Heart Sound Detection Based on Bidirectional Multilayer Recurrent Convolutional Neural Network

This research proposes an abnormal heart sound classification algorithm based on an improved Bidirectional Multilayer Recurrent Convolutional Neural Network (BMRCNN). Through the convolutional layer and recurrent layer of BMRCNN, more effective heart sound features are extracted from the image and timing features. This study was inspired by CRNN and modified the original network structure. Compared with previous studies, it improves accuracy and reduces misdiagnosis. Using the PhysioNet2016 dataset as the experimental object, the problem of imbalance between positive and negative samples is solved through data preprocessing. This research first uses framing technology and Butterworth filter as the preprocessing method for extracting Mel Frequency Cepstral Coefficient (MFCC) features, and then trains the neural network with the second-order differential MFCC feature data to classify abnormal heart sounds, and finally uses Neural Network Intelligence (NNI) ultrasonography. The parameter search framework optimizes model hyperparameters. Experimental results show that the trained model can successfully classify abnormal heart sounds, and the best accuracy rate reaches 99.35%.