Enhancing malignant transformation predictions in oral potentially malignant disorders: A novel machine learning
Jing Wen Li1, Meng Jing Zhang2, Ya Fang Zhou2
1Division of Oral & Maxillofacial Surgery, Faculty of Dentistry, The University of Hong Kong, Hong Kong SAR, China.
Iscience
|March 19, 2025
Summary
Predicting oral potentially malignant disorders (OPMDs) transformation risk is crucial. A novel Self Attention Artificial Neural Network (SANN) model significantly outperformed traditional methods, improving early detection for high-risk patients.
Area of Science:
- Oral Medicine
- Oncology
- Artificial Intelligence in Healthcare
Background:
- Oral potentially malignant disorders (OPMDs) carry a risk of malignant transformation.
- Accurate risk prediction is essential for timely intervention and improved patient outcomes.
- Current statistical methods may have limitations in capturing complex predictive patterns.
Purpose of the Study:
- To compare the predictive performance of machine learning (ML) models against traditional statistical methods for OPMD malignant transformation risk.
- To develop and validate a novel Self Attention Artificial Neural Network (SANN) model for enhanced risk prediction.
- To assess the generalizability and robustness of ML models in diverse clinical settings.
Main Methods:
- Retrospective analysis of 1,094 patients with OPMDs from three institutions (2004-2023).
- Development and validation of a Self Attention Artificial Neural Network (SANN) model.
- Comparison with other ML algorithms (ANN, RF, DeepSurv) and a Cox proportional hazards (Cox-PH) nomogram.
- Performance evaluation using AUC, sensitivity, specificity, accuracy, precision, ROC curves, calibration curves, and decision curve analysis.
Main Results:
- The SANN model achieved a superior Area Under the Curve (AUC) of 0.9877.
- SANN demonstrated high sensitivity, specificity, accuracy, and precision, all exceeding 0.96.
- The Cox-PH nomogram yielded significantly lower AUCs, ranging from 0.880 to 0.902.
- Comprehensive evaluations confirmed SANN's superior predictive efficacy, robustness, and generalizability.
Conclusions:
- Customized ML models, particularly SANN, show significant potential for improving the prediction of malignant transformation in OPMDs.
- SANN outperforms conventional statistical methods like Cox-PH nomograms in risk stratification.
- These findings support the integration of advanced ML for enhanced early identification and management of high-risk OPMD patients.
