Machine learning-based development of a prediction model for osteoradionecrosis in head and neck cancer

This project aims to develop and validate machine learning models for predicting Osteoradionecrosis (ORN) in patients with head and neck cancer treated with 3D Conformal Radiotherapy (3D CRT). ORN is a severe clinical complication, and its low prevalence often results in imbalanced datasets that challenge standard predictive modeling. The primary objectives are to evaluate the performance of six machine learning algorithms—Logistic Regression, Random Forest, Support Vector Machine, CatBoost, XGBoost, and TabPFN—within a small data clinical context, and to compare strategies for handling class imbalance, including probability threshold optimization via Youden’s Index and data augmentation using SMOTE-NC. This study will follow the TRIPOD + AI guidelines, and will use a retrospective cohort from the University Hospital of Brasília (HUB), integrating clinical records and EQD2-corrected dosimetric data. Validation is performed through repeated 5-fold cross-validation to ensure statistical robustness, model reliability, and the identification of the most stable clinical and dosimetric predictors of ORN.