Machine learning to predict osteoradionecrosis in head and neck cancer
Related Experiment Video
Updated: Nov 6, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
Vitoria Tavares Castro1, Luan Matheus Trindade Dalmazo2, Paula Elaine Diniz Dos Reis3
1University of Brasilia, Laboratory of Oral Histopathology, School of Health Sciences, Brasília, Brazil.
Objective:
Osteoradionecrosis is one of the most severe late adverse effects of head and neck radiotherapy. Predictive models may help identify patients at increased risk and support personalized treatment. This exploratory study aimed to develop a machine learning-based prediction model for osteoradionecrosis in a real-world cohort, following the TRIPOD+AI statement.
Methods:
Clinical variables and mandibular dose-volume histogram parameters were collected from head and neck cancer patients treated with radiotherapy. Forward Stepwise Logistic Regression was used for predictor selection. Six machine learning algorithms, Logistic Regression, Random Forest, Support Vector Machine, CatBoost, XGBoost, and TabPFN, were evaluated under class-weighted baseline modelling, threshold optimization, and data augmentation with repeated stratified 5-fold cross-validation. Performance was primarily assessed using sensitivity. Pairwise comparisons used the Nadeau-Bengio corrected paired t-test with Holm adjustment.
Results:
Ninety-seven patients were included, of whom 8 developed osteoradionecrosis (8.25%). Of the 80 available variables, V60Gy and tumor location in the oral cavity were selected as predictors. Without augmentation, Logistic Regression achieved the highest sensitivity (0.74 [95% CI: 0.40-1.00]), while several algorithms showed poor sensitivity. With data augmentation, sensitivity ranged from 0.72 to 0.77 across models, with TabPFN, CatBoost, Logistic Regression and Support Vector Machine reaching the highest sensitivity. Logistic Regression achieved the highest AUC under both strategies (0.87). However, F1 scores remained modest, reaching a maximum of 0.48.
Conclusion:
Machine learning models demonstrated exploratory potential for ORN risk assessment within an internal validation framework. Class-imbalance strategies were associated with greater minority-class detection in this cohort, but their generalizability remains uncertain. Larger multicenter cohorts and external validation are needed before clinical implementation.