In this paper, we propose a novel method based on self-supervised learning (SSL) to address the simplicity bias (SB) problem in electrocardiogram (ECG) analysis. The diagnostic value of ECG lies in its dynamic features, such as subtle waveform variations and rhythm fluctuations that evolve over time and frequency domains. However, existing supervised learning-based ECG models tend to overfit dominant and repetitive patterns, thereby overlooking subtle but clinically important clues, which are known as SB. In this study, we propose an SSL-based method that mitigates SB and captures features reflecting dynamic characteristics by utilizing time-frequency-aware filters and multi-particle prototype reconstruction. We construct a large-scale multicenter ECG dataset containing more than 1.53 million ECG records and conduct experiments on three subtasks. We show that the proposed method effectively reduces SB and achieves state-of-the-art performance. The code and dataset will be made public.
