Robust ECG Classification using Mamba and Self-Supervised Representation Learning

Electrocardiograms (ECG) offer a quick and noninvasive method to analyze heart disorders. However, analysis of ECGs under non-ideal or noisy situations presents challenges, even for experts. Recently, automated analysis of biosignals using machine learning has gained popularity. However, approaches require large amounts of data, and the scarcity of large public datasets can result in poor generalization and increased sensitivity to noise. This study proposes the self-supervised pre-training of a Mamba-based model to denoise ECG data, and use the resulting latent representation for downstream ECG classification tasks. Our results shows that pretraining a self-supervised Mamba-based model increases robustness to noise, increasing classification performance in noisy situations. Furthermore, we evaluate optimal noise parameters to pre-train the model, and show that the proposed approach offers better performance with reduced computational cost compared to an equivalent transformer-based approach.