Mixing Matrix Pseudostationarity and ECG Preprocessing Impact on ICA-Based Atrial Fibrillation Analysis

Abstract

Abstract. In this work two relevant considerations in the ICA-based estimation of atrial activity (AA) in atrial fibrillation (AF) episodes from real electrocardiogram (ECG) recordings are presented. Firstly, the impact of low-pass filtering preprocessing on the extraction quality of AA is analyzed, showing an average improvement over 17% in spectral concentration (SC) when low-pass filtering is applied after ICA with respect to the application of the same filtering before ICA. Secondly, it is demonstrated that the ICA mixing matrix obtained from one AF segment can also be used to estimate the AA present in different segments of the same recording, thus proving the pseudostationarity of the mixing matrix. Results over 32 AF segments show a mean cross-correlation of R dp = 81.5% between the directly estimated AA and the estimated using presudostationarity. Changes in spectral concentration from one case to the other (∆SC dp = 1.4%) are negligible.

ECG signal quantization effects in the analysis of atrial fibrillation

“…The respiratory induced axis-shifts of all the analyzed ECG recordings were previously eliminated by high pass filtering and the network interference was canceled by notch filtering. On the contrary, low pass filtering of ECGs was not applied as a previous step to ICA because of the proved higher performance of ICA when post-filtering is used instead of pre-filtering [8]. This fact is explained by the increase of signal gaussianity as a consequence of low pass filtering [9].…”