摘要
目的通过识别正常肺音、哮鸣音、捻发音和爆裂音4类肺音,将肺音和肺部疾病关联起来,预测每类肺音对应的呼吸疾病。方法将电子听诊器采集的正常和异常肺音经滤波和周期分段预处理,再采用韦尔奇功率谱估计和小波变换得到肺音信号统计特征值。比较神经网络和遗传神经网络两类分类器的性能,选择遗传神经网络识别算法进行肺音的识别。结果采用韦尔奇功率谱特征值的遗传BP神经网络平均识别率89.0%,优于BP神经网络的平均识别率(83.0%);用小波系数特征值的遗传BP神经网络平均识别率83.1%,优于BP神经网络的平均识别率(81.0%)。结论韦尔奇功率谱的特征提取方法有效,能较准确区分出肺音的类别。
Objective To predict the respiratory disease category by identifying 4 kinds of lung sounds including the normal lung sounds, wheeze, crepitants and crackles, which are related to lung diseases. Methods The normal and abnormal lung sounds collected by electronic stethoscope were filtered and periodically pretreated by sets. Then the Welch power spectrum estimation and wavelet transform method were adopted to obtain the statistical characteristics of the lung sound signals. After comparing the performance of neural network and ge- netic neural network, the genetic neural network identification algorithm was selected to identify the lung sounds. Results The average recognition rate of Welch power spectrum eigenvalues BP neural network was 89.0%, which was better than the BP neural network whose average recognition rate was 83.0%. While, the average recognition rate of wavelet coefficients eigenvalues BP neural network was 83.1%, which was also bet- ter than the BP neural network average recognition rate which was 81.0%. Conclusion Welch power spectrum feature extraction method is effective and can accurately identify the different lung sound categories.
出处
《航天医学与医学工程》
CAS
CSCD
北大核心
2016年第1期45-51,共7页
Space Medicine & Medical Engineering
基金
国家自然科学基金(31200709)
关键词
肺音
小波去噪
模式识别
遗传BP神经网络
lung sound
wavelet denoising
pattern identification
genetic BP neural network