GRASP: group-Shapley feature selection for patients

TL;DR

GRASP presents an interpretable feature-selection framework for medical prediction by coupling SHAP-based attribution with group-$L_{21}$ regularization in a logistic regression objective, optimized via proximal-gradient methods. It assigns group-aware penalties derived from SHAP-derived importances, enabling stable, non-redundant feature sets while preserving predictive performance. Across NHANES and UK Biobank mortality data, GRASP achieves compact feature sets with high stability, low redundancy, and calibration that aligns with clinical risk thresholds, outperforming or matching existing selectors. The approach enhances interpretability and clinical relevance of predictive models, facilitating reliable risk stratification in real-world healthcare applications.

Abstract

Feature selection remains a major challenge in medical prediction, where existing approaches such as LASSO often lack robustness and interpretability. We introduce GRASP, a novel framework that couples Shapley value driven attribution with group $L_{21}$ regularization to extract compact and non-redundant feature sets. GRASP first distills group level importance scores from a pretrained tree model via SHAP, then enforces structured sparsity through group $L_{21}$ regularized logistic regression, yielding stable and interpretable selections. Extensive comparisons with LASSO, SHAP, and deep learning based methods show that GRASP consistently delivers comparable or superior predictive accuracy, while identifying fewer, less redundant, and more stable features.

Figures (4)

• Figure 1: Workflow of GRASP algorithm.The procedure begins with feature importance calculation by combining input feature groups with Shapley values. These importance scores are then incorporated into a loss function that integrates $L_{21}$ loss and group-wise $L_{21}$ regularization. The final optimization is performed using a proximal-gradient algorithm.
• Figure 2: Main effect plots of Lactate dehydrogenase (LDH) using overlapping feature sets from GRASP, LASSO, SHAP and AFS. Red curves depict LOWESS curves and blue dots show Shapley values. Histograms indicate LDH distribution, and dashed lines mark estimated thresholds across methods.
• Figure 3: Calibration curve comparing predicted probabilities and observed risks for models based on feature sets from GRASP, LASSO, SHAP, and AFS. All results are calibrated using the Platt Scaling method. The dashed line indicates perfect calibration, with shaded areas representing 95% confidence intervals.
• Figure 4: Comparison of different feature selection methods using Kaplan-Meier curves in UKB dataset. Confidence bands represent the 95% confidence interval. The global log-rank test and test statistic are reported in each Kaplan–Meier plot.