Multiscale Kernel Residual Convolutional Neural Network to Detect Heart Valve Diseases

The preliminary detection of heart valve diseases (HVDs) is essential to avoid lethal cardiac arrest. This work presents a multiscale kernel residual convolutional neural network (MK-ResCNN) model for detecting HVDs using Phonocardiogram (PCG) signals. The proposed method pre-processes and segments the PCG signal into PCG cycles. Second, a multiscale kernel CNN (MK-CNN) architecture is deployed to capture the HVDs pattern from the PCG cycles at different scales. Third, residual learning is used with the MK-CNN to extract complex features from deep CNN layers without any degradation in performance. The proposed method is tested and validated using a multi-class murmur database accessible via GitHub. The proposed method shows a precision of 99.74%, recall of 99.73%, F1- score of 99.74%, and overall accuracy of 99.73%. The promising results demonstrate that the proposed method is reasonably competitive with the baseline methods available in the literature.