基于LSTM-CNN的睡眠呼吸暂停与低通气事件实时检测算法研究

睡眠呼吸暂停与低通气综合征(SAHS)严重影响睡眠质量, 是一种潜在的致死性呼吸疾病。为了兼顾对睡眠呼吸暂停与低通气(AH)事件检测的准确率与时间分辨率, 提出一种长短时记忆-卷积神经网络(LSTM-CNN)方法, 实现对AH事件的精准预测;同时基于事件检测结果, 提出一种呼吸紊乱指数(AHI)估计方法, 进而实现对SAHS严重程度的定量评估。选取美国国家心肺血液研究所睡眠健康数据库中54名受试者的腹部位移信号对LSTM-CNN算法进行测试。对于处理得到的超过90万数据片段, 正确率、敏感度、特异度分别为88.6%、88.2%、88.7%;54名被试的AHI预测结果与多导睡眠图(PSG)标注结果相比, 皮尔逊相关指数达到0.98;观察SAHS严重程度诊断结果, kappa系数达到0.95。结果表明, 所提出的方法不仅可以实现对AH事件的高精度检测, 而且可以对AHI指数与SAHS严重程度做出准确估计, 有望用于PSG检测之前SAHS的初步诊断以及成为家用SAHS长期监护工具。

Algorithm Study of Real-Time Detection of Sleep Apnea-Hypopnea Event Based on Long-Short Term Memory-Convolutional Neural Network

Sleep apnea and hypopnea syndrome (SAHS) is a potentially fatal disease as well as affects sleep quality. In order to balance the accuracy and time resolution of apnea and hypopnea (AH) event detection, this paper proposed a long-short term memory-convolutional neural network (LSTM-CNN) to predict AH event precisely, meanwhile an apnea-hypopnea index (AHI) estimation method based on event detection results was adopted to quantitatively assess the SAHS severity. The algorithm was tested using 54 subjects’ abdomen movement signals from National Heart Lung & Blood Institute. For over 900, 000 data fragments after preprocessing, the accuracy, sensitivity and specificity were 88.6%, 88.2% and 88.7% respectively. For the 54 subjects’ AHI, the Pearson correlation coefficient between estimated AHI and AHI scored from polysomnography (PSG) reached 0.98 and the Cohen's kappa coefficient for SAHS severity was 0.95. The results showed that this method not only realized a high-precision detection of AH event, but also accurately estimated the AHI and the severity of SAHS, therefore held the potential to be used for SAHS diagnosis before PSG or long-term monitoring of SAHS at home.