Multi-resolution hybrid sliding window approach for ECG arrhythmia detection using deep learning models

Abstract Early detection of cardiac arrhythmias is critical for preventing life-threatening events. This study proposes a novel multi-resolution hybrid sliding window strategy for ECG beat-level classification, extracting 1080-sample segments from three overlapping temporal windows (180, 360, and 540 samples) centered on annotated R-peaks. These enriched representations are fed into two deep learning models: a 1D CNN and a stacked LSTM, trained to classify seven arrhythmia types using the MIT-BIH dataset. The dataset was partitioned into 80% training and 20% testing, with class-wise stratification to ensure balanced evaluation. The CNN architecture consisted of two convolutional layers with dropout and ReLU activation, followed by dense layers, while the LSTM model included two stacked layers with 64 units each. Both models were trained for 10 epochs with a batch size of 32, using the Adam optimizer. The proposed CNN model achieved an accuracy of 99.17%, outperforming several recent ECG classification models. We further validated performance using macro-F1 scores and Grad-CAM visualizations to highlight relevant ECG regions, enhancing interpretability. Our method demonstrates strong generalizability and clinical relevance, offering a robust and interpretable solution for real-time arrhythmia detection.