Towards Feasible Home ECG Monitoring: AI-Driven Detection of Clinically Critical Arrhythmias Using Single-Lead...

Differentiating life-threatening arrhythmias, such as ventricular tachycardia and supraventricular tachycardia, from non-threatening ones is crucial for clinical applications. This study aimed to develop a deep learning model to classify five key Electrocardiogram (ECG) patterns: normal sinus rhythm, sinus tachycardia, sinus bradycardia, supraventricular tachycardia, and ventricular tachycardia. We collected 1500 single-lead 10 s ECG signals from public datasets, including PhysioNet/Computing in Cardiology (CiC) Challenge 2020 and the Malignant Ventricular Ectopy Database, for training and 2297 ECGs for testing. Each 10 s signal was decomposed into 1 s sliding windows with a 5-point stride, which served as the input for the proposed deep learning architecture utilizing temporal attention and Time2Vec embedding. The model performance achieved an overall accuracy of 95.2%. For the five classes-supraventricular tachycardia, sinus tachycardia, normal sinus rhythm, ventricular tachycardia, and sinus bradycardia-the model achieved sensitivities of 90.3%, 92.9%, 97.4%, 100.0%, and 99.0% and accuracies of 96.3%, 95.8%, 98.9%, 99.9%, and 99.5%, respectively. Specificities for all rhythm categories exceeded 97.4%. This simple and effective single-lead model can significantly support the growing trend of home healthcare and professional clinical decision-making.