ScholarWorks@성균관대학교

초록

In this paper, we tested the efficacy of frozen electrocardiogram (ECG) representation vectors from a foundation model in detecting Chagas disease, as part of the George B. Moody PhysioNet Challenge 2025. Our team, Seoul Mates, utilized a pre-trained ECG foundation model (ECG-JEPA) which was trained using self-supervised learning on various ECG datasets. This approach learns robust ECG representations by predicting masked portions of the signal in a latent feature space, avoiding the pitfalls of reconstructing noisy raw signals. For the challenge, we applied a linear evaluation protocol on the features extracted from the pre-trained model without any fine-tuning. Interestingly, the feature representations can recover important ECG parameters, such as QRS duration and heart rate, suggesting the model’s potential as an off-the-shelf screening tool for Chagas disease. Using these fixed representation vectors, our model achieved the challenge score score of 0.217 on the hidden test set, ranking 18th out of 41 teams.