Journal Article

1090 Deep Neural Network For Automatic And Causal Sleep Staging Based On A Single Eeg Channel Recorded At Home

E Bresch, U Grossekathofer and G Garcia-Molina


Published on behalf of American Academy of Sleep Medicine

Volume 41, issue suppl_1, pages A405-A405
ISSN: 0161-8105
Published online April 2018 | e-ISSN: 1550-9109 | DOI:

More Like This

Show all results sharing these subjects:

  • Neurology
  • Sleep Medicine
  • Clinical Neuroscience
  • Neuroscience


Show Summary Details




Consumer EEG sleep devices only use a few electrodes due to cost/space/comfort. This motivates research on single EEG based automatic sleep staging. Causal staging, in addition, enables real-time applications. Deep neural networks (DNN) can discover optimal features leading to high accuracy provided that sufficient data are available for training. We leverage DNN method in this research.


Dataset: 147 sleep recordings collected at home on 29 healthy volunteers (18F; 37 ± 6.8 yrs.; ~5 recordings/participant) using a wearable prototype that records frontal EEG and ocular signals. EEG and ocular signals were used for manual (reference), and only the EEG for automatic staging. DNN architecture: 1) a stack of convolutional layers (CONV) that processes 30-second long windows and acts as filter bank connected to 2) a stack of long short term memory layers (LSTM) that model temporal structure. The output layer produces five probabilities associated with Wake, N1, N2, N3, and REM. The highest probability determines the stage. Performance was evaluated using the Kappa statistic (κ) with manual staging as reference. The average κ characterizing inter sleep-technician agreement on full PSG is 0.75. We tested the effect on κ of the amount of training data, regularization method (norm-1, norm-2, dropout), architecture (number of CONV/LSTM layers, and LSTM units), and demographic information. Reported κ values are the average of a three-fold cross-validation (CV) procedure.


As the number of subjects in the training set increases from 3 to 19, κ increases from 0.63 to 0.72. Neither regularization method increased κ. Increasing the number of CONV layers from 0 to 3, increases κ from 0.34 to 0.72. Using 16 units in LSTM layers increased κ to 0.73. No significant change resulted from considering demographics.


The optimal DNN architecture requires serially connected CONV and LSTM layers to ensure a performance (robust to age/gender sleep variability) similar to inter sleep-technician agreement with a single EEG signal recorded at home.

Support (If Any)


Journal Article.  0 words. 

Subjects: Neurology ; Sleep Medicine ; Clinical Neuroscience ; Neuroscience

Users without a subscription are not able to see the full content. Please, subscribe or login to access all content. subscribe or login to access all content.