Abstract

Info

Author(s):: MA Jing; CAI Wenjie; YANG Li; College of Medical Instrument and Food Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China

Keywords:: Keywords: machine learning classification of heart sounds data mining model fusion

Abstract:: Abstract: The heart sound signals can reflect the pathological information of the heart and is therefore one of the important evidences for the diagnosis of heart diseases. In this paper, 145 features, such as time-frequency domain and Meier cepstrum coefficient, were extracted from heart sound signals as input data sets for machine learning, and then the best classifiers are selected from five classifiers, namely Random Forest, LightGBM, XGBoost, GBDT and SVM. The classifier is combined with the recursive feature elimination algorithm to conduct data mining, in order to find out the important feature set and compare and analyze its classification effect. Finally, the Stacking model fusion method is applied to optimize the model. The data mining feature subset improved the accuracy, recall rate, precision and F1 value by 33.51%, 14.54%, 20.61% and 24.04% respectively over the same number of feature subsets the fusion of LightGBM and SVM models improved the F1 value to 92.6%. In this paper, an effective heart sound recognition classification method is proposed to mine the eight most important features of heart sounds for clinical diagnosis.