Aims: Abnormal sound waves detection can find out whether the patient’s heart works normally. This study aimed to develop an algorithm to automatically identify the presence, absence, or unclear cases of murmurs in heart sound recordings.
Methods: All original phonocardiogram (PCG) signals were re-sampled from 4000Hz to 500Hz to make training faster and then fixed at 20 seconds in length. We extracted the first 20s of a recording if it was on less than 20s, Otherwise, zero-padding was applied. For patients with less than 4 recordings, missing recordings were zero-padded so that all signals have the same number of recordings. Biorthogonal wavelet transformation (bior2.6) was applied to reduce the noise in signals. We used SE_ResNet as our base model. It consists of one convolutional layer followed by 16 residual blocks, each of which contains two convolutional layers and a squeeze-and-excitation (SE) block. The SE block is used to model channel interdependencies. To solve the data imbalance issue, we used a Sign Loss function: the binary entropy loss function was extended to emphasize the loss of misclassified samples. Specifically, for the correctly classified labels, a coefficient smaller than 1 was multiplied by the default binary cross-entropy loss. By doing so, the accumulated loss from a large number of true negative labels became smaller, and the loss from the misclassified labels became more prominent.
Results: Our model received an average score of 1835 using the challenge evaluation metric on the offline 5-fold cross-validation datasets and 1449 on the online validation set. The name of our team is HeartBeats.
Conclusion: Our model can automatically identify the abnormal waves from heart sound recordings with good accuracy.