Rate-independent detection of atrial fibrillation by statistical modeling of atrial activity

• Published in: Biomedical signal processing and control Vol. 18; pp. 274 - 281
• Main Authors: Ladavich, Steven, Ghoraani, Behnaz
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2015
• Subjects: Atrial activity analysis; Atrial fibrillation; Electrocardiogram; Feature extraction and classification; Gaussian mixture model; P-wave absence; R–R interval variability; Atrial fibrillation; Electrocardiogram; Feature extraction and classification; Atrial activity analysis; P-wave absence; Gaussian mixture model; R–R interval variability
• ISSN: 1746-8094
• DOI: 10.1016/j.bspc.2015.01.007

•We developed a new statistical model to detect atrial fibrillation from ECG.•Unlike existing R–R interval-based algorithms, our method targets atrial activity and is heart rate independent.•The algorithm can work even if the patient has a pacemaker or is taking rate-control drugs, or if other heart rate issues occur simultaneously with AF.•The performance of the proposed method is demonstrated on real ECG data.•Our assessment showed a comparable performance to the R–R interval-based algorithms. In this study, we propose a P-wave absence (PWA) based method for atrial fibrillation (AF) identification over a short duration of electrocardiogram (ECG). The algorithm constructs a statistical model of normal sinus rhythm (SR) P-waves using a training set. Features extracted from P-waves are taken as an input to the Expectation–Maximization algorithm to create a Gaussian mixture model (GMM) of the P-wave feature space. The model is then used to identify PWA and detect AF. The algorithm performs AF identification in a single beat, and through post-processing of successive outputs using a majority voter determines the PWA over seven beats. The MIT-BIH Atrial Fibrillation Database was used to evaluate the algorithm. Classification using the majority voter showed a sensitivity of 98.09%, a specificity of 91.66%, a positive predictive value of 79.17% and an error of 6.88%. The performance of the proposed classifier is comparable to current R–R interval (RRI)-based algorithms, yet is able to detect short episodes of AF and performs rate-independent AF determination. The proposed algorithm targets atrial activity rather than ventricular activity that is targeted in RRI-based algorithms. It provides a patient specific detection of AF using a simple classifier, and can be leveraged as a tool to detect AF onsets/offsets over short AF episodes even when a patient's heart rate is controlled.