Visualization and classification of the heart sounds of patients with pulmonary hypertension

Abstract

Background Pulmonary hypertension (PH) is defined as a mean pulmonary artery pressure (PAP) of higher than 25 mmHg with a pulmonary capillary or left arterial pressure of less than 15 mmHg. PH is a non-curable disease commonly triggered by a preexisting disease and patients who develop PH usually have poorer prognoses than those who do not. Right heart catheterization, which is invasive and is not risk free, has been the gold standard in the diagnosis of pulmonary arterial hypertension. However, changes in the second heart sound induced by increased PAP can be used in clinical settings for the non-invasive estimation of PAP. This study thus aims to find a noninvasive method using heart sound classification to screen PH in primary care setting. Methods Thirty-two subjects undergoing right heart catheterization in three cardiac centers were recruited for this study and divided into two groups, a PH and a non-PH group, with a defined boundary of mean PAP at 25 mmHg. Recordings of 20 seconds duration were made at a sampling frequency of 44,100 Hz. The selected phonocardiogram was processed with time-frequency spectrum analysis and the normah'zed average Shannon energy versus time axis to extract the diagnostic features. Principal component analysis was performed to decrease the number of diagnostic features before designing the artificial neural network (ANN). The Perceptron neural network and the Multilayer perceptron - Back propagation neural network were used to classify diagnostic features to predict the PAP value. All of the networks had different layers with different numbers of hidden neurons, and they were all trained with different learning algorithms for 10 runs. A regression analysis of the network response between the network outputs and the corresponding target outputs specified by the PAP value was performed. Of all the different structures, the best and mean performances among the 10 runs for each algorithm were compared. Results Six principal components of heart sound features were used in the ANN training. The network using the Resilient Backpropagation algorithm with a log sigmoid transfer function at the two hidden layers including 10 hidden neurons in each layer and a linear transfer function at the output layer performed the best among all ANN design structures and achieved the highest R value of 0.86 between the predicted output and the target output specified by right heart catheterization measured PAP value. Conclusion An ANN-based heart sound classification for PH in human subjects was explored in this study with a promising result, achieved. This novel method of cardiopulmonary assessment is expected to lead to the development of an automatic noninvasive device for the high-volume screening of PH.