A Rhythm-Specific ECG Signal Quality Assessment Framework for Robust Cardiac Health Monitoring of AI-based Arrhythmia Classifer

Electrocardiogram (ECG) is an essential approach for monitoring cardiovascular diseases (CVDs). However, ECG signals obtained from wearable devices are often corrupted by noise and artifacts, particularly electrode movement (EM) artifacts that are hard to eliminate due to their similarity to ectopic heartbeats and result in significant false positive rates. Signal quality assessment (SQA) can mitigate this effect by identifying and filtering out noise-contaminated ECGs. Prior SQA studies relied on human defined signal quality standards and unsuitable for AI-based arrhythmia classifiers. Therefore, in this work, a rhythm-specific SQA framework for AI-based arrhythmia classifier is proposed and provides auxiliary reliability analysis of the classification results. Evaluated on the MIT-BIH and NSTB ECG datasets, it achieves 92.2%/98.5% noise detection sensitivity and specificity, and reduces up to 80.7%/55.4% false alarm rate for ‘SVEB’ and ‘VEB’ classification in noisy ECG signals. Cooperated with the AI-based arrhythmia classifier, this SQA method is suitable for application in wearable ECG devices and ensure robust cardiac health monitoring in daily life.