Explainable Deep Learning Framework For Oral Potentially Malignant Disorders Management

Oral potentially malignant disorders (OPMDs) carry a significant risk of malignant transformation, particularly in cases involving epithelial dysplasia (OED). However, reliable decision-support models for cancer risk assessment and follow-up optimization are lacking, while routine OED evaluation remains invasive and challenging. This study aims to develop and validate a fully automated, end-to-end explainable deep learning framework for non-invasive OED status prediction and time-to-event cancer risk estimation. Using a multi-institutional OPMD cohort (2003–2024) with histopathological confirmation, we developed a multi-modal multi-task fusion network integrating high-resolution clinical images and electronic medical record. Model performance was assessed via C-index, Area Under the Curve (AUC), sensitivity, specificity, accuracy, precision, and F1-score. Calibration, decision curve analysis (DCA), and external validation were conducted to evaluate clinical utility and generalizability. The fusion network achieved AUCs of 0.96 and 0.92 for cancer risk and OED identification respectively, with an overall C-index of 0.80 (95% CI: 0.7789–0.8105). Calibration and DCA confirmed strong clinical applicability, while external validation demonstrated robustness. Interpretability analysis identified key visual and clinical predictors, supporting real-world clinical integration. This explainable multidimensional framework provides accurate predictions for OPMDs malignant progression, demonstrating excellent calibration and discrimination. It shows significant potential as a decision-support tool, enabling optimized biopsy selection and reducing unnecessary procedures. Its applicability in low-resource settings may enhance global oral cancer screening and improve patient outcomes.