Aims: Heart murmur detection plays a crucial role in the early diagnosis of congenital and acquired heart diseases in children. This study aimed to construct a deep neural network architecture for detecting heart murmurs from PCG recordings. The model was created by the team "fly\_h” for the PhysioNet/Computing in Cardiology 2022.
Methods: The PCG signals collected from different auscultation positions were downsampled to 2000Hz, and then a sliding window method was used to clip the signal to 6000 samples. Next, the MFCC features of the PCG signals are extracted. To learn effective features, we constructed a ResNet with selective kernel convolution (SK-Conv). The SK-Conv was embedded into each ResBlock, which adaptively captures multi-scale features using convolution filters of different kernel sizes and applies a channel attention module (similar to Squeeze-and-Excitation) to emphasize the representation of important features. Our model was trained on record-level data and validated on patient-level data. For each patient-level data, it may contain 1-5 PCG recordings, i.e. AV, MV, PV, TV, Phc, and the final prediction result was selected from the corresponding record-level prediction results in the priority order of presence, absence and unknown.
Results: Using the scoring metric based on the costs for algorithmic prescreening for human experts for heart murmur identification, we scored 11331 in the official stage. In addition, the weighted accuracy of the proposed method reached 0.557.
Conclusion: The proposed heart murmur detection model performed well on the validation set. Such models may be used to assist physicians in diagnosing