Leveraging Compressed Sensing and Radiomics for Robust Feature Selection for Outcome Prediction in Personalized Ultra‐Fractionated Stereotactic Adaptive Radiotherapy

Personalized ultra‐fractionated stereotactic adaptive is an innovative radiation treatment paradigm. To fully harness its benefits, transitioning decision‐making and plan adaptation from empirical judgment to a data‐driven approach is essential. This study presents a novel compressed sensing (CS)‐based feature selection method, offering distinct advantages over conventional approaches such as the Least Absolute Shrinkage and Selection Operator (Lasso). Two feature selection models are developed—binary and Gaussian random projections—and applied them to a brain metastasis dataset. Using the top five features identified, support vector machine models are trained to classify whether lesions achieved ≥20% volume reduction at a 3‐month follow‐up. Integrating residual error with frequency‐based selection further enhances performance over weight coefficient‐based criteria. For 5‐feature sets, the CS‐Binary model outperforms Lasso across multiple metrics: AUC (0.937 vs 0.890), balanced accuracy (87.3% vs 83.0%), F1 score (79.4% vs 75.9%), Kappa coefficient (75.8% vs 69.0%), and Matthews correlation coefficient (78.0% vs 72.1%). The CS‐based framework shows great potential in streamlining feature selection and improving predictive accuracy, particularly beneficial in two scenarios: 1) early phase clinical trials with small datasets where traditional radiomics methods are prone to overfitting; and 2) emphasizing the selection of the most prognostically relevant features to help improve interpretability.