Robert C, Karasinski P, Natowicz R, Limoge A
Laboratoire d'Electrophysiologie, Université René Descartes Paris V, Montrouge, France.
Physiol Behav. 1996 Jun;59(6):1051-60. doi: 10.1016/0031-9384(95)02214-7.
Two multilayer neural networks were designed to discriminate vigilance states (waking, paradoxical sleep, and non-REM sleep) in the rat using a single parieto-occipital EEG derivation. After filtering (bandwidth 3.18-25 Hz) and digitization at 512 HZ, the EEG signal was segmented into eight second epochs. Five variables (three statistical, two temporal) were extracted from each epoch. The first network computed an epoch by epoch classification, while the second network also utilized contextual information from contiguous epochs. A specific postprocessing procedure was developed to enhance the vigilance state discrimination of the neural networks designed and especially paradoxical sleep state estimation. The classifications made by the networks (with or without the postprocessing procedure) for six rats were compared to these made by two human experts using EMG and EEG informations on 63,000 epochs. High rates of agreement (> 90%) between humans and neural networks classifications were obtained. In view of its development possibilities and its applicability to other signals, this method could prove of value in biomedical research.
设计了两个多层神经网络，用于使用大鼠单个顶枕脑电图衍生数据来区分其警觉状态（清醒、异相睡眠和非快速眼动睡眠）。在进行带宽为3.18 - 25赫兹的滤波以及512赫兹的数字化处理后，脑电图信号被分割为8秒的时段。从每个时段中提取了五个变量（三个统计变量、两个时间变量）。第一个网络逐时段进行分类计算，而第二个网络还利用相邻时段的上下文信息。开发了一种特定的后处理程序，以增强所设计神经网络的警觉状态辨别能力，特别是异相睡眠状态估计。将六个大鼠的网络分类结果（有无后处理程序）与两位人类专家利用肌电图和脑电图信息对63000个时段所做的分类结果进行了比较。在人类和神经网络分类之间获得了较高的一致率（> 90%）。鉴于其发展潜力及其对其他信号的适用性，该方法在生物医学研究中可能具有价值。
