Clustering of Human Sleep Recordings Using a Quantile Representation of Stage Bout Durations - Biomedical Engineering Systems and Technologies - page 346

distributions [23], and the parameter values in these models have been shown to be
affected by health conditions such as chronic fatigue syndrome [21].

Scope of the Present Paper. The work reported in the present paper uses a descrip-
tion of sleep dynamics based on the durations of continuous, uninterrupted bouts in
the different sleep stages, as well as in wakefulness episodes after sleep onset. This
representation captures temporal features of sleep that are not considered by standard
sleep composition variables alone. The present paper is a revised and extended version
of [34]. In addition to the durations of bouts in various sleep stages, it would be desir-
able to account for the specific stage to which a transition occurs at the end of each stage
bout. However, the information in a full night hypnogram appears to be insufficient to
adequately model such stage transitions [4].
In the present paper, the machine learning technique of Expectation-Maximization
(EM) clustering is used to group hypnograms into families based on the distributions
of their stage bout durations. Hypnograms within each family are more similar to one
another, in terms of their bout duration statistics, than are hypnograms from different
families. The prior work [20] also uses clustering to study sleep data, but considers only
stage composition, not bout durations nor other aspects of sleep dynamics.
Each of the families found through clustering is shown to be characterized by bout
duration statistics for specific sleep stages, the values of which are shown to be statisti-
cally significantly different from those of other families at the level p< 0 . 05 ,evenafter
a suitable correction has been made for the magnification of type I error due to multiple
statistical comparisons. The Benjamini-Hochberg framework [2] is used to bound the
overall false discovery rate in a rigorous manner.
Several potentially health-related variables not involved in defining the bout duration
families, such as a sensation of muscle weakness or paralysis that occurs in emotional
situations, are also shown to differ significantly among the bout duration families iden-
tified through machine learning ( p< 0 . 05 ). This is particularly noteworthy because, in
contrast to machine learning, the widely used statistical technique of multivariate linear
regression does not provide a good predictive model of this muscle paralysis variable
based on the same bout duration variables. Our results show that machine learning can
uncover interesting dynamical patterns in sleep data, and that such patterns may be used
to predict selected aspects of individual patient health based on an all-night sleep study.

2 Methods

2.1 Human Sleep Data

Fully anonymized human polysomnographic recordings were obtained from the Sleep
Clinic at Day Kimball Hospital in Putnam, Connecticut, USA. 244 recordings were
used. In addition to the polysomnographic recordings and sleep stage information (sec-
tion 2.2), health-related patient information was available through responses to a patient
questionnaire. Summary statistics for the collection of sleep data are as in Table 1. The
acronyms that appear in the header row of Table 1 have the following meanings. BMI:
Body-Mass Index, the ratio of body weight to height-squared; ESS: Epworth Sleepiness