Comparative Evaluation of Feature Selection Algorithms for Predicting Radiation-Induced Complications in Head and Neck Cancer Using Multimodal CT and MRI Radiomics

Head and neck cancer (HNC) presents a significant challenge due to its anatomic complexity and treatment-related complications such as sticky saliva and xerostomia, which critically impair patients’ quality of life. Recent advancements in imaging technologies, using the concept of radiomics, offer promising approaches for predicting these complications and ultimately reducing their incidence. In this regard, robust feature selection is critical in radiomics-based predictive modeling due to the high dimensionality, redundancy, and instability of extracted imaging biomarkers. This study aims to evaluate nine distinct feature selection algorithms—Boruta, Kruskal, Recursive Maximum Relevance minimum redundancy (R-MRmr), Relief, Recursive Feature Elimination (RFE), Forward Selection (FS), Backward Elimination (BE), Least Absolute Shrinkage and Selection Operator (LASSO), and Principal Component Analysis (PCA)—as a feature-transformation comparator for predicting sticky saliva and xerostomia in HNC patients using clinical, dosimetric, MRI and CT radiomics data. The study involved 250 HNC patients who underwent radiotherapy from three different institutions. From MRI and CT images, a total of 321 radiomic features (107 from each CT, T1, and T2-weighted MRI) along with 20 clinical and dosimetric features were extracted. The predictive robustness of feature selection methods was further evaluated across five machine-learning classifiers, including K-nearest neighbors (KNN), support vector machine (SVM), random forest (RF), logistic regression (LR), and eXtreme Gradient Boosting (XGBoost), were initially benchmarked using a common feature selection procedure, after which SVM was selected as the reference classifier for systematic comparison of the feature-reduction approaches. Model performance was assessed using metrics objectives such as Area Under the Curve (AUC), accuracy, sensitivity, and specificity. Among the evaluated feature selection approaches, R-MRmr and Relief generally yielded better discrimination, although performance varied according to toxicity endpoint and imaging modality. The highest AUCs were 0.85 for early sticky saliva (R-MRmr, T1-weighted MRI), 0.96 for early xerostomia (R-MRmr, T1-weighted MRI), 0.65 for late sticky saliva (Relief, combined CT+T1+T2), and 0.88 for late xerostomia (Relief, T1-weighted MRI). T1-weighted MRI frequently provided better individual-modality performance, whereas combining all three imaging modalities did not consistently improve discrimination. These findings demonstrate that the choice of feature selection strategy and imaging modality can materially influence radiomics-based toxicity prediction.