Multicenter Study On Noninvasive Ai Model For Opmd Risk Prediction

Objective: To develop a risk prediction model for oral potentially malignant disorders (OPMDs) using noninvasive diagnostic techniques. This study compares multiple machine learning models to determine the most effective approach for assessing malignant transformation risk. Methods: A total of 90 patients with oral mucosal lesions were included. All underwent histopathological examination (gold standard), TCT, DNA ploidy analysis, and gene methylation detection. Clinical data, including smoking, alcohol consumption, and betel quid chewing, were collected. Pathological results classified cases into mild (n=32), moderate (n=35), and severe dysplasia/OSCC (n=23). Ordinal logistic regression, random forest, and XGBoost models were trained, and performance was evaluated using ROC curves and AUC values. Results: The AUC values for ordinal logistic regression, random forest, and XGBoost were 0.72, 0.78, and 0.82, respectively. XGBoost showed the highest accuracy (76%) in predicting severe dysplasia/OSCC. Sensitivity and specificity for XGBoost were 74% and 79%, respectively. Conclusion: The integration of TCT, DNA ploidy analysis, and gene methylation detection provides a feasible noninvasive approach for OPMD risk assessment. While XGBoost performed slightly better than other models, overall predictive accuracy was moderate, suggesting the need for further optimization and validation in larger cohorts.