Reproducible feature selection in heterogeneous multicenter datasets via sign-consistency criteria.

Area of Science:

• Biostatistics
• Epidemiology
• Health Informatics

Background:

• Identifying disease risk features is vital for clinical decision-making.
• Between-center heterogeneity complicates traditional feature selection due to inconsistent covariate effects.
• Existing methods often struggle with data privacy and homogeneity assumptions.

Purpose of the Study:

• To propose a novel framework for selecting reproducible risk features with consistent effects across data centers.
• To address challenges posed by between-center heterogeneity in feature selection.
• To develop a method that protects data privacy and does not assume data homogeneity.

Main Methods:

• Developed a novel framework for reproducible risk feature selection.
• Quantified feature reproducibility using a sign-consistency criterion.
• Evaluated the method through extensive simulations and application to real-world data.

Main Results:

• The proposed method demonstrated greater power compared to existing feature selection techniques.
• Identified nine significant risk factors for depression in the China Health and Retirement Study Longitudinal Study (CHARLS).
• The sign-consistency criterion effectively balances heterogeneity and signal similarity.

Conclusions:

• The novel framework successfully identifies reproducible risk features even with between-center heterogeneity.
• The approach offers enhanced data privacy and robustness compared to traditional methods.
• This method provides a powerful tool for reliable risk factor identification in multi-center studies.