A high-performance, lightweight deep learning framework for accurate detection of single-lead ECG arrhythmias
Abstract
Cardiac arrhythmia, a common sign of cardiovascular disease, is a primary cause of morbidity and mortality globally. Timely detection and categorisation of arrhythmias are essential for successful therapeutic intervention. Here, we present a novel hybrid deep learning model that employs single-lead ECG signals to accurately and efficiently classify arrhythmias by merging CNN, channel attention modules, and GRU(Gated Recurrent Units). This technique is optimised for resource-constrained settings, including Internet of Things (IoT) based systems and wearable medical equipment. The proposed framework utilises a four-layer CNN for local feature extraction, attention methods to emphasise clinically pertinent segments, and GRUs to record temporal dependencies, effectively addressing issues such as vanishing gradients in sequential data modelling. Additionally, to solve class imbalance and improve predictive performance on minority classes, the SMOTE-Tomek hybrid sampling method is employed. The model underwent thorough evaluation on two benchmark datasets, MIT-BIH Arrhythmia and INCART, attaining classification accuracies of 99.23% and 99.58%, along with macro F1-scores of 95.11% and 96.80%, respectively. The proposed model outperforms existing ECG single lead classification models in accuracy, F1-score, and sensitivity, highlighting its potential for real-time arrhythmia identification in clinical and remote environments.