Enhanced ECG Classification using a Hybrid Signal Processing Framework

Accurate Electrocardiogram (ECG) classification is crucial for real-time cardiac monitoring. This study integrates static and wavelet-based scattering transform features to classify four heart rhythms: normal (N), other (O), atrial fibrillation (A), and noisy signals (~). Extracted features include mean, standard deviation, root mean square, skewness, kurtosis, and band power. The Symlet-2 wavelet, chosen for its symmetry, enhances classification accuracy by optimizing decomposition levels to balance feature retention and noise reduction. Zero-padding or symmetric padding mitigates boundary effects, preserving feature integrity. We evaluate LSTM, CNN-LSTM, TCN, and Transformer models using 5-fold cross-validation on the imbalanced PhysioNet 2017 dataset, employing F1-score and AUC-ROC metrics. Feature extraction improves accuracy, with CNN-LSTM achieving 66.83% accuracy, an F1-score of 0.67, and an AUC of 0.75–0.82. LSTM and TCN show moderate performance (F1-score 0.59, AUC 0.52–0.72). Without feature extraction, all models perform worse, though CNN-LSTM remains the best (AUC 0.67–0.83). Feature extraction also reduces inference time from 274.40s to 155.94s. Data augmentation (time warping, jittering, time masking, magnitude warping, scaling, cropping) further improves CNN-LSTM performance for Classes N, A, and O, though Class ~remains unaffected. These results underscore the effectiveness of hybrid feature extraction and augmentation in enhancing ECG classification for real-time health monitoring.