Date of Award

May 2021

Degree Type

Dissertation

Degree Name

Doctor of Philosophy

Department

Management Science

First Advisor

Huimin Zhao

Committee Members

Abhijeet Ghoshal, Sanjoy Ghose, Yang Wang

Keywords

Machine Learning, Predictive Modeling, Readmission Analytics, Readmission Costs

Abstract

The healthcare cost has continued to increase over the past few years despite various policies, efforts, and initiatives taken by the government. It is still projected to grow over the next few years by the Centers for Medicare and Medicaid Services (CMS). Readmissions have been a major contributor to the increase in costs and have always been a contributing factor. To get a perspective, considering the fact that at least 9% of individuals who had COVID-19 were likely to get readmitted shortly, according to a study by the Centers for Disease Control and Prevention (CDC) COVID-19 response team, along with their high estimated treatment cost, the problem of high healthcare costs will continue to grow. The implementation of the American Recovery and Reinvestment Act of 2009 has led to massive increase in digital health data facilitating various studies to utilize analytics to improve healthcare. The goal of the two essays in this dissertation is to address the identified research gaps in the literature in readmission analytics.In Essay 1, I deploy the term readmission in two different ways and then focus on building and identifying predictive models that are suitable for costs billed by hospitals for the identified readmission categories. By using a data-driven approach, my initial analysis revealed that 21% of readmitted individuals (regardless of the number of days to readmission) alone contributed to 48% of the healthcare cost. Apart from that, my analysis revealed that the readmission cost (for the identified readmission categories in this study) varied from the previous admission cost at both individual and aggregated levels. Deep learning-based models performed the best for all scenarios. In Essay 2, I focus on creating a multitask learning-based joint model for predicting different outcomes related to readmissions, namely, likelihood, cost, and length of stay. I then evaluate the performance of the joint model and analyze its usefulness. Analysis was done for the identified top three categories of readmission belonging to the same major diagnostic groups from Essay1. Results showed that the joint model performed slightly better than the single-task baseline model for specific scenarios. The joint model was also beneficial in determining predictors that were consistently important to predict all the outcomes related to readmissions regardless of not giving us a universally best model.

Download

Share

COinS