Machine learning models for predicting recurrence and malignant transformation of oral leukoplakia.

OBJECTIVE: This study aimed to evaluate machine learning models for predicting the recurrence and malignant transformation of oral leukoplakia (OL). METHODS: A total of 123 OL patients with biopsy-confirmed diagnoses were retrospectively enrolled, and their clinicopathologic data were carefully documented. The data underwent preprocessing to ensure uniformity and reliability. Five machine learning models (Logistic Regression, Random Forest, XGBoost, Artificial Neural Network, and Decision Tree) were evaluated using nested cross-validation, and model performance was assessed using accuracy, precision, recall, F1-score, specificity, and area under curve (AUC). SHapley Additive exPlanations (SHAP) values were computed to assess feature importance and interpret model predictions. RESULTS: Logistic regression was the best-performing model for recurrence prediction (AUC: 0.65; accuracy: 0.59; recall: 0.65; F1-score: 0.37), using an optimized threshold of 0.454. For malignant transformation, Artificial Neural Network achieved the best performance: balanced accuracy (0.82); recall (0.70); AUC (0.77). SHAP analysis identified recurrence, dysplasia degree, and treatment type as key predictors. CONCLUSION: Machine learning models showed potential in predicting OL outcomes with clinical data, with Logistic Regression and Artificial Neural Network offering the most balanced performance for recurrence and malignant transformation, respectively. However, variability across folds and limited sensitivity highlight the need for further model refinement and larger datasets before clinical application.