SCH: New Bayesian Causal Structural Discovery and Reinforcement Learning Methods for Mental health in HIV Care

While antiretroviral therapy (ART) has significantly reduced HIV-related mortality and increased life expectancy for people living with HIV (PWH), a range of comorbidities (e.g., kidney disease, mental health conditions, cognitive impairment) remain highly prevalent. Depression, the most common mental health comorbidity in PWH, affects 20% to 60% of this population, posing a major challenge to long-term HIV management. Modern combination ART (cART) regimens typically consist of three or more drugs from multiple classes with different mechanisms. Since PWH must remain on cART indefinitely once initiated, and its effects on depression vary across individuals, designing individualized, optimally effective cART regimens with minimal risk of worsening depression is critical in the emerging field of precision medicine for HIV. The availability of large-scale, longitudinal HIV cohort data, such as the MACS/WIHS Combined Cohort Study (MWCCS), spanning over 35 years, presents an unprecedented opportunity to investigate the effects of cART on both viral suppression and depression at an individual level. However, significant scientific challenges remain, including the need to accurately predict individuals' depression and other health outcomes, account for complex drug-drug interactions in estimating cART effects, and develop strategies for planning long-term, patient-tailored regimens that adapt to evolving health conditions. This proposal aims to develop novel statistical and machine learning methods to address these challenges, advancing NIAID's mission by leveraging real-world cohort data and innovative data science approaches to drive precision medicine for people with HIV. Specifically, we propose three aims: (1) Develop novel causal structural discovery models and robust prediction tools to effectively handle distribution shifts resulting from interventions in HIV-related health outcomes; (2) Develop a Bayesian model-based reinforcement learning (RL) framework to optimize personalized cART regimens and improve long-term mental health outcomes in PWH; and (3) Encapsulate the proposed statistical methods and computational algorithms into R and Python packages and develop a web interface for practical application and dissemination. RELEVANCE (See instructions): While antiretroviral therapy (ART) has reduced HIV-related mortality and increased life expectancy for people with HIV, mental health comorbidities, including depression, remain prevalent. Our proposed Bayesian causal discovery and reinforcement learning methods aim to accurately predict depression and other health outcomes while optimizing personalized ART. These advancements have the potential to reduce HIV transmission risk and assist physicians in making personalized treatment decisions.