Exploring spatial dose information in the parotid gland for xerostomia prediction and local dose patterns in head and neck cancer radiotherapy

Abstract Purpose . To investigate the relationship between spatial parotid dose and the risk of xerostomia in patients undergoing head-and-neck cancer radiotherapy, using machine learning (ML) methods. Methods . Prior to conducting voxel-based ML analysis of the spatial dose, two steps were taken: (1) The parotid dose was standardized through deformable image registration to a reference patient; (2) Bilateral parotid doses were regrouped into contralateral and ipsilateral portions depending on their proximity to the gross tumor target. Individual dose voxels were input into six commonly used ML models, which were tuned with ten-fold cross validation: random forest (RF), ridge regression (RR), support vector machine (SVM), extra trees (ET), k-nearest neighbor (kNN), and naïve Bayes (NB). Binary endpoints from 240 patients were used for model training and validation: 0 (N = 119) for xerostomia grades 0 or 1, and 1 (N = 121) for grades 2 or higher. Model performance was evaluated using multiple metrics, including accuracy, F 1 score, areas under the receiver operating characteristics curves (auROC), and area under the precision–recall curves (auPRC). Dose voxel importance was assessed to identify local dose patterns associated with xerostomia risk. Results . Four models, including RF, SVM, ET, and NB, yielded average auROCs and auPRCs greater than 0.60 from ten-fold cross-validation on the training data, except for a lower auROC from NB. The first three models, along with kNN, demonstrated higher accuracy and F 1 scores. A bootstrapping analysis confirmed test uncertainty. Voxel importance analysis from kNN indicated that the posterior portion of the ipsilateral gland was more predictive of xerostomia, but no clear patterns were identified from the other models. Conclusion . Voxel doses as predictors of xerostomia were confirmed with some ML classifiers, but no clear regional patterns could be established among these classifiers, except kNN. Further research with a larger patient dataset is needed to identify conclusive patterns.