Improved parametric estimation of time frequency representations for cardiac murmur discrimination

In this work a methodology of heart murmur detection by means of time-frequency representations (TFR) based on time-varying auto regressive (TVAR) modeling of phonocardiographic signals is proposed. Time-varying coefficients are estimated with Kalman smoother obtaining improved estimation precision and appropriate tracking of time-varying dynamics of phonocardiographic signals. TFRs derived from TVAR parameters are decimated with wavelet decomposition and taken to a feature space with PCA embedding (eigenfaces). Analysis of identification performance is accomplished for a database composed of 201 normal PCG records, and 201 murmurs. Results show that TFRs derived from Kalman smoother can discriminate normal heart sounds and murmurs better than other parametric TFRs obtained from LMS and RLS parameter estimation algorithms and non parametric TFRs based on Choi-Williams distribution.