Abstract
Radiation-induced xerostomia remains a common and debilitating side effect in head-and-neck cancer radiotherapy, despite advances in volumetric modulated arc therapy (VMAT). Traditional dose-volume histogram (DVH) metrics capture only part of the variation in toxicity, motivating the use of multimodal imaging biomarkers such as dosiomics and radiomics to characterize dose distribution and tissue response better. In this pilot study, we present an integrated framework combining DVH metrics, 3D dosiomics features, baseline planning CT (pCT) radiomics, and novel longitudinal delta-radiomics derived from daily cone-beam CT-based synthetic CT (sCT) images to predict post-treatment xerostomia severity. In a cohort of ten high-risk oropharyngeal cancer patients treated with VMAT at the Cleveland Clinic, wrapper-based feature selection yielded a compact set of 15 predictors (5 DVH, 3 dosiomics, 4 pCT radiomics, 3 Δ-sCT radiomics). Using cross-validation, four classifiers, including support-vector machine (SVM), regularized logistic regression (GLMnet), Naïve Bayes, and k-nearest neighbors, achieved consistently strong performance for discriminating grade I vs. grade II xerostomia, with AUC of 0.97-1.00, accuracy of 0.90-0.93, uniformly high sensitivity (1.00), specificity of 0.75-0.83, and F1 scores of 0.923-0.945. SVM and GLMnet showed the best overall balance of discrimination and robustness. These results demonstrate the potential of integrating dosiomics with multiphase radiomics, particularly time-resolved delta-radiomics, for individualized xerostomia risk prediction.