Predicting Lost-to-Follow-Up among ART Clients: Proposal for the Application of Machine Learning

Because these techniques capture complex relationships among variables, capturing nonlinear interactions among hundreds or even thousands of inputs, they align well with the extremely complex factors that lead to loss to follow-up (LTFU). The output from ML algorithms can then help HIV programs rank clients in terms of their risk for LTFU and prioritize those most likely to default without intensive intervention. The implementation strategy for predicting LTFU among ART clients involves several stages, including data collection and feature generation, model training and testing, model application, and model deployment. Data collection and feature generation involve acquiring, cleaning, and exploring data, as well as converting raw data inputs into variables that yield better insights into the outcome being modeled. Model training and testing involve using historical data to optimize models for both predictive accuracy on historical data and generalizability to new data. The Connected Health AI Network (CHAIN) technology, developed by macro-eyes, a partner of Data for Implementation (Data.FI), has been employed for the accurate prediction of client adherence to schedules in various scenarios, including routine health appointments in the United States and early childhood immunizations in Tanzania. The output from the ML algorithms developed for this work helps programs to rank and prioritize which patients should be targeted for more intensive intervention before the patient defaults. A case study in Mozambique and Nigeria demonstrated the effectiveness of ML in predicting LTFU among ART clients. The study used data from USAID-supported programs in these countries and applied ML algorithms to identify clients at risk of becoming lost to follow-up. The results showed that the ML approach was able to accurately predict LTFU and identify clients who required more intensive intervention. The implementation strategy for predicting LTFU among ART clients also involves integrating with local health systems, securing data privacy, providing technical assistance to accelerate future ML projects, and meeting ethical and other regulatory standards. The strategy presents the potential for both cost and program performance impact, clarifying the value proposition at hand. The implementation strategy aligns with and reflects frameworks and considerations laid out in the USAID Center for Innovation and Impact report, "Artificial Intelligence in Global Health: Defining a Collective Path Forward." The ML approach to predicting LTFU among ART clients has several benefits over traditional approaches. It is able to capture complex relationships among variables and identify nonlinear interactions among hundreds or even thousands of inputs. This allows for more accurate predictions and better identification of clients at risk of becoming lost to follow-up. Additionally, the ML approach can be replicated and tailored to data on a wide variety of client adherence scenarios, making it a versatile and effective tool for improving ART client adherence. In terms of systems integration, the ML approach to predicting LTFU among ART clients involves several components, including integrating with local health systems, decentralized deployment, centralized 'push' deployment, centralized 'pull' deployment, workflow integration, monitoring model performance, technical assistance, data privacy, regulatory standards, cost, and program performance impact. The implementation strategy presents the potential for both cost and program performance impact, clarifying the value proposition at hand.