Abstract
By personalizing healthcare to an individual’s specific requirements, precision health promises to maximize benefit and minimize harm, thereby maximizing value. We describe here, how in Phase 2 of the Million Veteran Program–Computational Health Analytics for Medical Precision to Improve Outcomes Now (MVP-CHAMPION), artificial intelligence (AI) and high performance computing (HPC) have been applied to Veteran’s electronic health records (EHRs) and genetic data to advance real-world precision health. Eight concept projects were selected on the basis of potential impact on high-burden conditions among Veterans, including heart failure, suicide, lung cancer, diabetes, post COVID-19 sequelae, medication toxicity, and obstructive sleep apnea. Achievements include new and more discriminating risk prediction models to inform medical decision making, multimorbidity-aware analytic frameworks, and development and deployment of reusable computational tools. The identification of novel risk factors from genetic data and unstructured text in the EHR has both informed risk prediction and offered new insights for medication repurposing and development. We not only confirmed the need for shared infrastructure, data management, and novel AI-based workflows to inform precision health, but also found that such programmatic improvements result in valuable mechanistic insights. By building on these foundations through expanded deployment, adaptive modeling, and broader partnerships, the VA-DOE collaboration is poised to transform not only the future of Veterans’ healthcare, but the broader national landscape of precision health.
Introduction
The United States healthcare system currently faces a critical challenge: despite record-high spending, life expectancy continues to lag behind other developed nations. This discrepancy is largely driven by a "one size fits all" approach to medicine that often fails to account for the complex reality of multimorbidity—the presence of multiple chronic conditions in a single patient. To address this, the Department of Veterans Affairs (VA) and the Department of Energy (DOE) established the MVP-CHAMPION (Million Veteran Program–Computational Health Analytics for Medical Precision to Improve Outcomes Now) collaboration. By integrating the VA’s massive clinical datasets with the DOE’s world-class supercomputing power, the partnership aims to move precision health from theoretical research into clinical practice.
Methods and Data Infrastructure
The scale of the data involved in this collaboration is unprecedented. The study leverages the Million Veteran Program (MVP), which contains genomic data from over one million Veterans, linked to 30 years of longitudinal electronic health records (EHR) for more than 25 million individuals. This includes over 4 billion unstructured clinical notes and integrated geospatial data regarding social determinants of health, such as air quality and housing stability.
To process this data, the collaboration utilized several cutting-edge AI and high-performance computing (HPC) tools:
- Exascale Computing: Leveraging the DOE’s "Frontier" supercomputer to perform massive-scale Natural Language Processing (NLP) and Genome-Wide Association Studies (GWAS).
- VA-LLM: A custom 1.62-billion parameter Large Language Model pretrained on 1.4 trillion tokens of VA clinical text. This domain-specific model was designed to extract insights from unstructured notes that traditional structured data often misses.
- Explainable AI (iRF-LOOP): The "Iterative Random Forest Leave-One-Out Prediction" framework was employed to ensure that AI models are not "black boxes," providing clinicians with interpretable reasons for specific risk predictions.
- Late Fusion Techniques: These techniques integrate structured EHR data (vitals, labs) with LLM-derived insights to create robust ensemble models for disease prediction.
Key Insights and Clinical Results
Phase 2 of MVP-CHAMPION focused on eight high-burden conditions, yielding significant improvements in risk prediction and mechanistic understanding:
Suicide Risk Prevention
Using deep sequential neural networks, researchers identified a top 0.1% risk tier where 99% of the identified Veterans attempted suicide within the following year. This model achieved an adjusted positive predictive value of 0.54, offering a highly targeted tool for clinical intervention.
Lung Cancer and Environmental Factors
The team achieved a c-statistic of 0.81 for predicting lung cancer diagnoses. Notably, the analysis identified air pollution as a stronger predictor of cancer risk than many traditional clinical markers, highlighting the importance of environmental data in precision health.
Renal Disease in Diabetes
A dynamic prediction model for end-stage renal disease (ESRD) was developed, achieving an AUROC (Area Under the Receiver Operating Characteristic curve) of greater than 93%. This significantly outperforms current state-of-the-art risk scores used in diabetic care.
Heart Failure and Genomics
By applying network analysis (GRIN & MENTOR), researchers identified novel gene clusters related to inflammation and fibrosis. These findings provide new targets for medication repurposing, particularly for patients with heart failure with preserved ejection fraction (HFpEF), a condition with limited treatment options.
Discussion: The Value of Shared Infrastructure
A primary finding of the collaboration was that shared infrastructure leads to exponential efficiency. For example, an NLP pipeline originally built to detect homelessness for suicide prevention was rapidly adapted to extract risk factors for obstructive sleep apnea. Furthermore, the success of the 1.62B parameter VA-LLM demonstrated that domain-specific training on high-quality medical data can outperform much larger general-purpose models like Llama 3 or Mistral.
The research also underscored the scientific necessity of diversity. Because the MVP includes a diverse Veteran population, researchers found that 12% of genomic risk loci were only significant in non-European ancestry groups, ensuring that the benefits of precision health are distributed equitably across different populations.
Conclusion and Future Directions
The VA-DOE collaboration has successfully demonstrated that massive-scale interagency partnerships can overcome the "translational gap" in precision medicine. Future work will focus on moving from static models to "continuously learning systems" that update in real-time. The next frontier includes the development of "digital twins"—virtual patient models used to simulate disease progression and test personalized interventions in silico. By expanding these foundations, MVP-CHAMPION is poised to transform the national landscape of healthcare value and precision medicine.
Summary updated