HybridArrhyNet: a CNN-BiLSTM framework with attention mechanism for robust arrhythmia classification from ECG signals

Abstract

Arrhythmia classification is critical for the early diagnosis of cardiac disorders that may lead to sudden cardiac arrest, stroke, or heart failure. However, accurate detection from electrocardiogram (ECG) signals remains challenging due to noise, class imbalance, and the need to identify diagnostically important signal segments. This study presents HybridArrhyNet, a deep learning framework that integrates convolutional neural networks (CNNs), bidirectional long short-term memory (BiLSTM) networks, and an additive attention mechanism for robust arrhythmia classification. The CNN extracts spatial features from ECG signals, while the BiLSTM captures temporal dependencies in both forward and backward directions. The attention mechanism selectively emphasizes clinically relevant ECG segments, improving classification performance and interpretability. To address class imbalance, random oversampling and weighted cross-entropy loss are employed. The proposed model was evaluated on the MIT-BIH Arrhythmia Database containing over 100,000 annotated beats across five arrhythmia classes (N, S, V, F, and Q). Experimental results demonstrate superior performance, achieving 98.7% accuracy, 98.5% precision, 98.6% recall, and 98.5% F1-score. Ablation studies further confirm the effectiveness of each component, highlighting HybridArrhyNet as a reliable framework for automated arrhythmia detection.