JMIR Publications

ABSTRACT

Background:

Most of the existing automated sleep staging methods rely on multimodal data and scoring a specific epoch requires not only the current epoch but also a sequence of consecutive epochs that precede and follow the epoch.

Objective:

We propose a deep learning model named SleepInceptionNet, which allows sleep classification of a single epoch using a single-channel electroencephalogram (EEG).

Methods:

SleepInceptionNet was based on our systematical evaluation of the effect of different methods of pre-processing EEG data, different EEG channels, and different available deep learning classifiers on automatic sleep staging performance. The evaluation was performed using polysomnography (PSG) data of 883 participants (937,975 thirty-second epochs). Raw data of individual EEG channels (i.e., frontal, central, and occipital) and three specific transformations of the data including power spectral density (PSD), continuous wavelet transform (CWT), and short-time Fourier transform (STFT) were separately used as the inputs of deep learning models. To classify sleep stages, seven sequential deep neural networks were tested for the 1-D data (i.e., raw EEG and PSD), and sixteen image classifier convolutional neural networks were tested for the 2-D data (i.e., CWT and STFT time-frequency images).

Results:

The best model —SleepInceptionNet that utilizes time-frequency images developed by the CWT method from central single-channel EEG data as input to the InceptionV3 image classifier algorithm— achieved a kappa agreement of 0.705±0.077 in reference to the gold standard PSG.

Conclusions:

The performance suggests that SleepInceptionNet may allow real-time automated sleep staging in free-living conditions using only single-channel EEG data and may eventually be used for on-demand intervention or treatment during specific sleep stages.