RDLHeart: Statistical Gene Selection with Deep Learning and SMOTE for Heart Failure Classification from Bulk RNA Sequencing

Abstract

Cardiovascular diseases are the leading cause of global mortality. Accurate molecular classification of heart failure etiology remains challenging due to extreme class imbalance and ultra-high dimensionality of genomic data. This study proposes a deep learning framework for multi-class heart failure subtype classification directly from real Next-Generation Sequencing (NGS) RNA-sequencing (RNA-seq) profiles of human left ventricular tissue. We utilized the GSE116250 dataset consisting of 64 samples: 14 Non-Failing (NF) controls, 37 Dilated Cardiomyopathy (DCM), and 13 Ischemic Cardiomyopathy (ICM) cases, yielding 54,675 gene expression features per sample after preprocessing. To prevent data leakage, stratified train/test splitting (80/20) was performed prior to any augmentation. The Synthetic Minority Over-Sampling Technique (SMOTE) was then applied exclusively to the training set to address severe class imbalance. Three sequential models were evaluated: one-dimensional Convolutional Neural Network (1D-CNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). On the held-out test set, the proposed 1D-CNN with SMOTE achieved 99.68% accuracy and 0.5255 loss, significantly outperforming the same architecture without SMOTE (87.50% accuracy, 0.6978 loss). LSTM and GRU models also showed consistent gains with SMOTE (approximately 11–13% accuracy improvement), confirming the general benefit of balanced training in NGS-based cardiology applications. This work demonstrates, for the first time, the successful integration of leakage-free SMOTE with deep sequential models on real human cardiac RNA-seq data and establishes a strong, reproducible benchmark for etiology-specific heart failure diagnosis directly from NGS transcriptomes.