Optimized Machine Learning for Cancer Classification via Three-Stage Gene Selection

Abstract

Gene selection from high-dimensional microarray data presents challenges such as overfitting, computational inefficiency, and feature redundancy. Despite significant advances, existing methods often suffer from limitations in scalability and interpretability, especially for precision oncology. This study introduces a novel Three-Stage Gene Selection (3SGS) strategy that addresses these issues through a combination of filter-based methods (signal-to-noise ratio, correlation coefficient, ReliefF) with accuracy-driven refinement and redundancy reduction. The 3SGS approach identifies minimal but highly predictive gene subsets, achieving 100% accuracy for leukemia and 98% for prostate cancer using only 3-4 genes. Compared to traditional methods, 3SGS enhances efficiency and interpretability, establishing itself as a scalable and robust solution for cancer classification.