ECE PhD Prospectus Defense: Yang Hu

Title: Data Driven Models In Health Care Management

Presenter: Yang Hu

Advisor: Professor Yannis Paschalidis, ECE

Chair: Professor Ayse Coskun, ECE

Committee: Professor Christos Cassandras, ECE; Professor Ayse Coskun, ECE; Dr. Rebecca G. Mishuris, BMC

Abstract: The growing role of machine learning in scientific and biomedical applications has facilitated the advancement of health and medical informatics. Machine learning-based predictive models have shown great promise in identifying important features and enabling early-stage treatment. This work investigates data-driven techniques to predict health-related events and elucidates important predictive variables; in some instances, revealing racial bias implied by the predictive models. This report considers three problems. First, identifying patients with not well-controlled hypertension by an easily implementable and interpretable predictive model of very high blood pressure among hypertensive patients, based only on demographic features and socioeconomic factors. Age, race, Social Determinants of Health (SDoH), mental health, marital status, cigarette use, and gender were found to be predictive of high SBP. Being Black or needing help with food, transportation, housing, and employment led to higher probability of poorly controlled SBP. The second problem considered concerns personalized hypertension management. A prediction-based prescriptive model is developed to analyze outcomes under each possible medication. The model incorporates robust regularized regression technique based on Distributionally Robust Optimization (DRO) on an ambiguity set constructed from the Wasserstein metric. Then K-Nearest Neighbors (K-NN) regression is applied to capture the nonlinear similarity between patients’ most predictive characteristics and model the future Systolic Blood Pressure (SBP) under each prescription. Our approach leads to a reduction of 14.28 mmHg in SBP, on average, which is 70.30% larger than the reduction achieved by the standard-of-care and 7.08% better than the outcome of the 2nd best alternative model using Ordinary Least Squares regression. The third project focuses on predicting important events in the treatment of COVID-19. Linear and nonlinear classification methods were applied to predict the following outcomes: 1) hospitalization, 2) ICU care, 3) mechanical ventilation, and 4) mortality. A timeline strategy was introduced to capture the dynamic evolution of vital signs, labs, and radiological findings for predicting ICU, mechanical ventilation, and mortality. A score based on a recurrent neural network and a transformer architecture was developed to capture the dynamic evolution of vital signs over the timeline. Hospitalizations can be predicted with an Area Under the ROC Curve (AUC) of 92% using symptoms, patient characteristics, and hospital occupancy. Parsimonious models to predict intensive care, mechanical ventilation, and mortality that used the most recent labs and vitals exhibited AUCs of 92.7%, 91.2%, and 94%, respectively.