Digital Phenotyping of Sleep Patterns Among Heterogenous Samples of Latinx Adults using Unsupervised Learning
Video
Team Information
Team Members
Ipek Ensari, Associate Research Scientist, The Data Science Institute, Columbia University
Billy Caceres, Assistant Professor, School of Nursing, Columbia University
Kasey Jackman, Assistant Professor, School of Nursing, Columbia University
Niurka Suero-Tejeda, Project Director, Assistant Professor, School of Nursing, Columbia University
Additional Team Members: Ari Shechter, Michelle Odlum, Suzanne Bakken
Abstract
This study aimed to identify homogeneous sleep sub-types (“phenotypes”) from diverse patient samples based on objectively-estimated sleep data using a flexible unsupervised machine learning technique. The study sample included 118 Latinx adults (Ages 19-77) enrolled into one of the pilot studies of Precision in Symptom Self-Management Center focusing on sleep health in individuals at increased risk for sleep disturbance. Data on total sleep time (TST), time in bed (TIB), wake after sleep onset (WASO), sleep efficiency (SE), and number of awakenings (NOA) were collected using wrist-mounted accelerometers. Cluster analysis of the sleep data (494 person-level days) was conducted using mixtures of multivariate generalized linear mixed models (MMGLMMs), which rely on the assumption that each individual is characterized by values of a set of latent random effects (i.e., predictors) and their probability of class membership (i.e., “individual component probability”; ICP) is estimated using Markov Chain Monte Carlo-based Bayesian inference. ICPs and thus the maximum likelihood of being classified in a cluster is based on the resulting posterior median probabilities. Results indicated a 3-cluster model to provide the best fit, based on the model fit indices of Penalized Expected Deviance (i.e., PEDΔ~ -20 from 2- to 3-cluster model) and model likelihood ratio (Pdiff ~ 0.92). Phenotype 1 (N=64) was the largest cluster, associated with shorter sleep time, but also shorter time in bed, and fewer nighttime awakenings, as well as a later bedtime than the other phenotypes. Phenotype 2 (N=12) was the smallest cluster consisting of all female sex participants. It was characterized by lower Sleep Efficiency, higher NOAs, and greater WASO and TIB than the other phenotypes. Phenotype 3 was characterized by a more variable sleep profile than the others, indicated by the significant t-scores associated with TIB, WASO, Sleep Efficiency. Finally, self-reported self disturbance scores did not differ across the 3 phenotypes.
These results suggest that robust digital data-driven modeling approaches can be useful for detecting sleep phenotypes from heterogenous patient populations, and have implications for designing precision sleep health strategies for management and early detection of sleep problems.
Contact this Team
Contact: Ipek Ensari (use form to send email)
