Successful implementation of antiretroviral treatment programs in sub-Saharan Africa has transformed HIV and AIDS from an emerging global health catastrophe to a manageable chronic condition. The HIV care cascade is a conceptual model describing key milestones in successful delivery of care and treatment for people living with HIV. These include diagnosis of HIV, linkage and retention in care, initiation of antiretroviral therapy, and suppression of viral load. Many care programs, such as AMPATH in western Kenya, have implemented electronic health records with point-of-care interface for visualizing patient records and entering data in real time. This has improved the ability of care providers to track outcomes and, where necessary, intervene to improve them.
In this talk we describe development and implementation of a Bayesian decision support module geared toward maximizing retention in care – a critically important component of the cascade. The goal of the project is to identify in advance those patients at high risk for missed visit, interruption in treatment, and loss to follow up. The project involves building and validating predictive models derived from an electronic health record system, embedding the models in the EHR back end to generate real-time predictions of missing a scheduled visit, and using the predictions to activate pre-visit outreach by clinic staff. Our model addresses several idiosyncrasies in the data, such as competing risks, discontinuous hazard functions, and missing predictors. We describe how to use the posterior predictive distribution to generate various types of insights, such as flagging patient-level features that explain risk classification and identifying optimal timing for the next appointment. We also show the implementation of the model at the point of care and describe our plans for evaluating the impact of the decision support process.

This is joint work with Arman Oganisian and Nick Lewis at Brown University, and Ann Mwangi at Moi University in Eldoret, Kenya.