Extraction of Desired Signal Based on AR Model with Its Application to Atrial Activity Estimation in Atrial Fibrillation

The use of electrocardiograms (ECGs) to diagnose and analyse atrial fibrillation (AF) has received much attention recently. When studying AF, it is important to isolate the atrial activity (AA) component of the ECG plot. We present a new autoregressive (AR) model for semiblind source extraction of the AA signal. Previous researchers showed that one could extract a signal with the smallest normalized mean square prediction error (MSPE) as the first output from linear mixtures by minimizing the MSPE. However the extracted signal will be not always the desired one even if the AR model parameters of one source signal are known. We introduce a new cost function, which caters for the specific AR model parameters, to extract the desired source. Through theoretical analysis and simulation we demonstrate that this algorithm can extract any desired signal from mixtures provided that its AR parameters are first obtained. We use this approach to extract the AA signal from 12-lead surface ECG signals for hearts undergoing AF. In our methodology we roughly estimated the AR parameters from the fibrillatory wave segment in the V1 lead, and then used this algorithm to extract the AA signal. We validate our approach using real-world ECG data.