Phenotyping obstructive sleep apnea patients and prognosis of their comorbidities through machine learning methods

Abstract

Obstructive sleep apnea (OSA) is conventionally diagnosed with a single metric, the apnea-hypopnea index (AHI). However, the AHI may not be the sleep characteristic most relevant to the associated comorbidities of OSA, which include cardiovascular, neurovascular, and metabolic disease. In this thesis, by using machines learning methods, we aim to discover comprehensive phenotypes of OSA based on whole polysomnography (PSG) scores to identify the groups of patients with prevalent comorbidity outcomes and the sleep characteristics of such groups. PSG data of patients that underwent the PSG test at a sleep center in a tertiary hospital in South Korea from 2004 to 2017 were used. We use the Dirichlet process Gaussian mixture models (DPGMM) to cluster the patients, where six clusters with different sleep characteristics and cardio-neuro-metabolic outcomes were discovered. In addition, to further investigate the relationship between the PSG variables and comorbidity prevalence, we use the random survival forest (RSF) to discover sleep characteristics that have high contributions in the prevalence of comorbidities.