Artificial Intelligence-Based Evaluation of Oral Potentially Malignant Disorders Using Photographic and Clinical Data: A Retrospective Exploratory Diagnostic Accuracy Study

Abstract

Objective: The objective of this study was to evaluate the diagnostic performance of GPT-5.2 Thinking in identifying oral potentially malignant disorders within a patient sample that also included oral carcinomas and other oral cavity lesions, using intraoral photographs and associated clinical data. Methods: This retrospective exploratory diagnostic accuracy study included 103 oral lesions: 69 oral potentially malignant disorders, 25 oral carcinomas, and 9 other oral pathologies. Each case was assessed in three independent sessions, and the majority diagnosis was compared with the final clinical and/or histopathological reference diagnosis. Results: For oral potentially malignant disorders, sensitivity was 92.75%, specificity was 88.24%, accuracy was 91.26%, positive predictive value was 94.12%, negative predictive value was 85.71%, and Cohen’s kappa was 0.804. Sensitivity was 81.82% for homogeneous leukoplakia, 78.57% for non-homogeneous leukoplakia, and 86.36% for oral lichen planus. For oral carcinomas, sensitivity was 84.00%, specificity was 97.44%, accuracy was 94.17%, positive predictive value was 91.30%, negative predictive value was 95.00%, and Cohen’s kappa was 0.837. Conclusions: GPT-5.2 Thinking showed promising diagnostic performance in this exploratory case-enriched dataset, but predictive values require validation in representative clinical populations.