A boundary-guided strategy via pseudo-anomaly feature perception for unsupervised ECG signal quality assessment

Electrocardiogram (ECG) signal quality assessment is vital for ensuring accurate ECG diagnoses and efficient medical services. Traditional supervised ECG signal quality assessment methods are limited by a restricted set of noise categories, which results in poor generalization ability. On the other hand, existing unsupervised ECG quality assessment methods lack guidance from abnormal features and effective hard-sample mining strategies, resulting in vague decision boundaries and poor discrimination between noise and true pathological ECG signals. To address these issues, this paper proposes an unsupervised ECG quality assessment method based on pseudo-anomaly boundary guidance. By constructing pseudo-anomaly features to explicitly guide decision boundary optimization, the proposed method enhances model accuracy. Specifically, we use an approximate iterative greedy kernel set algorithm to extract representative features from noisy ECG signals and build a noise feature library. Key noise features are then randomly sampled from this library and fused with high-quality ECG features to generate pseudo-anomaly features. The model is subsequently trained to constrain the high-quality ECG and pseudo-anomaly features within their respective threshold ranges. Additionally, we introduce a boundary guidance strategy based on pseudo-anomaly feature perception, which pulls high-quality ECG features and pseudo-anomaly features within the aliasing region toward their respective feature boundaries. This enables the model to establish a clearer decision boundary, thereby enhancing its ability to more accurately differentiate between noise and true disease-related ECG signals. Extensive experiments on multiple ECG signal quality assessment datasets demonstrate the effectiveness of the proposed method.