Noninvasive estimation of organization evidences differences between paroxysmal and persistent atrial fibrillation

Methods: The materials correspond to 14 synthetic recordings and 40 real recordings from the Clinical University Hospital, Valencia, Spain. The proposed method (frequency domain algorithm, or FDA) maximizes the power content in the range of frequencies of 3 to 10 Hz for the first uncorrelated recordings to assure the decoupling between the atrial and ventricular activities. The central frequency of this range of frequencies corresponds to an initial estimation of the atrial frequency from the T-Q intervals of lead V1. Fig. 1 shows the corresponding block diagram. The performance measures used to assess the quality are correlation (ρ), spectral concentration (SC), and kurtosis. In the case of real recordings, we also estimate the peak frequency of the extracted atrial activity. To compare the results, Fast ICA and PCA were also applied to the same recordings. Results: We have analyzed the performance of the 3 algorithms with synthetic and real recordings. The results are shown in the Table 1 (statistical analysis in the last line). The new method outperforms the classical blind source separation methods applied to the estimation of atrial activity, Fast ICA and PCA. We have also analyzed the behavior of the methods under realistic noise conditions. In this case, in terms of correlation, SC, and kurtosis, FDA is specially robust compared with other algorithms. In the last experiment, we have checked that FDA does not require an accurate prior estimation of the initial atrial rate to extract the atrial activity successfully. Conclusion: The proposed method presents the following advantages: simplicity and efficiency in computational time, because it is a fixed-point algorithm and a source extraction method; that is, it does not require the computation of the whole set of sources and the postprocessing identification of the atrial component among them. It has been applied successfully to nonnoisy and noisy simulated ECG and to real ECG, obtaining better results than other algorithms in terms of the proposed quality parameters. These results show that (i) the use of higher order statistics is not necessary to solve the problem (Fast ICA) and (ii) the use of decorrelation can be enhanced (PCA). Finally, the algorithm is robust to deviations in the initial estimation of the central frequency.