Machine learning-based prediction of severe oral mucositis from head and neck cancer radiation therapy

This study aimed to develop prediction models for severe oral mucositis (OM) (grade ≥ 3) from head and neck cancer (HNC) radiotherapy (RT) using machine learning (ML) techniques and different types of features, including clinical, dose-volume metric, and radiomic features.

Retrospective data of 148 HNC RT patients were collected and randomly divided into 80% training and 20% test datasets. Sixteen combinations of feature selection and classification algorithms were used for modeling with different combinations of feature types. The model performance was evaluated using a confusion matrix and the area under the receiver operating characteristic curve (AUROC). The difference in the error rates obtained from any pair of modeling approaches were evaluated with McNemar's test.

The best-performing model was based on the random forest classifier and minimum redundancy maximum relevance feature selection algorithm, achieving the accuracy of 0.930 (95% confidence interval [CI], 0.926 to 0.933) and AUROC of 0.901 (95% CI, 0.897 to 0.905). For this model, D45% of the oral cavity, subsite, V75Gy of the parotid glands and total dose to the target were found to be the most important features. All modeling approaches achieved the accuracy and AUROC of at least 0.756 and 0.722 and the error rates obtained by all modeling approaches were insignificantly different.

The inclusion of clinical and dose-volumetric features were most promising for ML-based prediction of severe OM, although information obtained prior to dose calculation (clinical or radiomic) may also be used exclusively with insignificantly different error rates.