Diabetic retinopathy (DR) is a complication from diabetes that affects the eyes. According to the National Eye Institute, early detection, timely treatment, and appropriate follow-up can reduce the risk of severe vision loss from DR by 95%. Yet, DR remains the leading cause of blindness among Americans aged 20-74. In the US, an estimated 899,000 diabetic adults have vision-threatening DR (VTDR) despite it being preventable with timely treatment. VTDR is difficult to catch due to its slow progression and dependence on patients' care seeking behavior. Here I present an overview of a project which takes an end-to-end approach to this problem. Working with a care coordination company, we (1) use medical record and healthcare claims data to predict VTDR risk (2) identify barriers, motivators, and the effects of interventions at the patient and population level, and (3) apply agent-based simulation to guide care coordination intervention choices. I will go into detail on the prediction task, in which we leverage 20+ years of electronic health records to construct and extend ensemble classifiers to identify patients that will develop DR and VTDR within the next year with high recall. In practice this classifier can personalize care coordination to improve utilization and timing without any additional patient actions.