Utilizing machine learning-based cluster analysis for sleep phenotypes identification in COPD patients

COPD patients with disturbed sleep have lower quality of life and are at risk of adverse outcomes. The goal of this study is to identify distinct phenotypic clusters of COPD patients using objective sleep parameters and evaluate the relationship between these groups and all-cause mortality to inform risk stratification. We performed a longitudinal observational cohort study on data from Veterans Health Administration. In patients with COPD, sleep parameters were extracted from polysomnography physician interpretation using a validated natural language processing algorithm. We employed cluster analysis using an unsupervised machine learning algorithm (K-means) and examined the association between clusters and mortality using Cox regression analysis. All the results were adjusted for potential confounders and visualized with Kaplan-Meier estimates. We discovered five distinct clusters in a cohort of 9,992 patients with COPD and a clinically indicated baseline polysomnogram, based on age, comorbidity burden and sleep parameters. Overall mortality and short-term mortality (< 5.3 years) increased from 9.4% and 3.4% in Cluster 1 to 42% and 24.3% in Cluster 5. Younger age (in Cluster 1), comorbidity burden (in Cluster 5) and total sleep time and sleep efficiency (in the other three clusters) showed significant association with mortality. Our data showed that machine learning is an effective tool to identifying COPD phenotypes. Total sleep time and sleep efficiency correlated with mortality in COPD. Five clusters showed the value of objective sleep parameters in predicting the mortality and phenotypic characterization in COPD population.