Atrial activity extraction from atrial fibrillation episodes based on maximum likelihood source separation

A novel non-invasive approach for the atrial wave estimation in atrial fibrillation (AF) episodes is presented. The method is based on the maximum likelihood (ML) solution of independent component analysis (ICA). The densities of the bioelectric independent sources corresponding to the ventricular and atrial activities are previously analyzed, and the prior knowledge extracted from them is considered in order to develop an appropriate separation model. As a consequence, the sources can be recovered through the optimization of the ML criterion. The algorithm is validated using a significant database of synthesized and real AF electrocardiograms (ECGs). A simulation model for the generation of realistic AF signals with known AA is also defined. The results show good performance in terms of SNR and correlation indices when estimated AA is compared to known AA. Regarding the real AF ECGs, the AA sources could always be estimated. Successful AA extraction was validated using spectral parameters. The main frequency of the atrial wave ranged from 4.9 to 7.4 Hz, and the spectral concentration was 49.8% on average.