A Machine Learning Approach to Predict HIV Viral Load Hotspots in Kenya Using Real-World Data

Area of Science:

• Public Health
• Machine Learning
• Epidemiology

Background:

• Machine learning is not routinely used for predicting HIV status.
• The study focuses on developing a machine learning model to predict HIV viral load (VL) hotspots in Kenya.
• Hotspots are defined as health facilities with ≥20% of people living with HIV (PLHIV) not achieving viral suppression, per WHO recommendations.

Purpose of the Study:

• To develop a machine learning model for predicting HIV viral load hotspots.
• To establish an early warning system for health administrators in Kenya.
• To optimize treatment and resource distribution for antiretroviral therapy (ART) programs.

Main Methods:

• A random forest model was developed using routinely collected data from Ministry of Health affiliates.
• Patient-level data (4 million tests, 4,265 facilities) were aggregated to the facility level after cleaning and outlier/multicollinearity checks.
• The facility-level dataset was split into training (75%) and testing (25%) sets for model development.

Main Results:

• The model achieved 78% accuracy in discriminating between hotspots and non-hotspots.
• The model reported an F1 score of 69% and a Brier score of 0.139.
• In December 2019, the model accurately predicted 434 additional VL hotspots beyond the observed 446.

Conclusions:

• The developed hotspot mapping model is crucial for antiretroviral therapy programs.
• It supports decision-makers in proactively identifying VL hotspots.
• The model utilizes cost-efficient, routinely collected data for timely interventions.