Introduction:
Alternatively to analyzing EEG frequency bands, sleep stages can also be separated
using momentary signal amplitudes of the different recording channels. Based on this
finding, we present a real time application that estimates the sleep stages by using
measured amplitude vectors.
Methods:
The sleep stage specific distributions of amplitude vectors correspond to clusters
in a 3-dimensional space. From these clusters we derive continuous probability distributions
by summing multivariate Gaussian distributions. With Bayesian statistics we compute
the current probabilities for all sleep stages from these probability distributions
of measured data. The more amplitude vectors are measured the more exact the classification
becomes, even for highly overlapping clusters of different sleep stages. Sudden changes
of sleep stage can be detected by multiplication with a Markov matrix.
Results:
In order to validate the method we measured three EEG channels (F4, C4, O2) from 40
subjects. A part of the 3-dimensional amplitude vectors served as training data set
for the algorithm. Subsequently, all amplitude vectors were used as test data set.
A comparison with manual classification resulted in an accuracy of more than 90%.
Conclusions:
In principle, using our method we can estimate the momentary sleep stages fully automated,
in real time and with high accuracy. In future, the performance of our method could
be further increased by recording from more than three EEG channels.
