Theses and Dissertations from DePaul University

Date of Award

Summer 2026

Degree Type

Thesis

Degree Name

Master of Science (MS)

College

College of Computing and Digital Media

First Advisor

Roselyne Tchoua

Abstract

Emergency Department (ED) readmissions remain a major challenge for healthcare systems, affecting both patient care quality and financial costs. Most prediction models depend largely on structured data from clinical tools that assign points to a small set of predefined factors – such as comorbidities, previous hospital visits, and behaviors such as smoking and drinking – and then sum those points to produce an overall risk score. These point-based tools leave out important information from Social Determinants of Health and a wealth of information from Community Health Workers (Community health workers (CHWs), trusted members of a community who help connect people to healthcare and social services. Our overall goal is to empower CHWs to interact with and interpret the data leading to improved ED readmission rates. As CHWs take notes and fill out surveys to improve patients’ health outcomes there is currently no way to see what parts of the notes have the highest impact towards predicting readmissions. A dashboard is being implemented to show the probability of being readmitted. As part of the responsible design of this dashboard, the aim is to explain and display the reasons for specific predictions. This work contributes to the overall goal by showing what words within the CHW notes have the highest impact on the prediction and in which direction. Prior work was done using Natural Language Models (NLMs) for text and Random Forest for the (structured) tabular data. Using explainable AI (XAI) methods (specifically SHAP and LIME), this thesis explores which words and features have the largest impact on the patient and explores ways to help predict what parts of the notes and survey give these predictions. Importantly, the work not only asks what model is more accurate but also what is most useful and explainable. Based on all modeling the use of the original notes with no additional preprocessing provides the best model to use and easiest to see with the XAI methods. This work shows that SHAP is the best for global word usage and allows management to have an idea of what is beneficial and why good notes are so helpful; at the same time LIME is best used for local views by CHWs to see what words have the highest impact and give these a score of impact, all color coded. The findings are summarized in insights from and recommendations to CHWs. Any improvement that allows CHWs to better understand how their notes impact the model, ultimately helps give better guidance to the patient. This, in turn, will impact their ability to help prevent readmittance and thus also reduce patient and hospital costs.

Download

Share

COinS