CNN-LSTM and Wavelet Deep Learning for Robust Arrhythmia Classification

Abstract

Deep learning has significantly advanced automated arrhythmia detection, but achieving optimal performance across diverse arrhythmia types remains challenging. This research presents two hybrid deep models, CNN-LSTM and DWT-CNN-LSTM, to classify ECG arrhythmia based on the combination of time, space, and frequency-domain features. CNN-LSTM attained 99.17% accuracy, and DWT-CNN-LSTM reached 99.46%, outperforming the benchmark CNN-LSTM model (99.29%). The suggested CNN-LSTM enhanced sensitivity (~0.8%) and specificity (~0.2%), and DWT-CNN-LSTM further boosted specificity to 99.83%, one of the highest reported. Class-wise analysis illustrated almost perfect identification of Normal rhythms (sensitivity: 99.74%, specificity: 97.90%) and strong performance for PVC identification (F1-score: 97.50%), with DWT-CNN-LSTM keeping false alarms to a minimum. For RBBB and LBBB, both models achieved more than 99%, and in atrial fibrillation detection, CNN-LSTM provided better sensitivity (88.89%) and F1-score (92.06%), whereas DWT-CNN-LSTM demonstrated better specificity. These results validate the robustness, clinical applicability, and superiority of the proposed models over the current CNN, LSTM, and hybrid architectures. CNN-LSTM provides balanced sensitivity, whereas DWT-CNN-LSTM has improved specificity, and both are appropriate for wearable monitoring in real time and clinical decision support.