Heart Murmur Severity Stages Classification Using Multikernel Residual CNN

Computer-aided diagnosis of heart murmurs (HMs) using phonocardiogram (PCG) signals facilitates early treatment of heart valve diseases (HVDs). This work proposes a novel multikernel residual convolutional neural network (MK-RCNN) model to classify HM severity stages. The model employs multikernel CNN-based feature extractors to capture scale-specific features from raw PCG segments. Furthermore, residual learning (RL) aids in extracting relevant features from deep CNN layers without degrading performance accuracy. The effectiveness of the proposed method is verified through comprehensive testing on three open-access databases. The proposed MK-RCNN model demonstrates an overall accuracy (OA) of 98.33%, overall precision (OPr) of 96.98%, overall recall (ORe) of 97.36%, and overall
$F1$
-score (OF1) of 97.05% using the Physionet Challenge 2022 database for HM severity stage classification. We have also evaluated the efficacy of the proposed model for multiclass categories of HMs using the heart sound murmur (HSM) database and binary classification as healthy or pathological using the Physionet Challenge 2016 database. The proposed model’s accurate predictions, low computational complexity (CC), and speedy response make it well-suited for reliable and fast HM diagnosis in primary healthcare units.