面向心音分割的个性化高斯混合建模方法

准确的心音分割是分析和处理心音信号的基本前提。主流的心音分割算法采用监督式预先训练的方法构建统计模型,它不仅依赖于繁琐的手工标注,还存在模型与被分割数据之间的不匹配问题。提出了一种面向心音分割的个性化高斯混合建模方法,避免了手工标注和预先训练,而且在线训练获得的个性化模型能够高度匹配被分割的心音数据。由于心音信号的周期在一段短时间内很稳定,因此假设在包含若干心动周期的分析窗内,心音信号具有稳定的周期性,通过主成分分析提取本征心动周期信号,通过无监督学习构建个性化的统计模型,根据模型实现窗内每一心动周期的分割。实验表明,算法的平均分割准确率比主流的LRHSMM算法高3%。

A personalized Gaussian mixture model modeling method for heart sound segmentation

Heart sound segmentation is a prerequisite for heart sound processing systems. Most methods utilize a supervised learning framework to construct the statistical model for heart sound segmentation,the major hindrancesof which are the hard work of hand-labeling training data and the mismatch between the training and testing datasets.The heart sound segmentation method based on a personalized statistical model is proposed, which does not rely on training data and enables the good match between the training and testing data. Heart sound signal is stationary in a short period. The Personalized Modeling Method(PMM) is based on an assumption that that the periodicity of heart sound signal is stationary in an analysis window which contains several cycles. The eigen signal for the cardiac cycle is extracted by making use of principal component analysis, based on which, the personalized model is constructed by an unsupervised learning framework. The heart sound signal is eventually segmented using the unsupervised model.Experiments showed that the proposed method outperforms the widely used LRHSMM by 3% in accuracy.