A Deep Learning Model for Recognizing Pediatric Congenital Heart Diseases Using Phonocardiogram Signals

Diagnosing congenital heart disease (CHD)in children through heart sound auscultation requires extensive medical training and understanding. However, the quality of PCG data may be compromised due to the sensor location, a child's developing heart, and the complex and changeable cardiac acoustic environment. This study proposes a one-dimensional Convolution Neural Network (1D-CNN) with a residual block that classifies PCG signals to predict heart abnormalities in 751 patients with PCG signals aged five months to twenty years. After assessing the signal quality, only good-quality signals are used as input features of the neural network. The study's results showed the accuracy of 0.93 accuracy and 0.98 sensitivity. The Receiver Operating Characteristic (ROC) plot yielded an Area Under Curve (AUC) value of 0.98, and the F1-score was 0.94. The proposed model required only 15 sec of the PCG signals to predict CHD cases (4.2 ms processing time). Thus, it can be implemented as a primary screening tool for remote-end pediatricians by providing cheaper and faster interpretations of PCG signals before referring the cases to specialists.