A stable feature selection method based on majority voting and SHAP for high-dimensional metabolomics data

Area of Science:

• Metabolomics
• Bioinformatics
• Computational Biology

Background:

• Metabolomics enables large-scale metabolite measurement for disease research.
• High-dimensional, small-sample metabolomics data requires robust feature selection.
• Current methods lack stability, impacting feature consistency.

Purpose of the Study:

• To develop a stable and accurate feature selection framework for metabolomics.
• To enhance the reliability of biomarker screening in complex diseases.

Main Methods:

• Proposed MVFS-SHAP framework integrating majority voting and SHAP values.
• Utilized cross-validation and bootstrap sampling for robust feature subset generation.
• Employed Ridge regression and Linear SHAP for feature re-ranking and selection.
• Evaluated stability using the extended Kuncheva index and predictive performance via PLS regression.

Main Results:

• MVFS-SHAP demonstrated superior stability and predictive accuracy over existing methods.
• Achieved high stability (0.90+) on Exo and Endo datasets, with 80% of results >0.80.
• Maintained stability (0.50-0.75) even on challenging datasets.
• Reduced RMSE values across various predictive models (Lasso, Random Forest, XGBoost).

Conclusions:

• MVFS-SHAP offers a stable and effective solution for feature selection in metabolomics.
• Framework shows robustness in handling noisy, complex data for reliable biomarker selection.
• Future work includes enhancing adaptability and applying to precision medicine and TCM research.