Hybrid Statistical-Spectral Sparse Feature Selection with Optimization for Robust and Generalizable Lung Adenocarcinoma Classification

High dimensionality, redundant features, and poor cross-dataset generalization hinder Lung Adenocarcinoma (LUAD) classification using gene expression data. This study proposes Hybrid Statistical-Spectral Sparse Feature Selection with Optimization (HS3FS+), a novel framework that integrates Mutual Information (MI) and Kullback-Leibler (KL) divergence for feature ranking, Kernel Principal Component Analysis (KPCA) for nonlinear transformation, pathway-guided filtering for biological validation, and Genetic Algorithm (GA)-based optimization for feature selection. The framework was validated on four independent datasets: The Cancer Genome Atlas (TCGA)-LUAD, Gene Expression Omnibus (GEO) datasets GSE19188 and GSE37745, and TCGA-Lung Squamous Cell Carcinoma (TCGA-LUSC), ensuring robust cross-platform evaluation. HS3FS+ achieved classification accuracy of 98.3% on TCGA-LUAD, 97.1% on GSE19188, 96.0% on GSE37745, and 94.8% on TCGA-LUSC. The selected gene signatures exhibit strong concordance with established LUAD biomarkers, supporting both biological relevance and model interpretability. Additionally, the method demonstrated a fivefold reduction in computational time compared to Deep Learning (DL)–based feature selection approaches. These findings confirm HS3FS+ as a robust, interpretable, and scalable solution for LUAD classification, with potential applications in biomarker discovery and precision oncology.