基于机器学习的心音识别分类研究

Research on heart sounds classification based on machine learning

心音信号可反映心脏的病理信息,是诊断心脏健康的重要依据之一。本文首先从心音信号提取时频域、梅尔倒谱系数等145个特征作为机器学习的输入数据集,然后在随机森林、LightGBM、XGBoost、GBDT、SVM共5种分类器中选出效果最佳分类器与递归特征消除算法结合进行数据挖掘,找出重要特征集并对其分类效果做比较与分析,最后运用Stacking模型融合方法优化模型。数据挖掘特征子集比同数量特征子集在准确率、召回率、精确率、F1值上分别提高了33.51%、14.54%、20.61%、24.04%;采用LightGBM和SVM模型融合可将F1值提高至92.6%。本文提出了一种有效的心音识别分类方法,挖掘出心音最重要的8个特征,为临床诊断提供参考。

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.