Heart Murmur Detection Using an Ensemble Model Based on Multi-Feature Fusion

Yue Zhao, Zhuoyang Xu, Tingting Zhao
PingAn Chuangke


Aims: We aim to develop a model with high performance of accuracy and generalization, which can automatically identify the presence, absence, or unclear cases of murmurs according to routine demographic information and heart sound recordings collected from multiple auscultation locations on human body. Methods: Wavelet transformation was applied to reduce the noise in recordings. In addition to the routine demographic information in the challenge data, we also extracted time domain features and frequency domain features from each recording of the patient. In order to eliminate the differences among patients with similar demographic features, we divided the patients into various groups and measured the specificity of each patient relative to others in the group. Specifically, for each patient, such specificity was calculated by subtracting the average of normal heart sound signals of all the patients in the group. Then, we trained a baseline model of boosting ensemble trees. In order to solve the problem of class-imbalance, we adopted a dynamic weighting strategy. By updating the sample weight dynamically, we increased the misclassification loss of the minority classes and decreased the loss of the majority classes. Therefore, the model would pay more attention to the samples of minority classes and reduce the overall missed diagnosis. Results: Our model achieved the average score 1954 for offline cross-validation and 2129 for online testing. The name of our team is sakura. Conclusion: Our model is capable of automatically identifying the presence, absence, or unclear cases of murmurs with good accuracy and generalization.