Sarah Urbut, MD, PhD , of the Mass General Brigham Heart and Vascular Institute , is the lead author of a paper published in Nature , “ A Bayesian framework for longitudinal EHR and genetic discovery .” Pradeep Natarajan, MD, MMSc , also of the Heart and Vascular Institute, is one of the co-senior authors, along with collaborators at Harvard Medical School, Harvard T.H. Chan School of Public Health, the Broad Institute of MIT and Harvard, and Dana-Farber Cancer Institute.
Q: What challenges or unmet needs make this study important?
In medicine, there is a tendency to view different specialties—or even different diseases within the same specialty—separately, rather than borrowing information across a patient's entire disease history. Different aspects of a patient’s care are often managed in silos, and their medical history is frequently treated as a snapshot in time rather than as a process that continuously evolves.
Moreover, patients are typically treated according to their diagnostic label (for example, “coronary disease” or “diabetes”) rather than the biological processes driving their condition. This is a major issue because two people with the same diagnosis can have very different mixes of underlying disease processes.
Until now, we have lacked a practical way to see the whole, evolving picture of a patient's health or to understand how their disease is developing over time and how it differs from others who share the same diagnosis.
Q: What central question(s) were you investigating?
In this study, we asked three key questions:
Q: What methods or approach did you use?
To help answer these questions, we developed an advanced generative model called ALADYNOULLI. The model borrows information across patient histories and hundreds of diseases to uncover underlying disease processes and predict individual outcomes in a dynamic fashion. Specifically, it analyzes patterns of diagnoses over time alongside age and genetic risk information to identify hidden disease signatures that influence how health conditions develop and progress.
Q: What did you find?
We found that ALADYNOULLI could reduce complex patient histories across 348 diseases into 21 reproducible, “latent” disease signatures—hidden patterns that capture underlying disease processes. When applied to data from three independent biobanks including more than 683,000 people with up to 52 years of follow-up, these signatures were highly similar across datasets and reproduced disease processes we already know and understand.
Genetic analyses of these signatures confirmed known genetic associations and uncovered others that were missed when researchers analyzed individual diseases separately.
The signatures were also strongly predictive of future disease risk, particularly when predictions were updated over time as new patient information became available.
Q: What are the real-world implications?
By pooling information across diseases and over time, ALADYNOULLI can dynamically predict a patient's future health outcomes while helping researchers uncover the biology underlying different disease signatures.
Because the model spans hundreds of conditions at once and updates as a patient's history grows, it also reflects how medicine is actually experienced: as the care of a whole person whose health evolves over time. And since the model transfers across health systems, those benefits do not depend on every hospital having access to a massive genetic dataset.
Q: What part of this work feels most meaningful to you personally?
As Bayesian statisticians, borrowing information across subgroups and using generative models to tell stories helps us better understand the world. As physicians, looking across patient stories helps us better understand each individual patient. It is exciting to bring those two philosophies together and succeed at both.
What excites us most is not just the improved risk prediction, but the ability to describe the variation in each patient’s unfolding disease trajectory in a concrete, multidimensional way. After all, two patients with the same diagnosis are not the same patient: this model shows they often have different underlying signature profiles, which can translate into different progression patterns and different responses to the same treatment.
Authorship: In addition to Urbut and Natarajan, Mass General Brigham authors include Tetsushi Nakao, Satoshi Koyama, Anika Misra, Whitney E. Hornsby and Jordan W. Smoller. Additional authors include Yi Ding, Xilin Jiang, Achyutha Harish, Leslie Gaffney, Alexander Gusev and Giovanni Parmigiani.
Paper cited: Urbut, S., et al . “A Bayesian framework for longitudinal HER and genetic discovery.” Nature. DOI: 10.1038/s41586-026-10780-5
Funding: This work was supported by NHLBI K08 grant (1K08HL183784), American Heart
Association Career Development Award (25CDA1444806) and Burroughs Wellcome Fund
award (1360373). A full list of funding sources can be found in the paper.
Nature
A Bayesian framework for longitudinal HER and genetic discovery
15-Jul-2026