Analytics Approaches to the Development of Diabetic Retinopathy Screening Policies

Diabetic retinopathy (DR) is the leading cause of blindness for working-age adults in the US. Over 60% of patients with type II diabetes and 90% of patients with type I diabetes develop DR within 20 years of diagnosis. Routine comprehensive screening examinations have proved effective in detecting early stages of DR and timely treatment can prevent up to 98% of DR-related vision loss. However, only 50-60% of diabetic patients adhere to the current annual screening guidelines. Recently, teleretinal imaging (TRI) has emerged as an accessible screening tool for patients with limited access. However, there exists no well-established guideline that incorporates TRI-based screening for such patients. In this thesis, we study a multi-pronged analytics approach to quantify and evaluate the advantages and limitations of TRI compared with traditional clinic-based screening (CS) and propose new screening policies for patients with limited access to eye care. First, we develop a simulation model that examines the health and cost benefits of various routine CS and TRI-based DR screening policies at different time intervals for various types of diabetic patients. Additionally, we identify patient subgroups who would truly benefit from TRI in terms of health benefits and cost savings. Second, we develop a partially observable Markov decision process (POMDP) model to generate personalized DR screening recommendations that exploit the dynamic interaction of TRI and traditional screening based on each patient’s unique health-related and behavioral factors. Lastly, we develop a decision tree model that establishes interpretable DR screening policies by transforming the complex, POMDP-driven personalized screening policies into policies that are more explainable, implementable, and adoptable in clinical practice.