Enhancing feature selection for ordinal outcomes using resampling-based sparse linear discriminant analysis

Area of Science:

• Biostatistics
• Bioinformatics
• Computational Biology

Background:

• High-dimensional biomedical data with ordinal outcomes present feature selection challenges due to predictor correlations and small sample sizes.
• Sparse Linear Discriminant Analysis (sLDA) is used for classification and feature selection but suffers from instability and lack of reproducibility, especially with collinearity.
• Existing sLDA methods often fail to identify robust and interpretable biological marker sets.

Purpose of the Study:

• To develop a resampling-based ensemble sLDA framework to improve the stability and reproducibility of feature selection in high-dimensional biomedical data.
• To enhance the identification of biologically coordinated and interpretable marker signatures.
• To improve predictive performance and biomarker discovery in precision medicine.

Main Methods:

• A resampling-based ensemble Sparse Linear Discriminant Analysis (sLDA) framework integrating bootstrapping and subsampling was developed.
• Feature selection was based on Variable Inclusion Probability (VIP) aggregated across multiple resampled datasets.
• The proposed method was evaluated using simulation studies and applied to kidney renal papillary cell carcinoma staging and glioma grading datasets.

Main Results:

• The ensemble sLDA framework demonstrated improved stability and reproducibility of selected feature sets compared to standard sLDA.
• Simulation studies showed more accurate and consistent recovery of true predictors with the ensemble method.
• Applications revealed enhanced predictive performance and the identification of biologically interpretable and reproducible feature sets.

Conclusions:

• The proposed resampling-based ensemble sLDA framework offers a more stable and reproducible approach to feature selection for high-dimensional biomedical data.
• This method enhances biomarker discovery by identifying robust and interpretable feature sets, contributing to precision medicine.
• The framework's ability to handle collinearity and improve predictive accuracy makes it a valuable tool for analyzing complex biological datasets.