HybridCardioNet: A CNN-LSTM-Based Deep Learning Framework for ECG Signal Classification and Cardiac Anomaly Detection

Abstract

Adaptive learning and automatic classification of ECG signals is one of the commonly processed and practical methods in cardiac anomalies detection with that area has a huge potential to teach clinicians for better clinical healthcare decision making and also remote health patient monitoring [7,8]. Past methods tackle issues like spatial–temporal feature extraction, class imbalance and dataset generalisation, but they are limited by a number of shortcomings: the traditional ML models depend on hand-crafted features and thus lack scalability, and the stand-alone deep models (CNNs or LSTMs) do not capitalize on spatial and sequential information simultaneously. To overcome these shortages, we present HybridCardioNet, a joint deep-learning framework that integrates CNN-based spatial feature extraction and LSTM based temporal-dependency modelling. The ECG was filtered using band-pass filtering (0.5–40 Hz) to remove the low-frequency baseline-wander, z-score normalisation, and segmented into single-beat segments using the R-peak detection from the MIT-BIH Arrhythmia Database; class imbalance was handled via class-weighted loss (random minority oversampling provided validation) HybridCardioNet with stratified cross-validation gives 98.39% accuracy with the same balanced precision, recall and macro-F1 score. Against popular protocols in recent literature on MIT-BIH, it also achieves competitive performance against internal baselines (CNN-only, LSTM-only and classical ML). Thus, hybridCardioNet solves one key limitation of the previously existing methods. With regards to the other two limitations, since hybridCardioNet is able to outperform the state-of-the-art for multi-class ECG classification, hybridCardioNet will be appropriate for real-time applications in terms of early detection & continuous ECG signals clinical/remote monitoring.