Clustering of Human Sleep Recordings Using a Quantile Representation of Stage Bout Durations

In this paper, a condensed representation of stage bout durations based on the

q

-quantiles of the duration distributions is used as a basis for the discovery of duration-related patterns in human sleep data. A collection of 244 all-night hypnograms is studied. Quartiles (

q

 = 4) provide a good tradeoff between representational detail and sample variation. 15 descriptive variables are obtained that correspond to the bout duration quartiles of wake after sleep onset, NREM stage 1, NREM stage 2, slow wave sleep, and REM sleep. EM clustering is used to identify distinct groups of hypnograms based on stage bout durations. Each group is shown to be characterized by bout duration quartiles of specific sleep stages, with statistically significant differences among groups (

p

 < 0.05). Several sleep-related and health-related variables are shown to be significantly different among the bout duration groups found through clustering. In contrast, multivariate linear regression fails to yield good predictive models based on the same bout duration variables used in the clustering analysis. This work demonstrates that machine learning techniques are capable of uncovering naturally occurring dynamical patterns in sleep data that also provide sleep-based indicators of health.