Background
Viral load suppression (VLS) is crucial for controlling HIV transmission and improving health outcomes in people living with HIV. While PEPFAR-supported programs have made substantial progress in improving VLS, innovative approaches are needed to sustain and enhance VLS in resource-limited settings. This analysis integrates machine learning and the difference-in-difference (DiD) method to forecast future VLS and assess the impact of programmatic efforts across 21 PEPFAR-supported countries in sub-Saharan Africa.
Methods
Historical aggregate PEPFAR program data from 2017 to 2024 were analyzed to identify key factors associated with VLS improvement. A program indicator matrix, aligned with PEPFAR’s Monitoring, Evaluation, and Reporting (MER) guidance, was computed for key HIV metrics. This matrix was integrated with socio-demographic, socioeconomic, and health indicators from the World Bank’s World Development Indicators to create a comprehensive analytical dataset. A DiD analysis was conducted to compare changes in VLS between 2017 and 2024. An XGBoost machine learning algorithm was used to develop a time-series model forecasting VLS from 2025 to 2030, tuned and validated using five-fold cross-validation with root mean square error (RMSE) as the primary performance metric.
Results
The XGBoost model exhibited strong predictive accuracy when validated against actual data from 2017-2024, achieving an RMSE of 1.61 on the training set and 2.30 on the testing set. This small difference indicates minimal overfitting and reliable generalization to unseen data. The model identified key predictors of VLS, including viral load coverage, new enrollments in PrEP, and age demographics. Notably, the 15-64 age group had a more significant impact on VLS outcomes compared to the pediatric population (ages 0-14). The DiD analysis revealed a statistically significant increase in VLS from 2017 (82.93%) to 2024 (95.59%) across all 21 countries, with an overall estimated 15.27% increase (p=0.0016). The forecast suggests a continued upward trend in VLS through 2030, contingent on sustained programmatic interventions identified.
Conclusions
This analysis demonstrates the potential of machine learning to identify targeted HIV interventions that can be modified to improve program performance. It underscores the need for innovative, data-driven strategies to enhance viral load testing, PrEP enrollment, and ART uptake for sustained VLS in sub-Saharan Africa.