Division of Sleep and Circadian Disorders, Brigham and Women's Hospital, Boston, MA, USA; Division of Sleep Medicine, Harvard Medical School, Boston, MA, USA.
University Hospital of Child and Adolescent Psychiatry and Psychotherapy, University of Bern, Bern, Switzerland; Chronobiology and Sleep Research, Institute of Pharmacology and Toxicology, University of Zurich, Zurich, Switzerland.
Sleep Med. 2018 Jul;47:126-136. doi: 10.1016/j.sleep.2017.11.1128. Epub 2017 Nov 29.
We present an automated sleep electroencephalogram (EEG) spectral analysis pipeline that includes an automated artifact detection step, and we test the hypothesis that spectral power density estimates computed with this pipeline are comparable to those computed with a commercial method preceded by visual artifact detection by a sleep expert (standard approach).
EEG data were analyzed from the C3-A2 lead in a sample of polysomnograms from 161 older women participants in a community-based cohort study. We calculated the sensitivity, specificity, accuracy, and Cohen's kappa measures from epoch-by-epoch comparisons of automated to visual-based artifact detection results; then we computed the average EEG spectral power densities in six commonly used EEG frequency bands and compared results from the two methods using correlation analysis and Bland-Altman plots.
Assessment of automated artifact detection showed high specificity [96.8%-99.4% in non-rapid eye movement (NREM), 96.9%-99.1% in rapid eye movement (REM) sleep] but low sensitivity (26.7%-38.1% in NREM, 9.1-27.4% in REM sleep). However, large artifacts (total power > 99th percentile) were removed with sensitivity up to 87.7% in NREM and 90.9% in REM, with specificities of 96.9% and 96.6%, respectively. Mean power densities computed with the two approaches for all EEG frequency bands showed very high correlation (≥0.99). The automated pipeline allowed for a 100-fold reduction in analysis time with regard to the standard approach.
Despite low sensitivity for artifact rejection, the automated pipeline generated results comparable to those obtained with a standard method that included manual artifact detection. Automated pipelines can enable practical analyses of recordings from thousands of individuals, allowing for use in genetics and epidemiological research requiring large samples.
我们提出了一个自动化的睡眠脑电图（EEG）频谱分析管道，包括一个自动化的伪影检测步骤，并测试了这样一个假设，即使用这个管道计算出的频谱功率密度估计值与经过视觉专家（标准方法）进行人工伪影检测的商业方法计算出的值相当。
我们对来自一个基于社区的队列研究的 161 名老年女性参与者的多导睡眠图的 C3-A2 导联的 EEG 数据进行了分析。我们通过对自动化与基于视觉的伪影检测结果的逐epoch 比较，计算了敏感性、特异性、准确性和 Cohen's kappa 度量；然后，我们计算了六个常用 EEG 频带的平均 EEG 频谱功率密度，并使用相关分析和 Bland-Altman 图比较了两种方法的结果。
对自动化伪影检测的评估显示出高特异性[非快速眼动（NREM）期为 96.8%-99.4%，快速眼动（REM）期为 96.9%-99.1%]，但敏感性较低（NREM 期为 26.7%-38.1%，REM 期为 9.1-27.4%）。然而，大的伪影（总功率>第 99 百分位数）的敏感性高达 87.7%在 NREM 期和 90.9%在 REM 期，特异性分别为 96.9%和 96.6%。两种方法计算的所有 EEG 频带的平均功率密度显示出非常高的相关性（≥0.99）。与标准方法相比，自动化管道的分析时间减少了 100 倍。
尽管对伪影的拒绝敏感性较低，但自动化管道生成的结果与包括手动伪影检测的标准方法相当。自动化管道可以实现对数千个人的记录进行实际分析，从而可以用于需要大样本的遗传学和流行病学研究。
