results: The periodontist panel showed almost perfect inter-rater agreement (Fleiss' kappa = 0.885; asymptotic 95% confidence interval [CI]: 0.827-0.934). Overall classification accuracy differed significantly among the evaluated models (p < 0.001). Gemini 2.5 Pro showed the highest accuracy, correctly classifying 49% of photographs, followed by ChatGPT 5 with 32%, Grok 3 with 26%, and Copilot GPT-5 with 24%. Conclusion: Under standardized single-prompt and photograph-only conditions, the evaluated multimodal large language models showed limited and inconsistent agreement with the periodontist panel visual consensus classification.
