GRASP: A New Feature Selection Method for Medical Prediction

Researchers have developed GRASP, a novel framework for feature selection in medical prediction that combines Shapley value attribution with group L21 regularization. The method first extracts group-level importance scores from a pretrained tree model using SHAP, then applies group L21 regularized logistic regression to enforce structured sparsity, producing compact and non-redundant feature sets. Comparative tests against LASSO, SHAP, and deep learning methods show GRASP achieves comparable or superior predictive accuracy while selecting fewer, less redundant, and more stable features. The approach addresses key limitations of existing methods like LASSO, which often lack robustness and interpretability. The paper is available on arXiv under computer science and machine learning categories.