Murmur Classification with U-net State Prediction

Heart sounds are recorded which is known as phonocardiography (PCG). However, characteristic of PCG and influence of noise, the detection and classification of PCG are challenging. Generally, murmur is detected in systolic of in diastole state. Our team predicted four states of PCG and classified heart sound into three class that are absent, present, and unknown. S1, systolic, S2, diastole state was predicted by U-net model. The results of U-net prediction were used as feature along with the denoised PCG signal for classification. Transformed denoised PCG signals were classified by ResNet multi-classification method. U-net accuracy of predicting four state of heart signal was 0.97. The mean squared error (MSE) score was 0.02. Our team MainLab's challenge training time according to the leaderboard was 2 hours 45 minutes, and Murmur weighted accuracy was 0.583 and outcomes was 18475 respectively. Ranking of our team was 31 out of 40. Novelty of our proposed method is that murmur is detected on systolic and diastole state and, our method predicts additional information. Additional information in this case would be four states of heart sound S1, systolic, S2, and diastole. And classify murmur from PCG signal combined with predicted state of PCG.