Self-Supervised Learning for Automated ECG Signal Quality Assessment

Dieko De Graaf, Maria Galanty, Hoel Kervadec, Clara I. Sánchez

Track: Type A (Regular Papers)

Keywords: Self-supervised learning (SSL), electrocardiogram (ECG), signal quality assessment

Abstract: This work explores self-supervised learning (SSL) for automated ECG signal quality assessment, a crucial task since low-quality signals can trigger false alarms in medical monitoring and weaken disease-detection algorithms. Existing approaches rely on annotated datasets, which are often limited in availability and inconsistent in their defined quality classes. Those limitations can be addressed by using SSL models to learn high-quality representations of ECG signals without requiring annotated data. The effectiveness of these representations for signal quality assessment is examined by evaluating three SSL models, SimCLR, BYOL, and SwAV. The Brno University of Technology ECG Quality Database is used, a single-lead dataset annotated for ECG signal quality. A comprehensive set of augmentations and parameters is explored for pre-training the SSL models and their performance is compared to supervised (KNN) and unsupervised (k-means) baselines. The results show that all SSL models outperform the baselines, with SwAV achieving the highest macro F1 on the test set with 81.85\%. However, divergence between validation and test performance for low-quality signals reflects dataset limitations; nonetheless, the results demonstrate that SSL has great potential for ECG signal quality assessment.

Serve As Reviewer: ~Maria_Galanty1

Submission Number: 51