Time-varying risk prediction for oral potentially malignant disorders using deep learning

Patients with oral potentially malignant disorders require repeated risk assessment after diagnosis because malignant transformation risk varies across lesions and over time. Existing tools rarely provide individualized, interpretable, time-resolved estimates to guide surveillance. A retrospective multicentre study was conducted with clinicopathologically confirmed oral leukoplakia or oral lichenoid conditions. Hong Kong cohorts from 2003 to 2024 were used for model development and internal testing, and independent United Kingdom and mainland China cohorts for external validation. OCRiskNet integrates demographic, exposure, clinical, treatment and histopathological data in a fixed deep survival pipeline combining probabilistic feature enrichment with an intrinsically interpretable free-knot Kolmogorov-Arnold network. Performance was evaluated using concordance index, area under the curve, Brier score, calibration and decision curve analysis. Interpretability was assessed using symbolic function extraction and feature attribution. Here we show that OCRiskNet provides strong internal discrimination for oral leukoplakia and oral lichenoid conditions, with concordance indices of 0.9209 and 0.9207. External validation supported model generalizability across geographically distinct cohorts. Interpretation analyses identified clinically meaningful contributors to risk. OCRiskNet also demonstrated favorable calibration and clinical utility, with clear stratification of risk groups over time. A web-based interface was further implemented to support practical use of individualized risk estimation in longitudinal management. OCRiskNet provided individualized, time-resolved estimates of malignant transformation risk in OPMDs and demonstrated consistent performance across multicentre cohorts. This framework may support more transparent and risk-adapted surveillance planning in heterogeneous clinical settings, with potential public-health value by enabling risk-adapted surveillance of high-risk patients and more efficient allocation of clinical follow-up resources. Oral potentially malignant disorders (OPMDs) represent an important window for oral cancer prevention. These conditions may persist for years, with malignant transformation risk changing across the disease course, making long-term risk-adapted surveillance essential. Specialist oral medicine services remain limited in many regions, and general dental practitioners may face challenges in determining which patients require closer monitoring. We developed OCRiskNet using longitudinal real-world clinical and histopathological data from multicentre cohorts with extended follow-up. The model estimates how an individual patient’s future cancer risk changes over time using information routinely collected during care. This low-burden approach may help resource-limited settings prioritise surveillance for high-risk patients, use specialist services more efficiently, and strengthen oral cancer prevention.