Wearable electrocardiogram (ECG) devices enable long‑term cardiac monitoring but are subject to diverse noise sources—baseline wander, power‑line interference, and motion artifacts—that reduce signal usability and risk misdiagnosis. The proposed method, NA‑SQA (global‑key region collaborative noise‑aware ECG signal quality assessment), targets unsupervised settings where labeled anomalous data are scarce. NA‑SQA combines global quality prediction with a key‑region branch that uses the reconstruction residual from an autoencoder as a focused noise representation. By jointly optimizing global, key‑region, and consistency losses, the model integrates overall signal context with discriminative noise‑related features. Evaluation on four public ECG quality datasets (BUTQDB, Icentia11K, EHOQA, and EAWQA) yields F1 scores of 91.42%, 79.62%, 81.59%, and 81.56%, respectively, and demonstrates improved robustness and generalization compared with prior unsupervised methods.
Wearable ECG monitoring is increasingly widespread due to portability and comfort, facilitating continuous recordings. However, realistic monitoring environments introduce mixed and variable artifacts—baseline wander, power‑line interference, and motion artifacts—that can degrade signals and potentially lead to missed or incorrect clinical interpretations. Automated ECG signal quality assessment (SQA) helps identify and remove unusable recordings prior to downstream tasks, improving diagnostic accuracy and reducing clinical workload.
Supervised deep learning approaches have advanced SQA but depend on large volumes of labeled data, which are expensive to produce. Unsupervised anomaly detection approaches, including autoencoders and Deep‑SVDD, avoid the need for labeled anomalies by modeling normal data distributions and flagging deviations. Yet, many unsupervised methods focus primarily on global characteristics and use reconstruction error or feature distances directly for anomaly scoring, often failing to localize and exploit discriminative noise‑related regions. This limitation reduces their capacity to distinguish noise contamination from normal physiological variability and weakens generalization to unseen noise conditions.
Prior SQA work falls into traditional signal‑processing methods and deep‑learning methods. Traditional approaches extract statistical and morphological signal quality indices (SQIs) such as power spectral measures, QRS peak kurtosis, inter‑detector agreement, and inter‑lead consistency, then classify quality using classical classifiers. Deep‑learning approaches have incorporated residual blocks with recurrent modules, convolutional architectures with data augmentation, and time‑frequency image representations classified with CNNs. Representative unsupervised approaches include autoencoders for reconstruction‑error‑based detection, Deep‑SVDD to model compact feature space descriptions, and methods that use pseudo‑anomalous samples or dual transformation modeling. However, most lack focused modeling of local noise regions and therefore may not fully capture discriminative features required for robust generalization.
NA‑SQA is an unsupervised framework that explicitly couples a global quality prediction branch with a key‑region branch derived from reconstruction residuals. The workflow comprises three main steps:
Data synthesis and labeling: Various artificial noises are injected into high‑quality ECG signals to generate low‑quality samples. The injected noise types serve as interference labels for constructing pseudo‑noise supervision.
Autoencoder reconstruction: All signals (original and synthesized) are passed through an autoencoder trained to reconstruct high‑quality ECG signals. The autoencoder output is used to compute a residual signal defined as the difference between the original input and the reconstructed signal.
Dual‑branch prediction and joint optimization: The original input is fed to a prediction model to obtain a global quality prediction. The residual (reconstruction difference) is treated as the key noise region and processed by a parallel prediction model to produce a key‑region quality prediction. Training jointly optimizes three objectives: a global prediction loss, a key‑region prediction loss supervised by the unified quality labels, and a consistency loss that encourages agreement and collaborative learning between global and key branches.
Key design considerations: the residual branch is supervised not only by pseudo‑noise labels but also by clean‑signal labels, allowing it to learn both large noise‑induced deviations and subtle variations present in high‑quality signals. The consistency constraint leverages contextual information from the global branch while focusing the key branch on local discriminative responses, improving the model’s capacity to separate noise artifacts from physiological variability.
NA‑SQA was evaluated on four public ECG quality assessment datasets: BUTQDB, Icentia11K, EHOQA, and EAWQA. The proposed method achieved F1 scores of 91.42% on BUTQDB, 79.62% on Icentia11K, 81.59% on EHOQA, and 81.56% on EAWQA. Across these datasets, NA‑SQA consistently outperformed state‑of‑the‑art unsupervised ECG quality assessment methods reported in the study, demonstrating enhanced robustness and generalization to diverse and unseen noisy conditions encountered in real‑world wearable monitoring.
Extensive experiments included comparisons to conventional autoencoder residual approaches and other unsupervised strategies. The results indicate that modeling the residual explicitly as a key noise region under global supervision and enforcing consistency between branches yields better discrimination of noise contamination and reduces false positives arising from subtle physiological variations.
The paper discusses limitations and proposes directions for continued research. Specific limitations and future work items were outlined in the source; details on experimental constraints, scalability, deployment, or additional augmentation strategies were described there. Where the source omitted particular implementation or hyperparameter details, those specifics were not reported.
NA‑SQA introduces a global‑key region collaborative and noise‑aware approach to unsupervised ECG signal quality assessment. By combining global context with a targeted residual‑based key region supervised under unified quality labels and enforcing a consistency constraint, the framework improves detection of noise‑related contamination while preserving sensitivity to legitimate physiological variations. Evaluations on four public datasets show improved F1 performance over prior unsupervised methods, supporting NA‑SQA’s potential for more robust ECG quality evaluation in wearable and remote monitoring applications.