Author: Mary Stanfill, MBI (2026)
Primary advisor: Susan Fenton, PhD
Committee members: Ming Huang, PhD and William R. Hersh, MD
DHI Translational Project, McWilliams School of Biomedical Informatics at UTHealth Houston.
ABSTRACT
Population health management of patients in risk-adjusted payment plans requires new tools and techniques to control costs, enhance care, and ultimately improve patient outcomes. Effective population health management of risk-adjusted patients with complications of chronic conditions first requires identifying all the risk-adjusted patients with such complications, but common structured data sources in the electronic health record (EHR) to identify such patients are notoriously inaccurate. Thus, the mechanisms used today, based only on structured data to retrieve patients with complications of a chronic condition, are prone to problems. This makes current information retrieval (IR) approaches inadequate to support population health management that requires broader data sources to reliably identify patients who need care for chronic conditions such as microvascular complications of type II diabetes mellitus (T2DM).
This project focused on improving the identification of patients with ophthalmic microvascular complications of T2DM, an important public health focus in the United States. (Hwang et al., 2025). This project used artificial intelligence (AI) to process structured and unstructured EHR data to identify patients with ophthalmic microvascular complications of T2DM for a defined Medicare Advantage (MA) patient population. AI agents to identify diabetic retinopathy (DR) and/or diabetic macular edema (DME) were developed and improved using prompt engineering techniques and human-in-the-loop review. The AI agents were subsequently tested on real-world EHR data and evaluated for use in the population health management workflow for MA patients.
Keywords: diabetic retinopathy, population health management, risk-adjustment, unstructured data, prompt engineering, AI agents