EQA-MDL: Wearable ECG Signal Quality Assessment via Multi-Scale Difference Learning

Signal quality assessment (SQA) is crucial in data collection and analysis based on wearable devices. However, most existing electrocardiogram (ECG) quality assessment methods rely on labeled data and struggle to accurately locate noisy areas. To address this issue, we propose a self-supervised method based on multi-scale difference learning, called EQA-MDL, which utilizes a noise generation module to provide pseudo-anomaly samples for difference learning and combines a reconstruction framework with multi-scale difference learning to capture key features for accurate anomaly localization. Specifically, we first design an ECG noise generation method to simulate real-world noise scenarios, thereby generating corresponding pseudo-anomalous samples along with their pixel-level labels. Second, a reconstruction-based self-supervised framework is introduced that leverages pseudo-anomalies to adversarially learn optimal latent feature representations of clean ECG signals. Finally, we design a multi-scale difference learning function to amplify the representational differences between noisy and high-quality ECG features, while incorporating representational correlations from the observation space into the feature space, thereby extracting more robust features. Extensive experiments conducted on two public datasets and three private datasets demonstrate that the proposed method significantly outperforms state-of-the-art approaches.