False Alarm Reduction in Atrial Fibrillation Detection Using Deep Belief Networks

Abstract

We propose and validate a novel method to reduce the false alarm (FA) rate caused by poor-quality electrocardiogram (ECG) signal measurement during atrial fibrillation (AFib) detection. A deep belief network is used to differentiate acceptable from unacceptable ECG segments. To validate the method, eight different levels of ECG quality are provided by artificially contaminating ECG records, from the MIT-BIH AFib database, with motion artifact from the MIT-BIH noise stress test database. ECG segments classified as “unacceptable,” in terms of signal quality, are restricted from AFib detection process. Results are evaluated for each level of quality and compared to AFib detection algorithm performance when ECGs of each level of quality are applied to it without performing any classification. Our results show that AFib detection performance for ECG with high signal-to-noise ratio (SNR) is minimally affected by this FA reduction approach. For clean ECG (no added noise), the AFib detection accuracy was 87%, without and with FA reduction. For ECG, with an SNR of −20 dB, the performance of AFib detection is markedly decreased with an accuracy of 58.7%; however, with FA reduction (using our method) the accuracy was increased to 81%.