Deep Learning-Driven ECG Classification Using CNN-LSTM-Transformer and Metaheuristic Optimization

Abstract

Cardiovascular diseases (CVDs) represent a major source of sickness and death across every part of the globe; therefore, accurate diagnoses at an early stage are needed for treatments that work well. Electrocardiograms (ECGs) are a key way to find heart problems, but reading them by hand is time-consuming and can lead to mistakes. For complete automated ECG classification, this study proposes a deep learning-based approach that uses a hybrid CNN-LSTM and Vision Transformer (ViT-LSTM) architecture. CNN catches spatial features from ECG signals, and LSTM/Transformer modules allow a strong classification of cardiac arrhythmias and CVD indicators by catching temporal dependencies. The model was trained on benchmark ECG datasets like MIT-BIH Arrhythmia along with PTB-XL. For optimal feature selection, the model uses M-Square normalization in addition to Chaotic Dwarf Mongoose Optimization (CH-DMO). The experimental results show that the proposed approach achieves a classification accuracy of 94–98%, which is superior to what is achieved by conventional CNNs and LSTMs. The system's quite high sensitivity and truly outstanding specificity make it a completely viable solution for actual real-time, IoT-based smart health monitoring. Future work will thoroughly broaden dataset diversity and considerably optimize model efficiency for advanced edge computing to greatly improve generalization.