Abstract
We propose and validate continuous, entirely probabilistic models of the all night sleep process.
The models are implemented as a hierarchical and a separator Gaussian Mixture Model (GMM). Features
extracted from recordings following a polysomnographic setting are used. In the study we focus on
describing sleep and transitions to sleep as a continuous process. The output of a GMM is a set of
curves representing probability of each sleep or wakefulness state at a given time point. This is
in contrast to the standard discrete Rechtschaffen & Kales (R&K) scoring system generally applied
to score sleep in human subjects. The expected outcome of the new sleep modelling effort is additional
information with respect to sleep quality, pathologies and other clinically relevant aspects not
obtained by the R&K scoring. We validate the new sleep representation through a comparison with the
R&K sleep profiles. We correlated the features extracted from both the discrete R&K and continuous
GMM sleep profiles with external criteria of sleep represented by a set of psychometric variables
collected after sleep. We demonstrate that the continuous sleep model possess the same or higher
level of information about the sleep process and can successfully complement the R&K standard for
a more comprehensive description of sleep.
Go back