Caracterización de Espectrogramas Usando Análisis de Componentes Principales y Medidas de Energía para Detección de Soplos Cardíacos

Abstract

A methodology based on spectral analysis of digitized auscultation signals is presented. Those signals are known as phonocardiographic signals (PCG), oriented to detection of cardiac murmurs originated by valvular pathologies. Initially, a filtration system based on the wavelet transform is developed to reduce the disturbances that usually appear in the acquisition stage. A between-beats segmentation algorithm is developed which uses information of the ECG signal previously acquired in a synchronous way to hook the beginning of the QRS complex with the beginning of the S1 sound of the PCG signal. Intra-beat segmentation is proposed for detecting S1, S2, systole and diastole based on the relationship analysis of energy and threshold. Features derived from the spectral analysis are extracted using Principal Component Analysis applied to the spectrograms and energy measures over the segments were the murmurs are located. Feature effectiveness is evaluated by a k-NN type classification model for separating the classes: normal and murmur. The database of PCG records used belongs to the National University of Colombia; 164 records of this labelled database were used: 81 records labelled “normal” and 83 records labelled “murmur”. Finally, with the help of specialist doctors, 180 representative normal beats and 180 representative beats with evidence of cardiac murmur were chosen. Classification precision, sensitivity and specificity results were obtained. The best result of classification precision was 95,6% with sensitivity and specificity values equals to 96,1% and 95%, respectively.