Improving predictive machine learning for retention in HIV care via synthetic data augmentation

The National Institute of Mental Health awarded Vanderbilt University $106,092 on August 11, 2026, under the Mental Health Research Grants program (CFDA 93.242) to study synthetic data augmentation for improving machine learning models that predict retention in HIV care.

The project will develop and evaluate synthetic data augmentation methods—using generative artificial intelligence to create additional training data—to enhance machine learning models' ability to identify people living with HIV at risk of missing clinical visits and discontinuing antiretroviral therapy. Current machine learning approaches suffer reduced effectiveness when applied across different patient populations or healthcare settings. The research will address computational and biomedical validity limitations of synthetic data augmentation, clarify the mechanisms underlying its effectiveness, and test whether the approach can improve model generalizability across diverse HIV care contexts, thereby extending the benefits of predictive technologies to all people living with HIV.

The recipient is Vanderbilt University's Sponsored Programs Administration division, located in Nashville, Tennessee. The project period extends through August 7, 2029. The award is a project grant, a non-renewable assistance type under the NIMH program.

Abstract

This research proposal aims to enhance the effectiveness of machine learning (ML) in improving adherence to antiretroviral therapy (art) among people living with HIV (PWH). Adherence to art is critical for viral suppression, a key goal in the ending the HIV epidemic plan for the united states. While modern ML methods have been shown to identify individuals at risk of missing their visits with HIV providers and potentially discontinuing art, these ML models are only as effective as the data they are trained on. As a result, ML models can be ineffective when applied to different patient populations or different healthcare settings, potentially worsening health outcomes for PWH. To address this challenge, this project will study synthetic data augmentation (SDA), a method that uses generative artificial intelligence (AI) to create additional data to train ML models. This approach has shown promise in various applications of ML, but has yet to be evaluated in, or adapted for, the domain of HIV care. My hypothesis is that SDA can improve ML models' ability to predict retention in HIV care, ensuring the benefits of these technologies extend to all PWH. However, there are limitations to SDA that must first be addressed. First, SDA is a new method and can be performed with many different computational approaches. Thus far, these varied methodologies for SDA have produced mixed results, making IT unclear what the precise mechanism is behind how SDA works and in what circumstances SDA will be most effective. Second, the generative AI models that SDA utilizes do not have any external knowledge, learning only what they observe in the data. This makes generative AI models prone to creating unrealistic or biomedically implausible synthetic data. While a well-trained generative model does this infrequently, this can still occur and harm the performance of ML models trained via SDA. The proposal is structured around three specific aims. Aim 1: develop a foundational framework for understanding the impact of SDA on ML performance, particularly in the presence of issues common to HIV datasets. Aim 2: enhance the quality of synthetic HIV data by embedding biomedical knowledge into generative AI models. Aim 3: apply SDA to a large cohort dataset of PWH in the united states and evaluate the effectiveness of SDA in improving ML performance across different healthcare sites and populations. Successful completion of this project will lead to more reliable and generalized ML systems, ensuring that advancements in ML technology benefit all individuals living with HIV.