Selección de características 2D en representaciones tiempo frecuencia para la detección de soplos cardíacos

Abstract

In this paper is proposed a methodology for dimensionality reduction of time-frequency representations (TFRs) aimed to nonstationary biosignal classification that deals directly with large quantity of irrelevant and redundant data, combining a stage of feature selection with a stage of dimensionality reduction by linear decomposition methods extended to bidimensional data. The methodology is tested on a set of parametric TFRs computed from a phonocardiographic signal database (PCG) for detection of heart murmurs. Results show an improvement compared with other methodologies that do not account for irrelevant and redundant data in these representations and demonstrate that the use of bidimensional linear decomposition methods adequately reduce redundancy on TFRs, obtaining a new feature set of lower dimension than the original dataset.