Nothing about US without US: integrating the voices of people who use drugs into HIV resource allocation using large language models

The National Institute on Drug Abuse awarded the University of California, San Diego $2.237 million on May 15, 2026, under the Drug Use and Addiction Research Programs (CFDA 93.279) to develop and apply a community-informed modeling approach for HIV and hepatitis C virus prevention resource allocation among people who use drugs.

The project will develop large language models trained on behavioral and biological data from multiple people-who-use-drugs cohorts to generate artificial-intelligence-powered simulations, termed digital twins, that reflect intervention preferences and locally-specific behaviors. These digital twins will be integrated into an agent-based epidemic and economic model that simulates HIV and HCV transmission, overdose, and related complications across multiple U.S. jurisdictions. The integrated framework will evaluate the impact and cost-effectiveness of different HIV/HCV prevention and treatment strategies under various funding scenarios. Epidemic model outputs will be disseminated through an interactive dashboard co-designed with public health departments, paired with a policy-support language model to help decision-makers interrogate results and generate tailored communication materials. The project employs a community-based participatory process incorporating the preferences and voices of people who use drugs into resource allocation decisions.

Work is performed in La Jolla, California. The period of performance runs through March 31, 2031.

Abstract

This project will develop and apply a novel, community-informed modeling approach to guide HIV and hepatitis c virus (HCV) prevention resource allocation among people who use drugs (pwud) in the united states. Despite proven harm reduction and major advances in HIV prevention and treatment, including long-acting injectable art, the united states continues to fail to meet the needs of pwud, a population central to intersecting HIV, HCV, and overdose crises. A key reason is the absence of a rapid, scalable, and comprehensive way to capture and act on the preferences and behaviors of pwud. Revolutionary advances in large language models (llms) now make IT possible to create high-fidelity "digital twins" - artificial-intelligence powered simulations of individuals that reflect their characteristics, attitudes, and likely behaviors. Using a community-based participatory process, we will train an LLM on behavioral and biological data from multiple pwud cohorts to generate pwud digital twins that simulate real-world attitudes, behaviors, and service uptake. These digital twins will be integrated into an agent-based epidemic and economic model (simdemic) that simulates HIV and HCV transmission, overdose, and related complications, across numerous jurisdictions in the united states. The integrated modeling framework will be used to evaluate the impact and cost-effectiveness of different HIV/HCV prevention and treatment strategies, including responses to potential funding constraints. Model outputs will be shared through an interactive simdemic dashboard, CO-designed with public health departments, featuring a policy-support LLM to help decision-makers interrogate results and generate tailored communication materials. Our project includes the following activities: 1) CO-design, development, and validation of a pwud-informed LLM. 2) digital twin simulations within an agent-based HIV/HCV transmission and overdose model (simdemic) to inform resource allocation. 3) dissemination and implementation of simdemic to county health departments. The proposed research will enable rapid, data-driven evaluation of prevention strategies and resource allocation options. IT aligns with the NIH office of aids research priority to reduce new HIV infections and address HIV coinfections.