Heart Sound Classification Method based on Ensemble Learning

Heart sound classification aims to explore a reasonable classification method of pathological events. The single-scale character only shows part of the information on heart sound signal. Different machine learning methods have a difference in performances. This paper studies the performance differences between multi-resolution features of different classifiers, and uses an ensemble learning method that combines features and classifiers to improve the accuracy of automatic heart sound classification. In this work, we extract the time domain, frequency domain, and statistical characteristics as feature vectors, then use the Principal Component Analysis (PCA) technology to reduce dimensions for avoiding redundant features, and apply feature fusion to splice features from multiple fields. Finally, this paper uses a stacking method for heart sound classification. It consists of two layers, SVM, Random Forest, and KNN are the primary classifiers as the first layer, and the secondary is a logistic regression model. The performance evaluation has an evident improvement with the ensemble learning method, its accuracy, sensitivity, specificity, and the overall score are 97.28%, 98.45%, 96.12%, 97.285% respectively.