Classification of phonocardiograms using residual convolutional neural network and MLP

Guohui Peng, Haitao Zou, Jin Wang
Jiangsu University of Science and Technology


Our method does not require signal segmentation or segment identification, it is simpler than identifying and segmenting S1, systole, S2 and diastole. Our team is JUST_IT_Academy1, in this competition, we design our strategy aiming at avoiding signal segmentation and segment identification. During our data preprocessing, each audio data is converted to a 128-dimensional vector by computing its Mel-scaled spectrogram. After fixed-length processing, we input such data into a residual convolutional neural network (RESNET), and input age, height, weight, and other characteristics to Multilayer Perceptron (MLP). Then we connect the output of RESNET and MLP, and assign the result to the fully connected layer for classification. The loss is calculated by binary cross entropy. Our method is applied to the 2022 George B. Moody PhysioNet Challenge. In the cross-validation, our murmur classification has the Accuracy of 0.872, AUROC of 0.923, and AUPRC of 0.874. The cross-validation result of the outcome classifier has the Accuracy of 0.637, AUROC of 0.707, and AUPRC of 0.701. The weighted accuracy of our method in the official competition is 0.723 (higher is better), the cost of the outcome classification is 10,692 (lower is better).