Machine learning can be employed to automate detection of diseases and patterns that are not detectable by traditional ECG analyses. However, contemporary machine learning tools require large labeled datasets, which can be scarce for rare but serious diseases. Self-supervised learning (SSL) can address this data scarcity.
Using an extensive clinical data set of 36,519 ECGs, we implemented the Momentum Contrast framework, a form of SSL. We assessed the learning using low Left ventricular ejection fraction (LVEF) detection as the downstream task using 1\%, 5\% and 10\% of total ECGs. We compared the SSL improvement of LVEF classification across different input augmentations: 1) Gaussian Noise 2) Gaussian Blur 3) Scaling 4) Magnitude Warping 5) Baseline Warping 6) Time Warping 7) Window Warping
Downstream performance varied across hyperparameters, and optimal hyperparameters varied across training set sizes, with gaussian blur showing the best results compared to no SSL baseline. Across augmentations, contrastive SSL produced a marginal improvement in AUC compared to the baseline, contrary to literature expectations. Our findings suggest that the augmentations applied to the ECG data did not markedly enhance the model's ability to discern patterns related to low LVEF detection. Future studies will address limitations, such as larger datasets for pre-training, evaluation on multiple encoders, and exploring SSL contrastive frameworks.