Time-Frequency Domain Modified Vision Transformer Model for Detection of Atrial Fibrillation using Multi-lead ECG Signals

Atrial fibrillation (AF) based cardiac disease is interpreted based on the appearance of abnormal atrial activity and irregular inter-beat duration in the electrocardiogram (ECG) signal. The sustained AF episodes are more likely to develop severe cardiac disorders such as cardiomyopathy and myocardial infarction. Therefore, AF detection at an early stage using ECG signals is challenging in clinical standards to monitor the progression of this pathology. In this article, a novel time-frequency (TF) domain deep learning approach is proposed for the automated detection of atrial fibrillation (AF) using multi-lead ECG signals. The wavelet-based synchrosqueezing transform (WSST) is used to evaluate the TF image of each lead ECG signal. Furthermore, the modified vision transformer (ViT) model detects AF arrhythmia using the WSST-based TF images of the three best lead ECG signals. The proposed approach is tested using a public database’s multi-lead ECG time series. The results reveal that the suggested WSST-based modified ViT model has obtained the accuracy, precision, recall, and F1-score values of 95.50%, 93.36%, 98.00%, and 0.957, respectively. Compared to TF images of other lead ECG signals, the WSST-based TF images of lead II ECG signals yielded the ViT model an accuracy value of 94.50%. To detect AF using multi-lead ECG signals, the proposed approach outperformed continuous wavelet transform (CWT) and short-time Fourier transform (STFT) domain convolutional neural network (CNN) models in terms of classification performance.