Global-key region collaborative noise-aware for electrocardiogram signal quality assessment

Wearable electrocardiogram (ECG) devices enable long-term cardiac monitoring due to their convenience and comfort. However, various types of noise, such as baseline drift, power-line interference, and motion artifacts, can severely degrade signal quality and compromise the reliability of automated diagnosis. Consequently, ECG signal quality assessment (SQA) plays a crucial role in ensuring accurate downstream analysis. Existing supervised SQA methods rely heavily on large-scale annotated datasets, which are costly and time-consuming to obtain in real-world applications. Although unsupervised anomaly detection methods alleviate the dependence on labeled data, most existing approaches primarily focus on global signal characteristics and directly utilize reconstruction errors or feature-space distances for anomaly detection, while overlooking discriminative noise-related regions. As a result, their ability to distinguish noise contamination from physiological signal variations remains limited, leading to suboptimal generalization performance. To address this limitation, we propose a global-key region collaborative noise-aware ECG signal quality assessment method (NA-SQA) for unsupervised settings. NA-SQA integrates both global information and key noise-region features to enable more accurate quality evaluation. Specifically, we generate low-quality signals and corresponding quality labels by injecting various types of artificial noise into high-quality ECG signals. All signals are then passed through an autoencoder to reconstruct high-quality ECG signals. Subsequently, the original signals are used for global quality prediction, while the difference between the original and reconstructed signals is treated as the key noise region for key-region quality prediction. By jointly optimizing the global and key quality prediction tasks, the proposed model effectively integrates global information with more discriminative noise-region features, thereby enhancing its generalization capability to unseen low-quality ECG signals. Extensive experiments on four public ECG quality assessment datasets, namely BUTQDB, Icentia11K, EHOQA, and EAWQA, demonstrate the effectiveness of the proposed approach. NA-SQA achieves F1 scores of 91.42%, 79.62%, 81.59%, and 81.56% on the four datasets, respectively, consistently outperforming state-of-the-art unsupervised ECG quality assessment methods and demonstrating superior robustness and generalization capability for ECG signal quality assessment in real-world noisy environments.