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

Motivation: High-dimensional biomedical datasets with ordinal outcomes-such as cancer stages or treatment responses-pose significant challenges for feature selection due to strong predictor correlations and limited sample sizes. Sparse Linear Discriminant Analysis (sLDA) is widely used for simultaneous classification and feature selection. However, concerns about model stability and reproducible feature selection persist, particularly in the presence of pronounced collinearity inherent in biomedical data. Consequently, direct application of sLDA often fails to capture a reproducible set of biologically coordinated markers, resulting in signatures that lack robustness and interpretability. Results: We propose a resampling-based ensemble sLDA framework that integrates bootstrapping and subsampling to improve the stability of feature selection. By aggregating results across multiple resampled datasets, the method identifies features based on Variable Inclusion Probability (VIP) rather than relying on coefficients from standard sLDA. Compared with standard sLDA, this ensemble strategy reduces sensitivity to data perturbation and improves the stability and reproducibility of selected feature sets. Simulation studies demonstrate that the proposed ensemble framework achieves more accurate and consistent recovery of ground-truth predictors compared with the standard (non-resampled) sLDA. Applications to kidney renal papillary cell carcinoma staging and glioma grading datasets further suggest that this framework can improve predictive performance and identify biologically interpretable and reproducible feature sets, highlighting its potential utility for reliable biomarker discovery in precision medicine. Availability and implementation: The source code used in the study is available via GitHub at https://github.com/ryan-wng/RE-sLDA.