Vision Transformer for Arrhythmia Classification

In this study, we introduce a novel methodology for classifying four different classes of arrhythmia using one-lead Electrocardiogram (ECG) recordings. Our approach integrates a time-frequency analysis technique, specifically the Stockwell Transform, which is well-suited for biomedical signal processing, along with a Vision Transformer network. The synergy of these two components enhances the accuracy and efficiency of arrhythmia classification. Following the recommendations of the AAMI standard, we target four classes of Normal, Ventricular Ectopic, Supraventricular Ectopic, and Fusion beats. We evaluated the model's performance using the MIT-BIH arrhythmia database under intra-patient paradigms. We achieved an accuracy, precision, and sensitivity of 98.80%, 98.81%, and 98.80%, respectively.