COCOH : A Multimodal Deep Learning Framework for Cancer Risk Assessment of Oral Potentially Malignant Disorders

Abstract Oral potentially malignant disorders (OPMD) are mucosal diseases with an increased risk of progression to cancer, although not all cases develop cancer in patients’ lifetimes. Although epithelial dysplasia (ED) grading is the current approach for assessing the risk of malignant transformation (MT), cancer risk prediction is limited by its subjective interpretation and inaccuracy. To address these challenges, this study developed COCOH ( C omprehensive O ral C ancer predictor in O PMD using H istopathology and immunohistochemistry), a multimodal attention-based deep multiple instance learning (MIL) framework that integrates disease information, hematoxylin and eosin (H&E), and Keratin-13 (KRT13) immunohistochemistry whole-slide images (WSIs) to predict MT risk in OPMD. Trained on 2,780 WSIs and over 2.7 million image patches from cases with oral leukoplakia and oral lichenoid disorders, COCOH achieved an AUC of 0.904 (0.843-0.965) and Brier score of 0.089 (0.044-0.134) during testing, with excellent discrimination and calibration maintained at prospective validation (AUC: 0.951 [0.897-1], Brier score: 0.062 [0.002-0.122]). Notably, the deep learning model outperformed two ED grading systems in assessing cancer risk in oral leukoplakia, yielding a significant net reclassification improvement of 20.2%. Decision curve analysis confirmed COCOH ’s clinical utility in identifying high-risk OPMD cases for intervention and close monitoring. These findings demonstrate that COCOH provides accurate and well-calibrated risk predictions for MT in OPMD, and its integration into clinical workflows could enhance prevention strategies and promote earlier detection of oral cancer.