Session P33.2

Improved Parametric Estimation of Time Frequency Representations for Cardiac Murmur Discrimination

LD Avendaño-Valencia, JM Ferrero*, G Castellanos-Domínguez

Universitat Politècnica de València
Valencia, Spain

In this work, a methodology of estimation of improved precision is carried out, based on Kalman smoother approach, for spectral components dynamics of phonocardiographic (PCG) signals, in order to discriminate between normal events and murmurs. The automatic discrimination system is oriented to clinical diagnosis, where murmurs must be differentiated from normal events with highest sensitivity. In this paper, estimation of time-frequency representations (TFR) is done by parametric approach which has as advantages representation parsimony, as models may be potentially specified by a limited number of parameters, improved precision and resolution, and improved tracking of time-varying dynamics, which makes them suitable for analysis of non stationary processes, particularly, those related with heart sounds. The proposed methodology, firstly, consists on adaptive estimation of time varying autoregressive process parameters with Kalman smoother, resulting in a time-frequency surface containing dynamic information of signal’s spectral components. Secondly, in order to reduce the dimension of representation, the obtained surface is characterized by means of a simple PCA model (eigenfaces), whose parameters are estimated and used as signal’s features. Analysis of identification performance is accomplished for a database composed of 201 PCG records, labeled as normal, and 201 as murmurs. For the sake of benchmarking of Kalman smoother, the database is characterized also by LMS and RLS recursive estimators, besides it is analyzed non parametric TFR estimation by means of Choi-Williams distribution. As a result, classification performance of proposed methodology reaches 90%, being better than the other studied cases of recursive and non parametric TFR estimation; the former one showing an overall performance of just 70%. Besides, computational cost of Kalman smoother is near of RLS estimator, which is the lowest, and definitely lower than the processing cost for Choi-Williams distribution.

(Abstract Control Number: 228)