Deephiv-AOU: a multi-modal deep learning framework leveraging the all of US research program for HIV precision prevention

The National Institute of Environmental Health Sciences, part of the National Institutes of Health, awarded the University of Texas Health Science Center at Houston $200,268 on September 1, 2026, under the Medical Library Assistance program (CFDA 93.879) for Project Grant R21LM015664.

The recipient will develop DEEPHIV-AOU, a multi-modal deep learning framework designed to improve HIV prevention through personalized risk assessment and intervention planning. The framework integrates four heterogeneous data streams—temporal electronic health records sequences, genomics, behavioral surveys, and non-medical drivers of health—drawn from the All of Us Research Program's nationwide cohort of over 400,000 participants. Aim 1 develops the framework to generate a dynamic HIV acquisition risk score and an intervention non-response likelihood. Aim 2 evaluates consistency using error parity metrics and ensures clinical interpretability via transformer-based attention maps. Aim 3 assesses clinical utility and health economic impact through microsimulation modeling and cost-effectiveness analysis. The project addresses the gap between current preexposure prophylaxis utilization ratios and population-level impact thresholds required to support federal mandates to end the HIV epidemic through precision public health.

Work is performed in Houston, Texas. The period of performance runs from September 1, 2026, through August 31, 2028. The assistance type is a Project Grant.

Abstract

The united states is currently falling short of its ending the HIV epidemic targets, with HIV incidence declining by only 4% annually despite a 90% reduction goal. This stagnation is driven by a critical GAP in preexposure prophylaxis (PREP) utilization and persistence, where the current ratio of PREP users to new diagnoses (15.6) remains far below the estimated threshold (112) required for population-level impact. Existing prevention efforts are limited by traditional predictive models that rely on unimodal data, lack temporal depth, and often fail to generalize across different populations. To address these barriers, we propose deephiv-AOU, a next-generation multi-modal deep learning (MDL) framework leveraged through the all of US (AOU) research program. AOU provides a large-scale, nationwide cohort of over 400,000 participants with linked longitudinal electronic health records (EHR), reflecting the characteristics of the U.S. Population. Our innovative architecture, using an intermediate fusion strategy, integrates four heterogeneous data streams: temporal EHR sequences, genomics, behavioral surveys, and non-medical drivers of health. Aim 1 will develop the MDL framework to generate a dynamic HIV acquisition risk score and an intervention non- response likelihood. Aim 2 will evaluate the framework for consistency using error parity metrics and ensure clinical interpretability via transformer-based attention maps. Finally, aim 3 will assess clinical utility and health economic impact through microsimulation modeling and cost-effectiveness analysis. By successfully bridging the GAP between high-dimensional data and clinical action, this project provides a foundation for personalized HIV prevention and directly supports federal mandates to end the epidemic through precision public health.