Predicting Periods of Hypertension for Patients Using Remote Monitoring Data

Abstract

Hypertension is directly linked with an increase in mortality risks. This increase is especially observed in rural areas with uninsured or under-insured patients. Remote patient monitoring is a technological development that can help monitor hypertensive patients and detect changes within their blood pressure levels. The technological limitation is the inability to predict periods of hypertension. In this thesis, we address this issue by presenting a software system designed to successfully predict periods of hypertension and alert clinicians. Central to our software system is a framework for preparing data for predictive machine learning models, and an adaptive model for different sized sliding windows, to enable adverse health event predictions before the optimal in-put length has been reached. Using this framework, an XGBoost model only tuned for unbalanced data achieved an area under the receiver operating characteristic curve score of 0.77 and area under the precision recall curve score of 0.76. Feature importance plots demonstrate that our framework can extract the most impactful features, which are common across models. This demonstrates the ability to transform previously unsuitable data into data well-suited for periods of hypertension prediction and use it to alert clinicians of hypertensive periods to facilitate early interventions.