Machine Learning Approches to Predict Blood Pressure Level and Variability Using Polysomnography Data

Blood pressure level has been studied as a medium between sleep apnea and cardiovascular disease because previous studies indicate that sleep apnea is associated with high blood pressure (BP) and high visit-to-visit blood pressure variability (BPV), both of which are risk factors of certain cardiovascular diseases. However, a limited number of studies have been conducted to predict high BP and visit-to-visit BPV using in-lab sleep study data, and the conclusions obtained in those studies are conflicting. This study’s objective is to develop a predictive model for BP and BPV using machine learning methods based on in-lab sleep study data collected from the Sleep Disorders Center (Sleep Laboratory) at the University of Virginia (UVA). After pre-processing and combining the sleep data with other available patient data, including demographic and clinical information, a multi-step feature selection procedure was applied, resulting in eighteen variables for the next step. Furthermore, various machine learning models were developed, and their performances were compared. A multiple imputation method for dealing with missing data and feature reduction methods was studied during the process of developing models. The results indicate that this in-lab sleep study data can be employed to build a promising classification model for the high-BP group (systolic BP 130 mm Hg), although it is unable to predict high visit-to-visit BPV. The model's feature importance indicates that sleep-related features and blood oxygen saturation (SpO2)-related features are associated with high BP.