Feature Extraction and Recognition of Heart Sound Signal Based on CEEMDAN Sample Entropy

Abstract

针对心音信号的非平稳特性和易被噪声干扰的特点，本文提出一种基于自适应噪声的完备经验模态分解(CEEMDAN)与IMF样本熵结合的特征提取方法。将信号进行CEEMDAN自适应分解为若干个IMF分量，并计算各阶IMF分量的样本熵作为特征向量。在此基础上提出一种基于因子分解机(Factorization Machines, FM)的推荐模型，能更好的处理稀疏大数据的缺点，较好的解决了样本熵的稀疏性。为了验证该模型的优劣，进行了AUC曲线分析。通过对600例先心病病例心音和600例正常心音实验数据分析，证明该方法能够改善信号特征提取的效果，对先心病心音类型上的判断表现出较高的识别率。 Due to the nonstationary characteristics of heart sound signal which was often disturbed by noise, a feature extraction method based on complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) with IMF sample entropy was proposed in this work. The heart sound signals were adaptively decomposed into several IMF components by using CEEDDAN, and the sample entropy of each order IMF component was calculated as the feature vector. A recommendation model based on Factorization Machines (FM) was proposed, which can deal with the disadvantages of sparse big data and solve the sparsity of sample entropy better. In order to verify the pros and cons of the model, AUC curve analysis was performed. 600 heart sounds of congenital heart disease and 600 normal heart sounds were analyzed. It is proved that the method can improve the signal feature extraction and show a higher recognition rate for the heart sound of congenital heart disease.