Answers are grounded in this paper and cite their sources.
Background: Oral potentially malignant disorders (OPMDs), including leukoplakia, erythroplakia, oral submucous fibrosis, and lichen planus, are established precursors to oral squamous cell carcinoma. Artificial intelligence (AI) models provide a potential to improve predictive accuracy with real-world, multiparametric data. Objective: The objective of this study was to create and validate a machine learning-based predictive model for the prognosis of malignant transformation in OPMD patients based on demographic, clinical, and histopathological factors from real-world data. Materials and Methods: The retrospective multicenter analysis comprised 346 patients with OPMDs and at least 3 years of follow-up. Three machine learning models, supervised and trained, were tested on the dataset: Logistic Regression, Random Forest, and XGBoost. Performance metrics were the area under the ROC curve (AUC), sensitivity, specificity, accuracy, precision, and Brier score. Predictors were chosen by recursive feature elimination and SHapley Additive exPlanations value analysis. Kaplan–Meier and Cox regression analysis were employed for time-to-event analysis. Results: The XGBoost model had the most accurate prediction with AUC = 0.912, sensitivity = 81.6%, specificity = 89.1%, and Brier score = 0.078. Multivariate logistic regression found moderate-to-severe dysplasia (adjusted odds ratio [OR]: 4.92), tobacco consumption (OR: 2.78), and tongue lesions (OR: 2.23) to be important independent predictors. Cox regression also found these as important predictors of earlier malignant transformation. Kaplan–Meier plots showed significantly lower transformation-free survival among patients with high-risk features. Conclusion: AI models, specifically XGBoost, are able to accurately stratify the risk of malignant transformation in OPMDs from real-world clinical and pathological information. Incorporation of such tools into clinical practice could assist with the early detection of high-risk patients and direct personalized surveillance plans.
No takes yet. Share an insight, caveat, or question.
Veeraraghavan et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: