The SIESTA Analyzer

Design

The SIESTA sleep analyzer in it's current implementation maps all-night EEG recordings to 3 probability plots. The probabilities of being awake, in deep sleep or in REM sleep are modeled with temporal resolutions down to one second. It’s entirely probabilistic approach offers the advantage that all uncertainties (e.g. from artifact contamination of the EEG recording) will lead to less certain decisions about the state of the sleeping subject.

As the result of extensive investigations with different feature subset selection techniques, complexity measures and autoregressive (AR) model parameters were preferred as opposed to classical FFT based features.

simple classifiers (experts) are built on top of each EEG electrode (thus, the curse of dimensionality of fitting one large model, is avoided).
Fusing multiple experts increase the reliability of the overall system, because information from such segments where the experts disagree will be downweighted.
The architecture can be extended very efficiently.

Here is an example of the analyzer's output:

Legend (from top to bottom):
R&K hypnogram (by human experts), proability plots for wake, slow-wave-sleep and REM, continuous spindle density plot.

find more information about the analyzer in:

Sykacek P., Roberts S., Rezek I., Flexer A., Dorffner G.: A Probabilistic Approach to High-Resolution Sleep Analysis, in Dorffner G., et al.(eds.), Artificial Neural Networks - ICANN 2001, International Conference, Vienna, Austria, Lecture Notes In Computer Science 2130, Springer, pp. 617-624, 2001.[available online at http://www.robots.ox.ac.uk/~sjrob/Pubs/sleep_icann01.ps.gz]