Classification of heart sound recordings using convolution neural network

Abstract

Aims: This study proposes a cardiac diagnostic model using convolution neural network (CNN). This model can predict whether a heart sound recording is normal or not by classifying phonocardiograms (PCGs) from both clinical and non-clinical environments - in accordance with the '2016 Physionet/CinCChallenge'. Methods: Heart sound recordings in the training data set are filtered by using Windowed-sinc Hamming filter algorithm to remove signals regarded as noise. The filtered recordings are then scaled and segmented. Using the filtered and segmented recordings, CNN is trained to extract features and construct a classification function. The CNN is trained by back propagation algorithm with stochastic gradient descent and mini-batch learning. To classify one sound recording, the signal should be filtered and segmented. Each segment of the signal is then classified by the trained CNN model. The model assigns each segment signal a relative probability between normal and abnormal labels. By accumulating these relative probability values for all the segmented signals, one can reliably and robustly determine whether the target signal is normal or abnormal. Results: The proposed model achieved an overall score of 79.5 with a sensitivity of 70.8 and a specificity of 88.2.