Arrhythmia Classification from ECG Signals Using Transformers and Data Balancing Techniques

Abstract

While many arrhythmias pose minimal threat, certain heart rhythm irregularities elevate the potential for stroke or heart failure. The complexity arises particularly with the supraventricular premature heartbeat which has a resemblance to a normal beat and occurs infrequently. Consequently, this research proposes a data balancing and classification technique that enhances the accuracy of identifying mentioned hard-to-classify heartbeats while maintaining robust metrics for other classes. The study introduces a deep learning framework combined with a multi-head attention transformer, for balancing – under-sampling and synthetic minority oversampling are used. To evaluate the proposed model, various experiments based on real data were conducted. The results were compared with an existing model used in chest belt heartbeat monitoring, and the results show that the transformer model achieved better performance for supraventricular premature heartbeats, at the same time reaching high overall and per-class metrics.