Feature learning augmented with sampling and heuristics (FLASH) improves model performance and biomarker...

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Large biological datasets, like gene expression profiles, often contain redundant features.
• Redundancy degrades model performance and limits generalization, especially with class imbalance and hidden sub-clusters.

Purpose of the Study:

• Introduce FLASH, a novel feature selection method.
• Address challenges of redundant features in biological datasets.
• Improve model performance and generalization.

Main Methods:

• FLASH combines filtration and heuristic-based systematic elimination.
• Utilizes multiple statistical tests (t-test, ANOVA, Wilcoxon Rank-Sum, Brunner-Munzel, Mann-Whitney) on random samples.
• Ranks features using machine learning model coefficients and recursively eliminates features with cross-validation.

Main Results:

• FLASH preserves predictive performance on independent datasets.
• Outperforms dRFE, Mutual Information, MRMR, ElasticNet, NeuralNet, Permutation test, and SAGA.
• Selected features show greater biological relevance, with higher overlap with disease-associated genes from DisGeNET.

Conclusions:

• FLASH is an effective feature selection method for biological datasets.
• Improves model generalization and identifies biologically relevant features.
• Offers a robust alternative to existing feature selection techniques.