Modeling and Discovery of Patterns in Human Sleep

This research project aims to develop machine learning techniques for modeling and automated discovery of meaningful patterns in human sleep data. In recent work in this direction, we developed a deep convolutional neural network approach to sleep stage classification with interpretation of the emergent internal features (Sokolovsky et al, 2018; Parasrisomsuk et al, 2018). In prior work, we demonstrated discovery of naturally occurring subgroups of sleep studies, or "sleep types", based on the stage composition of sleep (Khasawneh et al, 2010 and 2011). Our earlier work yielded an association mining approach for exploratory analysis of sleep data (Laxminarayan et al, 2005), including tight bounds on the false discovery rate (Laxminarayan et al, 2006), and construction of a terabyte-scale database of anonymized polysomnographic recordings and health histories.

Sleep Dynamics

Our work has also included modeling of the dynamics of sleep. An initial step in this direction is presented in (Usher et al, 2012). A major challenge to dynamical modeling of sleep is the scarcity of dynamical stage-transition events within a single all-night sleep recording for a given person. The paper (Alvarez and Ruiz, 2013) presents a general E-M framework for dynamical modeling of time series data (sleep in particular) that contain limited dynamical information. Subsequent work (Wang et al, 2014) applied this general approach to sleep using partially observable Markov models as the class of dynamical models.

Selected publications (asterisks denote student co-authors)

Project Personnel


Current Students

Former group members