CardioSyncNet: A Scalogram-Driven Ensemble of CNN, BiLSTM, and Transformer for Robust ECG Arrhythmia Classification

The accurate delineation of EKG signals is paramount for automated arrhythmia detection, bolstering clinical decision-making. The MIT-BIH Arrhythmia Database, featuring five beat types- NB, SPB, PVCB, FVNB, and UB, presents hurdles due to class imbalance and intricate temporal-spatial complexities, often eluding single-model approaches in capturing holistic features. This investigation unveils CardioSyncNet, a sophisticated deep learning ensemble amalgamating Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and Transformer models, leveraging scalogram-based insights for precise classification. ECG signals are transformed into 224×224 RGB scalograms via Continuous Wavelet Transform (CWT) with Morlet wavelet, with the dataset harmonized through downsampling and class weight optimization. The CNN, fortified with residual connections, extracts spatial signatures, which are refined by BiLSTM for temporal continuity and a custom multi-head self-attention Transformer (tailored for MATLAB) for extended context. Employing 5-fold cross-validation, the ensemble merges predictions via validation accuracy-weighted fusion. Performance is gauged using accuracy, precision, recall, F1-score, weighted F1-score, and ROC-AUC, yielding a 99.72% mean cross-validation train accuracy (with 99.72% precision, recall, and F1-score), a 92.28% test accuracy, a 72.07% weighted F1-score, and a 0.97 average AUC, affirming robust generalization despite imbalances. CardioSyncNet’s pioneering scalogram and attention integration delivers exceptional accuracy and efficiency, marking it as a potent candidate for clinical arrhythmia solutions.