NIH is turning more than 12 petabytes of multimodal health data into a format usable by artificial‑intelligence tools, a task that depends on making the information interoperable across decades of research.
Building a “converter box” for legacy data
The effort centers on BioData Catalyst, a cloud‑based ecosystem created by the National Heart, Lung, and Blood Institute (NHLBI) together with the National Library of Medicine (NLM) and the Office of Data Science Strategy (ODSS). “NIH has over 12 petabytes of data that’s all multimodal—everything from genomics, clinical imaging, sleep, sensor data, all different modalities,” said Sweta Ladwa, chief of the Scientific Solutions Delivery Branch at NHLBI’s Information Technology and Applications Center.
That collection includes long‑running studies such as the Trans‑Omics for Precision Medicine (TOPMed) program, which follows roughly 180,000 participants. Simply storing the data does not make it ready for AI analysis. The main obstacle is ensuring that a cardiovascular measurement recorded in a 1990s Framingham Heart Study can be understood the same way as a similar metric from a recent pulmonary‑fibrosis trial.
To address this, NHLBI has built a pipeline based on the Linked Data Modeling Language (LinkML). Ladwa described it as a “converter box approach, where you can plug in the data from the source and it will put it in that [format] for your analysis.” The system maps information across standards such as LOINC, FHIR and the Human Phenotype Ontology (HPO).
Automated mapping is paired with clinical validation. “We have pulmonologists who we’re working with to really clinically determine the concepts,” Ladwa said, noting that the goal is to confirm that different drug names refer to the same therapeutic class. The AI‑assisted process relies on “publicly available metadata” rather than patient‑level data, reducing privacy concerns while still harmonizing terminology.
Related: WHAT IS HORMONE PELLETS THERAPY?
Linking research standards to everyday clinical records
While NHLBI focuses on making existing research data interoperable, ODSS tackles the complementary challenge of bringing research‑grade standards into the electronic medical record (EMR) systems that generate new data daily. Susan Gregurick, NIH’s associate director for data science and director of ODSS, explained a push to align NIH research standards with the United States Core Data for Interoperability (USCDI), the benchmark used for EMR accreditation.
Work that began in oncology is expanding to other disease areas, including a new cardiovascular partnership with NHLBI. The aim is for EMR systems to capture cardiovascular phenotypes during routine patient encounters in a format that researchers can immediately use.
“The impact for that sort of cross‑agency collaboration is really huge,” Gregurick said. “I think that’s almost apart from AI, but it’s going to be something that drives AI in the future.”
Underlying much of this work is the NLM, the world’s largest biomedical research library. Lisa Federer, acting director of the NLM’s Office of Strategic Initiatives, said the library provides “the substrate for future work in AI” through tools such as Medical Subject Headings (MeSH) and Common Data Elements, which standardize how research data is described and collected across NIH.
Federer noted a shift in who consumes the data. “We’re not just thinking about humans. We’re thinking about how machines are consuming data as well,” she said. “You do have to consider not just a human consuming that, but how is agentic AI going to be consuming this information?”
Related: Shane Schaffer Elaborates on Competitive Market Dynamics
Standardizing Common Data Elements has become more urgent as the NLM observes an increase in bots and other AI agents crawling its repositories. The distinction between curating data for human readers versus machine agents influences how metadata is structured and tagged.
From a broader perspective, the drive to harmonize decades of health data reflects a growing recognition that data silos limit scientific progress. By translating older study formats into modern, machine‑readable standards, researchers can combine historical cohorts with contemporary real‑world evidence, potentially accelerating discovery and improving patient outcomes.
NIH invested nearly $400 million last year in AI‑related research grants, according to Gregurick. Yet the less visible investment—building pipelines, ontologies and standards that enable data to move across institutional boundaries—may have an even larger impact.
“When you’re able to connect that data with the real‑world data, with other data that exists out in the research space or in the health data fabric across the nation or even internationally, you just increase the power of that data to be able to do more,” Ladwa said, adding that the ultimate goal is to help those affected by diseases and disorders.
Without interoperable standards, 70 years of health data stays locked in the formats it was born in. With the new standards, it becomes the foundation for a new generation of AI‑driven medical research.