Fig. 1: Performance evaluation of OMMT-PredNet model for OED identification and cancer risk assessment.

A and B ROC curve for OED [AUC = 0.9219 (95% CI = 0.9088, 0.9349)] and malignant transformation [AUC = 0.9592 (95% CI = 0.9491, 0.9693)] prediction, illustrating the model’s sensitivity and specificity at various thresholds. C and D Calibration plot for multi-tasks, showing the agreement between predicted probabilities and actual outcomes. The plots feature a diagonal line representing perfect calibration, where predicted probabilities match observed frequencies. Points above the line indicate overestimation of risk, while points below suggest underestimation. E and F Decision curve analysis (DCA) for multi-tasks, demonstrates the net benefits of using the model across different probability thresholds, comparing it to the strategies of treating all patients or treating none. DCA revealed that OMMT-PredNet model offers a net benefit compared to both the treat-all and treat-none strategies within clinically relevant thresholds, highlighting its potential utility in clinical decision-making. G The training and validation loss curves. The training loss consistently declined from approximately 1.05 to 0.85 over 100 epochs, indicating effective learning. The validation loss showed a similar trend, with both losses closely aligned, suggesting good generalization and no overfitting. The stability of the loss curves also indicates effective convergence, demonstrating the model’s efficacy and potential for reliable performance. H Kaplan–Meier analysis is performed to visualize patient stratification of low- and high-risk patients for individuals.