Authors:
Shirin Riazy
1
;
Tilo Wendler
1
;
Jürgen Pilz
2
;
M. Glos
3
and
T. Penzel
3
Affiliations:
1
Hochschule für Technik und Wirtschaft, Germany
;
2
Alpen-Adria Universität Klagenfurt, Austria
;
3
Charité - Universitätsmedizin Berlin, Germany
Keyword(s):
Automatic Sleep Staging, Two-channel Measurement, Bayesian Statistics, Hidden Markov Model, MAP.
Abstract:
In this paper, we shall introduce an algorithm that classifies EEG data into five sleep stages, relying only on two-channel sleep measurements. The sleep of a patient (divided into intervals of 30 seconds) is assumed to be a Markov chain on the five-element state space of sleep stages and our aim is to compute the most probable chain of this hidden Markov model by a maximum a posteriori (MAP) estimation in the Bayesian framework. Both the prior distribution of the chains and the likelihood model have to be trained on manual classifications
made by professionals. For this purpose, the data is first preprocessed by a Fourier transform, a log transform and a principal component analysis for dimensionality reduction. Since the number of possible chains is immense (roughly 10^335), a heuristic approach for the computation of the MAP estimator is introduced, that systematically discards unlikely chains. The sleep stage classification is then compared to the classification of a professional
, who scores according to the AASM and uses a full polysomnography. The overall structure of the hypnogram can adequately be reconstructed with error rates around 25%.
(More)