This AI Found 15.4% More HIV Cases by Rethinking How We Map Social Networks
New algorithm finds 15.4% more HIV cases using 25% fewer tests.
A collaboration between Harvard, the WHO, and the University of the Witwatersrand has produced a novel AI framework to make HIV testing far more efficient in resource-constrained environments. The paper introduces Policy-Embedded Graph Expansion (PEGE), which bypasses traditional network reconstruction by embedding a generative distribution over graph expansions directly into the decision-making policy. This is paired with Dynamics-Driven Branching (DDB), a diffusion-based model designed to handle sparse, forest-structured referral data common in real-world testing programs.
In experiments on real HIV transmission networks, PEGE + DDB outperformed all baselines: a 17.3% improvement in discounted reward and 15.4% more HIV detections when only 25% of the population had been tested. The approach directly supports UN Sustainable Development Goal 3.3 (ending the AIDS epidemic) and demonstrates how graph-based AI can drive public health impact without requiring complete network data.
- PEGE + DDB improves discounted reward by 17.3% over existing methods in HIV testing simulations.
- Achieves 15.4% more HIV detections with just 25% of the population tested.
- Designed for data-limited settings where referral networks form natural tree structures.
Why It Matters
Graph-based AI can dramatically improve late-stage disease detection in low-resource settings, saving lives efficiently.