SViT-ECG: Spectrogram Vision Transformer for Detection of Short-Term Atrial Fibrillation from ECG Signals

Atrial fibrillation (AF) is a prevalent cardiac arrhythmia that can cause severe complications, such as stroke and heart failure, highlighting the critical importance of early and precise diagnosis. Timely detection and intervention are essential, as they can markedly reduce the risk of these adverse outcomes. This study proposes a novel approach to detecting short episodes of AF in patients' electrocardiogram (ECG) signals using transformer-based deep learning techniques. Initially, the ECG segments were transformed into spectrograms, which served as input for the proposed fine-tuned Vision Transformer model (SViT-ECG). The model demonstrated impressive performance, with an average 5-fold cross-validation accuracy of 98.14% and an F₁-score of 95.81%. When validated on an unseen ECG dataset, the SViT-ECG model achieved an accuracy of 95.97% and an F₁-score of 91.14%. These promising results suggest that the proposed method could be effectively extended to real-time AF detection applications, offering a significant advancement over current state-of-the-art algorithms.