A Deep Ensemble Gene Selection and Attention-guided Classification Framework for Robust Cancer Diagnosis from Microarray Data

Abstract

Listen

Translate article

Microarray technology has enabled unprecedented insight into cancer diagnosis through large-scale gene expression analysis. However, the high dimensionality and complexity of microarray datasets pose significant challenges, as only a small subset of genes is typically informative, with the remainder introducing noise and complicating classification. Traditional gene selection methods, including filter, wrapper, and hybrid techniques, have achieved promising results but often fail to capture complex gene interactions, suffer from computational inefficiencies, or lack interpretability. This study presents DEGS-AGC (Deep Ensemble Gene Selection and Attention-Guided Classification), a novel integrated framework for gene selection and classification. DEGS-AGC is designed to address these limitations through two primary components: Deep Ensemble Gene Selection (DEGS), which leverages ensemble learning with Random Forest, XGBoost, and Deep Neural Networks to select relevant genes while reducing redundancy via sparse autoencoders, and Attention-Guided Classification (AGC), where an attention mechanism dynamically assigns weights to genes to improve interpretability and classification precision. The DEGS-AGC framework was evaluated against traditional methods, using consistent classification models for robust comparisons. Evaluation metrics demonstrated the potential of DEGS-AGC as an effective tool for high-dimensional biomedical data analysis. The results highlighted the ability of DEGS-AGC to offer accurate, interpretable, and computationally feasible solutions for cancer diagnosis, advancing the development of data-driven personalized approaches in healthcare.