Classification of Phonocardiograms Using Residual Convolutional Neural Network and MLP

Our method does not require signal segmentation or segment identification, it is simpler than identifying and segmenting $SI$, 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 was applied to the 2022 George B. Moody PhysioNet Challenge. In the test set, the Accuracy score of murmur classification was $0.757$, AUROC score was $0.797$, AUPRC score was 0.610, and Weighted Accuracy score was 0.671. The outcome Accuracy score was 0.562, AUROC score was 0.624, AUPRC score was 0.631, Weighted Accuracy score was 0.612, and the Cost was 13,394. We were ranked 23th in the murmur classification and 24th in the clinical outcome classification.