Multimodal Artificial Intelligence: Opportunities and Challenges in HIV Clinical Care


Lori A. J. Scott-Sheldon, Ph.D.
Division of AIDS Research


The goal of this concept is to encourage the use of multimodal artificial intelligence to accelerate HIV diagnosis, prevention, and treatment.


Novel, data-driven approaches will be necessary to fulfill the HHS Ending the HIV Epidemic in the U.S. Initiative key strategies to diagnose, treat, and prevent HIV. The use of transformative technologies in artificial intelligence (AI) may allow for fully flexible, reusable models that offer the ،ential to expand capabilities for ending the HIV epidemic. This concept, consistent with the HHS AI Strategy , seeks to strengthen the generation, adoption, and use of more trustworthy AI systems to improve the overall health and well-being of people living with or at risk for HIV.

Recent advances in highly flexible, reusable AI models have generated much interest and discussion about the promise of AI to transform research, clinical care, and public health. This concept seeks to leverage these advancements in generative and interactive AI by encouraging the development, adoption, and use of cutting-edge multimodal AI models to augment existing narrow AI/ma،e learning approaches used in clinical care that will enable better target prevention efforts, optimize treatment decisions, and improve patient experience. One challenge is that most multimodal AI models lack interpretability. Multimodal AI enhanced with HIV-specific medical domain inputs using knowledge graphs has the ،ential for more accurate, explainable, and interpretable systems that can be used in a wide range of data-driven applications, including comprehensive patient histories, early HIV detection, personalized HIV treatment plans, real-time patient support, epidemic surveillance, and more to address HIV prevention, treatment, and care needs.

This concept aims to encourage the use of multimodal AI in clinical care that can expand our capacity to address the dynamic, complex, and evolving HIV epidemic with three goals: (1) development, adaptation, and use of accurate, safe, efficient, and unbiased multimodal AI models to support a diverse range of HIV applications; (2) creation of knowledge graphs to strengthen model interpretability and explainability for application in HIV research, clinical care, and public health; and (3) exploration of synergistic integration of knowledge graphs and multimodal models for more precise model outputs and increased usability in HIV clinical care. Cross-disciplinary teams, including end-users, data scientists, domain experts, software developers, and engineers are encouraged. Teams are expected to use best practices and meaningful stake،lder engagement to ensure data privacy, security, and consent. Inherent in this concept is the need for the use of a human-centered AI approach in the design, development, and implementation to ensure that the system is fair, ethical, and unbiased for a more transformative, meaningful, and sustained impact on HIV clinical care. Finally, NIMH encourages approaches to address HIV and mental health priorities across the prevention and care continuum.