Cardio Ensemble Net for Automated Classification of Normal and Abnormal Heart Sounds Using Phonocardiogram Signals

Abstract

Cardiovascular diseases CVDs are still considered as one of the major causes of death among people on the earth, and the early diagnosis based on cardiac auscultation is usually restrained by subjectivity and unavailability of specialists. Phonocardiogram signals offer an automated cardiac screening technique that is non-invasive, cost effective, but automatic classification of the normal heart sound and abnormal heart sound is a problem because of noise, non-stationarity, and inter patient variability. This paper presents Cardio Ensemble Net CENet, a hybrid soft voting ensemble system of automated normal and abnormal heart sounds classification by combining multi domain features. It is aimed to create a strong, general, and computationally effective system that can detect abnormalities with reliability using short PCG recordings without ECG gating. The suggested model is a combination of Support Vector Machine, radial basis kernel, Random Forest and Gradient Boosting classifiers using probabilistic soft voting using a 138 dimensional feature set generated using time domain, frequency domain, MFCC, chroma, wavelet and envelope descriptors. The PhysioNet 2016 dataset is evaluated on stratified 5 fold cross validation performance. CENet has a mean accuracy of 98.47 %, sensitivity of 97.56 %, specificity of 98.91 % and F1 of 98.01 %. Standalone SVM, Random Forest and Gradient Boosting compared to baseline models had average percentage improvements in accuracy of 2.3, 1.9 and 2.6 respectively. The results show high strength and lower false negative. The suggested design has great opportunity to be implemented in a handheld electronic stethoscope and community health screening system.