Compressed sensing for the extraction of atrial fibrillation patterns from surface electrocardiograms

The non invasive analysis of atrial fibrillation (AF) arrhythmia represents a challenge nowadays. The fibrillatory pattern of AF, known as f-wave, is partially masked by the ventricular activity of the heartbeat in the surface electrocardiogram (ECG). Classical techniques aiming to extract the f-wave are based on average beat subtraction (ABS) or blind source separation (BSS). They present limitations in performance and require long ECG records as well as multi-channel records in the case of BSS. The originality of the present work consists in exploiting the sparsity of the atrial activity (AA) signal in the frequency domain to extract the fullf-wave using a recent data acquisition technique called compressed sensing (CS). The present contribution takes a step forward in the extraction of the f-wave by exploiting the time rather than the space dimension. We intend to recover the AA signal with a variant of CS where classical random sampling is replaced by a block sampling scheme. Our breakthrough finding consists in the ability of our method to accurately extract the AA from a short ECG record of just one heartbeat, with a normalized mean squared error of 15%, which is unfeasible with ABS, BSS and other variants that require longer observation windows.