Robust Hybrid Deep Learning for ECG Classification and Heart-Disease Risk Prediction

Abstract

Extracting information of a given electrocardiogram (ECG) signal is crucial to the early detection and treatment of cardiac diseases, but automated detection of the results on a large scale is still impeded by signal noises, inter-patient variations, and intricate time-dependent relationships. In order to overcome these difficulties, a hybrid deep-learning system of convolutional neural networks, bidirectional long short-term memory units, and attention machines is designed to jointly encode morphological and temporal representations of ECG signals. Noise resilience is improved by comprehensive preprocessing, adaptive segmentation, and ECG-specific data augmentation, and also class imbalance is reduced. Experimental analysis indicates performance on par with the state of art, reaching an average F1-score of 98.36% and AUC = 0.995 on the MIT-BIH dataset, and 97.08% F1-score on PTB-XL, which indicates good generalization and convergent behaviour. The low misclassification between morphologically close rhythms and interpretable images of attention reflects the clinical reliability of the model, and this formed a strong and explainable base of automated arrhythmia detection and heart disease risk prioritization.