Session PB3.4

Convolutive Multiband Blind Separation to Dissociate Atrial from Ventricular Activity in Atrial Fibrillation

C Vayá, JJ Rieta, R Alcaraz*

Universidad Politécnica de Valencia
Gandía, Spain

In order to use the ECG as a tool for the characterization of atrial fibrillation (AF), we need to dissociate atrial activity (AA) from ventricular activity. On the other hand, the reduced number of leads recorded from a Holter system limits the necessary spatial diversity required by Blind Source Separation (BSS) techniques to accurately extract the AA. The Wavelet Blind Separation (WBS) method partially solved this difficulty by including a wavelet decomposition of Holter leads prior to Independent Component Analysis (ICA). In this work, Convolutive Multiband Blind Separation (CMBS) is introduced as an improvement of WBS where the ICA stage has been replaced by the convolutive BSS algorithm Infomax. A first scenario of 15 synthetic ECG recordings and a second scenario of 18 real Holter recordings were considered. Three levels of signal to noise ratio (SNR) were studied: no noise, 15 dB and 5 dB. Four performance indexes were computed from the extracted AA: temporal correlation (Rt), spectral correlation (Rs), main peak frequency (fp) and the spectral concentration (SC). Concerning the first scenario and considering noise-free signals, the mean Rt obtained by CMBS, 0.8526, is higher than mean values obtained by WBS, 0.7801, and ICA, 0.8021. The mean Rs reached by CMBS, 0.8371, is higher than the ones obtained by WBS, 0.7066, and ICA, 0.8225. The mean fp estimated by CMBS, 5.49 Hz, is the best approach to the mean fp of the original mixed AA sources, 5.46 Hz. Also the mean SC obtained by CMBS, 62.82%, improves the results of WBS, 57.00%, and ICA, 61.78%, and it is the best approach to the mean SC of the mixed AA sources, 62.92%. In the second scenario, the estimated mean SC for real Holter ECG obtained by CMBS, 71.83%, exceeds the values obtained by WBS, 67.13%, and ICA, 43.23%. Furthermore, the level of noise affects less the estimation of fp by CMBS than by WBS or ICA. Results of CMBS were also better than those of WBS and ICA when the damaging effect of noise was considered. Our analysis shows up that CMBS improves the extraction performance of WBS and ICA in both scenarios so that a high accuracy of the estimated AA for synthetic and real AF ECG episodes is accomplished. In addition, results assure that CMBS preserves the original AA spectral parameters.

(Abstract Control Number: 54)