🤖 machine learning

GRASP: group-Shapley feature selection for patients

The work presents GRASP, a novel feature selection framework that combines Shapley value assignment with group- $L_{21}$ regularization to achieve stable, interpretable, and non-redundant feature sets for medical predictions, thereby outperforming existing methods such as LASSO in both accuracy and feature stability.

Original authors: Yuheng Luo, Shuyan Li, Zhong Cao

Published 2026-05-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yuheng Luo, Shuyan Li, Zhong Cao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a doctor tasked with predicting which patients are at risk for a serious health event. You have a massive filing cabinet containing thousands of notes for each patient: their age, blood pressure, diet, genetic markers, and thousands of other details.

The problem is that most of these notes are either useless, repetitive, or so confusingly similar to one another that it becomes difficult to find the real warning signs. Conventional methods for sifting through this chaos are like trying to find a needle in a haystack by simply grabbing handfuls of hay; they may work, but they often grab too much useless stuff, miss the needle, or find a different needle each time you try.

This article introduces GRASP, a new, smarter method for searching through this medical filing cabinet.

The Problem with Old Methods

Imagine old feature selection methods (like LASSO) as a strict editor who deletes words based on a simple rule: "If a word does not appear often enough, delete it."

The Flaw: This editor might delete two words that mean the same thing (redundancy) or keep a word that sounds important but is actually not useful. Worse still: if you ask this editor to perform the task twice, he might make a slightly different word selection each time, rendering the story unstable.

The GRASP Solution: A Detective Team in Two Steps

GRASP acts like a detective team in two steps, combining intuition with strict discipline.

Step 1: The Intuition (SHAP Values)

First, GRASP uses a tool called SHAP (Shapley Additive exPlanations). Imagine SHAP as an extremely wise consultant who has already read the entire filing cabinet and knows exactly which notes are most important for predicting the outcome.

Instead of guessing, SHAP assigns a "score" of importance to each individual note.
GRASP groups similar notes together (such as all "blood test" notes). It asks the consultant: "Which groups of notes actually drive the prediction?"

Step 2: The Discipline (Group L21 Regularization)

Once the consultant has provided the scores, GRASP brings in a strict trainer (the Group L21 Regularization).

This trainer says: "We only want the most important groups. If a group of notes is not critical, we delete the entire group. If a group is critical, we keep all notes within it, but we ensure we do not keep too many redundant ones."
This ensures that the final list is compact (short and concise), stable (you get the same list every time you run it), and interpretable (easy for a human to understand).

The Analogy: The Ultimate Travel Packing List

Imagine you are packing for a trip but have 1,000 items to choose from.

Old methods might give you a suitcase full of 50 items, including three different types of socks and two umbrellas that do exactly the same thing. It works, but it is heavy and messy.
GRASP considers your trip, asks an expert what you actually need, and then packs a tiny, perfect bag with only 23 essential items. It ensures you have the right socks and the right umbrella, but no duplicates.

What the Article Found

The authors tested GRASP against old methods using real data from two massive medical databases (NHANES and UK Biobank). Here is what they discovered:

Less is better: GRASP selected the fewest number of features (on average only 23) compared to the others (which selected 40 to 60).
Less clutter: The features selected by GRASP did not repeat each other. The old methods were full of "redundant" features (such as three different ways of measuring the same thing), which confused the models.
Stability: If you ran the test 1,000 times, GRASP selected almost the same 23 features every time. The other methods constantly changed their minds.
Just as precise: Despite using fewer features, GRASP predicted patient outcomes just as well as the methods that used hundreds of features.
Real-world relevance: When examining a specific medical marker called "Lactate Dehydrogenase" (LDH), GRASP identified a danger threshold that aligned much better with real clinical guidelines than the other methods. It found the "tipping point" where the risk actually changes, whereas the others found a slightly different number.

The Conclusion

GRASP is a new tool that helps doctors and researchers cut through the noise of massive medical data. By combining the "intuition" of AI (SHAP) with a "disciplined" mathematical filter, it creates a short, stable, and easy-to-understand list of the most important health indicators without the confusion of redundant data.

1. Problem Statement

In medical prediction, high-dimensional phenotypic data (from electronic health records, imaging, wearables) pose significant challenges for knowledge discovery due to noise and dimensionality. Existing feature selection methods suffer from several critical limitations:

Instability and Redundancy: Methods like LASSO frequently generate unstable feature sets across different data splits and cannot effectively handle multicollinearity.
Lack of Interpretability: While downstream tools like SHAP offer insights, they are often decoupled from the selection process, making it difficult to align selected features with specific disease mechanisms.
Clinical Relevance: Many approaches prioritize prediction accuracy over clinical interpretability and stability, which are crucial for medical acceptance.

The authors propose GRASP (GRoup-shApley feature Selection for Patients) to bridge these gaps by unifying prediction performance, stability, and clinical interpretability.

2. Methodology

GRASP is a novel framework that couples Shapley-value-driven attribution with Group L2,1 regularization within a penalized logistic regression framework. The workflow consists of three main phases:

A. Calculation of Feature Importance (SHAP Integration)

Pre-training: An XGBoost classifier is trained on the training fold.
SHAP Calculation: SHAP values are computed on a held-out validation fold to determine the contribution of each feature.
Group Aggregation: Features are partitioned into $G$ disjoint groups. The importance of a group ( $s_g$ ) is calculated as the mean of the absolute SHAP values of its member features.
Weight Transformation: Group importance values are converted into penalty weights ( $\omega_g$ ) using an exponential decay function:
$\tilde{\omega}_g = \exp(-s_g/\tau_0) + \epsilon$
This ensures that groups with higher SHAP values (higher importance) receive lower penalty weights and are thus more likely to be retained.

B. Construction of the Loss Function

GRASP uses logistic regression with Group L2,1 regularization. The objective function $J(\beta)$ combines the cross-entropy loss with a structured sparsity penalty:
$J(\beta) = L(\beta) + \lambda \sum_{g=1}^{G} \omega_g \|\beta_g\|_2$

$L(\beta)$ : Standard cross-entropy loss for binary classification.
$\|\beta_g\|_2$ : The Euclidean norm of the coefficient vector for group $g$ .
Effect: This penalty enforces sparsity at the group level (entire groups are either selected or discarded), while allowing dense coefficients within selected groups. This addresses multicollinearity by keeping correlated features together.

C. Optimization

The objective is optimized using a Proximal Gradient Algorithm employing an Armijo backtracking line search:

Gradient Step: Updates coefficients based on the gradient of the loss function.
Proximal Step: Applies the Group Soft-Thresholding operator to enforce sparsity.
Adaptive Step Size: The step size is dynamically adjusted to ensure sufficient reduction of the objective function.
Selection: Groups with non-zero norms ( $\|\beta_g\|_2 > 0$ ) are retained; all other features are discarded.

3. Main Contributions

Unified Framework: For the first time, SHAP-based feature attribution is directly coupled with regularization, so the optimization algorithm inherently preserves clinically interpretable features.
Group-Based Selection Mechanism: Introduction of a strategy where correlated features are jointly included in or excluded from the model, significantly improving stability and generalization.
Comprehensive Evaluation: Beyond simple accuracy metrics, redundancy (VIF, correlation), stability (Jaccard Index, Adjusted Stability Measure), and clinical alignment (threshold analysis) are evaluated.

4. Experimental Results

The study was evaluated on two large-scale datasets: NHANES (development) and UK Biobank (external validation). GRASP was compared with LASSO, standard SHAP selection, and the AFS algorithm.

Feature Set Efficiency: GRASP selected the smallest feature set (an average of 23 features) compared to LASSO (44), SHAP (43), and AFS (59).
Stability & Redundancy:
- Stability: GRASP achieved the highest Adjusted Stability Measure (0.593), significantly outperforming LASSO (0.382) and SHAP (0.398).
- Redundancy: GRASP had the lowest Variance Inflation Factor (VIF = 2.942), while LASSO and SHAP exhibited multicollinearity (marked as "—" in the results).
Prediction Performance:
- GRASP achieved comparable or superior accuracy across logistic regression, Random Forest, and XGBoost models on both datasets (NHANES and UKB).
- In the UKB dataset, GRASP maintained high accuracy (0.755 for LR) with fewer features, demonstrating robust generalization.
Clinical Interpretability:
- Analysis of Lactate Dehydrogenase (LDH) showed that GRASP identified a transition point of 168.8 U/L, which was closest to established clinical thresholds compared to LASSO (144.66 U/L) and SHAP (144.04 U/L).
- Calibration curves and Kaplan-Meier survival analyses confirmed that GRASP offered better risk stratification and alignment with observed clinical risks.

5. Significance

GRASP represents a significant advancement in medical machine learning by bridging the gap between predictive power and clinical interpretability.

Robustness: By enforcing group sparsity and utilizing SHAP values for weighting, it overcomes the instability and redundancy problems common in LASSO.
Clinical Utility: The method generates feature sets that are not only accurate but also better aligned with known biological mechanisms and clinical thresholds, fostering trust and acceptance in healthcare.
Scalability: The use of Proximal Gradient optimization makes the method computationally feasible for large-scale electronic health record data.

The authors conclude that GRASP offers a stable, interpretable, and efficient solution for feature selection in precision medicine, with future work aiming to optimize efficiency for ultra-high-dimensional datasets.