Pseudo anomalies enhanced deep support vector data description for electrocardiogram quality assessment.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Machine Learning

Background:

• Wearable electrocardiogram (ECG) recordings face noise interference, degrading signal quality.
• Traditional supervised signal quality assessment (SQA) methods struggle with diverse noise and unknown anomalies.
• High variability in ECG signals and noise challenges traditional SQA method generalization.

Purpose of the Study:

• To develop a simple, effective, unsupervised SQA method for ECG signals.
• To enhance the detection of anomalies and improve the generalization of SQA methods.
• To provide a reliable measure of ECG signal quality from wearable devices.

Main Methods:

• Modeled SQA as an anomaly detection problem using a pseudo-anomalies enhanced deep support vector data description.
• Introduced novel ECG noise-generation methods to simulate real-world noise scenarios.
• Learned a generalized hypersphere of high-quality ECG data in a self-supervised manner using generated pseudo-anomalies.

Main Results:

• The proposed unsupervised method effectively measures ECG quality based on distance to the learned hypersphere center.
• Demonstrated effectiveness across multiple public datasets and a real-world 12-lead ECG dataset.
• Achieved superior performance compared to traditional methods in detecting signal anomalies.

Conclusions:

• The unsupervised anomaly detection approach offers a robust solution for ECG SQA in wearable devices.
• The pseudo-anomalies enhanced deep support vector data description improves model generalization.
• This method provides a reliable and effective way to assess ECG signal quality, even with unknown noise types.