中山大学学报自然科学版 ›› 2018, Vol. 57 ›› Issue (3): 128-134.

• 论文 • 上一篇    下一篇

基于心率变异性与机器学习的睡眠呼吸事件分类

梁九兴1,张湘民2,黄少雄1,曾令紫1,罗语溪1   

  1. 1. 中山大学工学院,广东 广州 510006;
    2. 中山大学附属第六医院, 广东 广州 510655
  • 收稿日期:2017-08-17 出版日期:2018-05-25 发布日期:2018-05-25
  • 通讯作者: 罗语溪(1983年生),男;研究方向:医疗设备;E-mail:luoyuc@163.com

Classification of sleep respiratory events based on heart rate variability and machine learning

LIANG Jiuxing1, ZHANG Xiangmin2, HUANG Shaoxiong1, ZENG Lingzi1, LUO Yuxi1   

  1. 1. School of Engineering, Sun Yat-sen University, Guangzhou 510006, China;
    2. The Sixth Affiliated Hospital of Sun Yat-sen University, Guangzhou 510655, China
  • Received:2017-08-17 Online:2018-05-25 Published:2018-05-25

摘要:

为了提供一种针对睡眠呼吸暂停低通气综合征(sleep apnea hypopnea syndrome, SAHS)患者的筛查方法,本研究把心率变异性(heart rate variability, HRV)应用于睡眠呼吸模态的分类问题。通过构建和训练概率神经网络(probabilistic neural network, PNN)对HRV各特征值进行有无异常睡眠呼吸事件的判别,以期实现对该病征进行初步筛查的目的。首先,对标注的有无呼吸事件的多导睡眠监测数据提取其心电的HRV特征值,再经过归一化后作为特征向量;其次采用PNN分类算法对特征向量进行训练及分类输出;最后,对模型的预测分类性能进行评价。对于准确率、灵敏度、特异性、受试者工作特性曲线下面积(area under the receiver operating characteristic curve, AUC)及分类耗时等评价指标PNN分类器的结果分别为:75.97%,82.51%,76.22%,0.7936,0.63 s。与广义回归神经网络(generalized regression neural network, GRNN)及极限学习机(extreme learning machine, ELM)分类算法相比,PNN分类算法在灵敏度、特异性、AUC及分类耗时评价维度上均取得最优。本研究基于HRV及PNN分类算法实现了对有无异常睡眠呼吸事件的判别,提供了一种低生理负荷SAHS筛查的途径。

关键词: 心率变异性, 睡眠呼吸事件, 机器学习, 分类算法

Abstract:

To provide a method for screening patients with sleep apnea hypopnea syndrome (SAHS), the heart rate variability (HRV) was applied to the classification of sleep respiratory events. The probabilistic neural network (PNN) was proposed to classify the normal and abnormal sleep respiratory events according to the HRV features to achieve the purpose of preliminary screening of the disease. In this classification process, the HRV features of ECG were firstly extracted from the polysomnographic monitoring data related to the normal and abnormal sleep respiratory events, and then normalized as the features. Then, PNN classification algorithm was used to train and classify the features. The prediction and classification performance of the model was finally evaluated. The results of the PNN classifier for the accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC) of the subjects and time consumption for classification were respectively: 75.97%, 82.51%, 76.22%, 0.7936 and 0.63 s. Compared with generalized regression neural network (GRNN) and extreme learning machine (ELM) classification algorithms, PNN classification algorithm is optimal in sensitivity, specificity, AUC and time-consumptions. In this study, HRV and PNN classification algorithm were used to classify the presence or absence of abnormal sleep respiratory events, thus providing a low physiological load SAHS screening method. The study has a certain practical significance for the initial screening of the disease.

Key words: heart rate variability, sleep respiratory event, machine learning, classification algorithm

中图分类号: