Laxminarayan Srinivas, Wang Chao, Oyama Tatsuya, Cashmere J David, Germain Anne, Reifman Jaques
Department of Defense Biotechnology High Performance Computing Software Applications Institute, Telemedicine and Advanced Technology Research Center, United States Army Medical Research and Development Command, Fort Detrick, MD, United States.
The Henry M. Jackson Foundation for the Advancement of Military Medicine, Inc., Bethesda, MD, United States.
Front Psychiatry. 2020 Oct 30;11:532623. doi: 10.3389/fpsyt.2020.532623. eCollection 2020.
Previously, we identified sleep-electroencephalography (EEG) spectral power and synchrony features that differed significantly at a population-average level between subjects with and without posttraumatic stress disorder (PTSD). Here, we aimed to examine the extent to which a combination of such features could objectively identify individual subjects with PTSD. We analyzed EEG data recorded from 78 combat-exposed Veteran men with ( = 31) and without ( = 47) PTSD during two consecutive nights of sleep. To obviate the need for manual assessment of sleep staging and facilitate extraction of features from the EEG data, for each subject, we computed 780 stage-independent, whole-night features from the 10 most commonly used EEG channels. We performed feature selection and trained a logistic regression model using a set consisting of the first 47 consecutive subjects (18 with PTSD) of the study. Then, we evaluated the model on a set consisting of the remaining 31 subjects (13 with PTSD). Feature selection yielded three uncorrelated features that were consistent across the two consecutive nights and discriminative of PTSD. One feature was from the spectral power in the delta band (2-4 Hz) and the other two were from phase synchronies in the alpha (10-12 Hz) and gamma (32-40 Hz) bands. When we combined these features into a logistic regression model to predict the subjects in the set, the trained model yielded areas under the receiver operating characteristic curve of at least 0.80. Importantly, the model yielded a -set sensitivity of 0.85 and a positive predictive value (PPV) of 0.31. We identified robust stage-independent, whole-night features from EEG signals and combined them into a logistic regression model to discriminate subjects with and without PTSD. On the set, the model yielded a high sensitivity and a PPV that was twice the prevalence rate of PTSD in the U.S. Veteran population. We conclude that, using EEG signals collected during sleep, such a model can potentially serve as a means to objectively identify U.S. Veteran men with PTSD.
此前，我们确定了创伤后应激障碍（PTSD）患者和非PTSD患者在群体平均水平上睡眠脑电图（EEG）频谱功率和同步特征存在显著差异。在此，我们旨在研究这些特征的组合能够在多大程度上客观地识别出患有PTSD的个体。我们分析了78名有战斗经历的退伍军人男性在连续两晚睡眠期间记录的EEG数据，其中31人患有PTSD，47人未患PTSD。为了避免手动评估睡眠分期的需要并便于从EEG数据中提取特征，对于每个受试者，我们从10个最常用的EEG通道计算了780个与睡眠阶段无关的全夜特征。我们进行了特征选择，并使用该研究中前47名连续受试者（18名患有PTSD）组成的数据集训练了一个逻辑回归模型。然后，我们在由其余31名受试者（13名患有PTSD）组成的测试集上评估该模型。特征选择产生了三个不相关的特征，这些特征在连续两晚是一致的，并且对PTSD具有鉴别性。一个特征来自δ频段（2 - 4Hz）的频谱功率，另外两个特征来自α频段（10 - 12Hz）和γ频段（32 - 40Hz）的相位同步。当我们将这些特征组合到一个逻辑回归模型中以预测测试集中的受试者时，训练后的模型在受试者工作特征曲线下的面积至少为0.80。重要的是，该模型在测试集上的敏感性为0.85，阳性预测值（PPV）为0.31。我们从EEG信号中识别出了稳健的与睡眠阶段无关的全夜特征，并将它们组合到一个逻辑回归模型中以区分患有和未患有PTSD的受试者。在测试集上，该模型产生了高敏感性和PPV，其PPV是美国退伍军人人群中PTSD患病率的两倍。我们得出结论，使用睡眠期间收集的EEG信号，这样的模型有可能作为一种客观识别患有PTSD的美国退伍军人男性的手段。
