Recognition of wake-sleep stage 1 multichannel eeg patterns using spectral entropy features for drowsiness detection

• Published in: Australasian physical & engineering sciences in medicine Vol. 39; no. 3; pp. 797 - 806
• Main Authors: Sriraam, N., Padma Shri, T. K., Maheshwari, Uma
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Netherlands 01.09.2016; Springer Nature B.V
• Subjects: Algorithms; Biological and Medical Physics; Biomedical and Life Sciences; Biomedical Engineering and Bioengineering; Biomedicine; Biophysics; Databases as Topic; Electroencephalography; Electroencephalography - methods; Entropy; Humans; Medical and Radiation Physics; Neural Networks (Computer); Neurobiology; Sleep; Sleep - physiology; Sleep Stages - physiology; Spectrum analysis; Technical Paper; Wakefulness - physiology; Spectral entropy (SE); Electroencephalogram (EEG); Back propagation (BP) algorithm; Multilayer perceptron-feed forward (MLP-FF) neural network; Polysomnograms (PSG)
• ISSN: 0158-9938; 1879-5447; 1879-5447
• DOI: 10.1007/s13246-016-0472-8

Electroencephalographic (EEG) activity recorded during the entire sleep cycle reflects various complex processes associated with brain and exhibits a high degree of irregularity through various stages of sleep. The identification of transition from wakefulness to stage1 sleep is a challenging area of research for the biomedical community. In this paper, spectral entropy (SE) is used as a complexity measure to quantify irregularities in awake and stage1 sleep of 8-channel sleep EEG data from the polysomnographic recordings of ten healthy subjects. The SE measures of awake and stage1 sleep EEG data are estimated for each second and applied to a multilayer perceptron feed forward neural network (MLP-FF). The network is trained using back propagation algorithm for recognizing these two patterns. Initially, the MLP network is trained and tested for randomly chosen subject-wise combined datasets I and II and then for the combined large dataset III. In all cases, 60 % of the entire dataset is used for training while 20 % is used for testing and 20 % for validation. Results indicate that the MLP neural network learns with maximum testing accuracy of 95.9 % for dataset II. In the case of combined large dataset, the network performs with a maximum accuracy of 99.2 % with 100 hidden neurons. Results show that in channels O1, O2, F3 and F4 (A1, A2 as reference), the mean of the spectral entropy value is higher in awake state than in stage1 sleep indicating that the EEG becomes more regular and rhythmic as the subject attains stage1 sleep from wakefulness. However, in C3 and C4 the mean values of SE values are not very much discriminative of both groups. This may prove to be a very effective indicator for scoring the first two stages of sleep EEG and may be used to detect the transition from wakefulness to stage1 sleep.