Session P33.3

Effective Phonocardiogram Segmentation Using Nonlinear Dynamic Analysis and High-Frequency Decomposition

AF Quiceno, E Delgado, M Vallverdú*, AM Matijasevic, G Castellanos

Universitat Politècnica de Catalunya
Barcelona, Spain

The acoustic signals involved in the cardiac dynamic have high within-class variability. Effective segmentation is a very complex task due to the high distortion induced by cardiac murmurs in the temporal trace. In this study, an effective methodology for segmenting the temporal trace of phonocardiographic signals (PCG) is presented. Initially, between-beat segmentation is carried out using the DII lead of the ECG recording for locating the occurrence of the first heart sound (S1), as the beginning of S1 occurs together with the origin of the respective QRS complex. Next, the within-beat segmentation is achieved by using recurrence time statistics, which is sensitive to changes of the reconstructed attractor in a state space derived from nonlinear dynamic analysis. The recurrence time statistics are obtained using a sliding window of 2000 points over the sampled signal (fs=44100 Hz) with 90% overlap for taking well-resolution in time-domain. The dynamic attractor is reconstructed by using time-delay embedding, where the embedding parameters (i.e., embedding dimension and time lag) had to be estimated. The embedding dimension was obtained using the false neighbor method and the time lag was calculated minimizing the mutual information between each state variable and maximizing the distance from each point to the main diagonal of the state space. By recurrence time thresholding together with physiological rules, the preliminary segmentation is achieved. The results are tested using clinical requirements, with the aim of founding the cases where failures of the algorithm can be presented in the detection of the second heart sound (S2). Therefore, an alternative segmentation is proposed using thresholding over the Shannon envelogram extracted from the high-frequency decomposition. The database of PCG records which was used belongs to the National University of Colombia. The between-beat segmentation accuracy was 100% over all PCG recordings. Taking into account 1440 segments from 360 PCG beats, where a set of 180 beats were strongly disturbed by different types of cardiac murmurs, the within-beat segmentation yielded a sensibility result of 100% and a positive predictive value of 97.7%.

(Abstract Control Number: 226)