Deep Learning Models for Analyzing Phonocardiogram Signals to Detect Heart Abnormalities

Phonocardiography (PCG) is an extremely powerful tool for recording heart sounds. Auscultation is usually utilized by cardiologists to detect cardiovascular diseases (CVDs). However, it is prone to errors, and it is not convenient for longtime monitoring. Hence, automatic analysis of PCG recordings is needed to avoid human errors and facilitate longtime monitoring. In this paper, a method is proposed for detecting heart abnormality by analyzing PCG signals. Four deep learning models have been investigated, namely VGG16, LeNet-5, AlexNet, and Google Net. The PCG signals are filtered, segmented, and analyzed using Mel Scale spectrogram. Thereafter, binary classification is performed using the considered deep learning model. The best average accuracy achieved is 87.09% using Google Net. All experiments have been carried out using PhysioNet dataset.