Session P91.5

Combined Analysis of Time and Frequency Series Regularity Applied to the Study of Atrial Fibrillation

C Vayá, JJ Rieta*

Universitat Politècnica de València
Valencia, Spain

Atrial Fibrillation (AF) is the most common arrhythmia encountered in the clinical practice. AF has an estimated prevalence of 0.4% to 1% in the general population, with a continuous increase along with the age, reaching the 8% in people older than 80 years old. In this work, a new method based on electrocardiogram (ECG) signal processing is carried out in order to distinguish between AF episodes that will terminate immediately and those that will sustain. This new method is based on a combined analysis of the atrial activity (AA) series regularity in both time and frequency domains. The organization is measured by using the non-linear regularity estimator Sample Entropy (SampEn).
The signal database consisted of 50 surface ECGs in AF properly annotated by cardiologists as terminating or non-terminating episodes.. The ECG recordings analysis is completed in five main steps: extraction of the AA, computation of the spectrogram, construction of spectral parameter series, SampEn computation in time and frequency domains and discriminant analysis. We have used Average Beat Subtraction (ABS) to separate the AA from the rest of the cardioelectric signal as a preliminary step. The spectrogram is computed using Hamming windows of 1024 samples in length and 75% overlap. Twelve spectral parameters were computed from every AA spectrogram so that twelve numerical series are obtained. Then the SampEn of these series of spectral parameters and the SampEn of the AA in the time-domain are combined in a multivariate analysis. The variable selection was performed by forward stepwise analysis and minimization of the Mahalanobis’ distance.
Twenty of these recordings were used as learning set to create the mathematical model and the rest as test set. As a result, the discriminant function obtained from the learning set is a plane given by the equation x3 = 0.0355 •x1 + 1.6 • x2 + 0.4653, where x1, x2 and x3 represent the SampEn of the main peak amplitude (fp1), the SampEn of the normalized distance between fp1 and the second largest peak (fp2), and the SampEn of the AA, respectively. All of the cases used to create the model, i.e. the learning set, were correctly classified. Regarding the test set, 28 out of 30 cases were correctly classified. In global percentages, forty eight out fifty recordings (96%) were classified correctly by the discriminant analysis, which provided an improvement of 6% with respect to the univariate analysis of SampEn.

(Abstract Control Number: 92)