←
Sarah Urbut150 posts
Sarah Urbut avatar✓

Sarah Urbut ✓

@tigerstatdoc

Cardiologist and scientist @MGHHearthealth @harvardmed fascinated by statistics, genomics @broadinstitute, and exploring the world on two wheels.

◷ Joined August 2013▧ 8 Photos and videos
Sarah Urbut avatar
Sarah Urbut@tigerstatdoc·Jul 15

One of my first consults as a cardiology fellow: a 34-year-old, textbook MI. A day earlier, no risk model would have flagged him for prevention. That paradox has driven my work ever since — our models miss how disease actually evolves, dynamically, on top of a genetic background. @Nature nature.com/articles/s41586-0…

Post attachment
Eric Topol avatar
⟳ Sarah Urbut retweeted
Eric Topol@EricTopol·Aug 4

Accurately predicting an individual's health arc is a dream that has not been previously possible. Now we're headed there, thanks to a new @nature paper by @tigerstatdoc and team. Our conversation and annotations in the new Ground Truths

Post attachment
Sasha Gusev avatar
⟳ Sarah Urbut retweeted
Sasha Gusev@SashaGusevPosts·Jul 16

Very exciting new method for inferring latent disease trajectories, led by the tireless Dr. Urbut with Natarajan and Parmigiani labs. If you're considering a postdoc in this area talk to Sarah!

Post attachment
Eric Topol avatar
⟳ Sarah Urbut retweeted
Eric Topol@EricTopol·Jul 15

Predicting a person's health and disease arc many years in advance from their electronic health records and polygenic risk scores, along with the biological mechanism for the diseases just published @Nature Will be discussing the paper with the brilliant lead author @tigerstatdoc this Friday in Ground Truths 10:30 AM PST nature.com/articles/s41586-0…

Post attachment
Sarah Urbut avatar
Sarah Urbut@tigerstatdoc·Jul 15

Two things evolve smoothly in time: the signatures themselves (φ), and each person's loading on them (λ) — anchored in germline genetics, so the model builds from and reveals new biology, rather than black box artifacts. Then Bayesian updating does the work: every new diagnosis refines the estimate, turning past into present into a prediction of the future. Every parameter is interpretable and described in our generative framework, scaled to work with biobank level data.

Post attachment
Sarah Urbut avatar
Sarah Urbut@tigerstatdoc·Jul 15

But patients don't sit still. How much a signature matters to you depends on when you ask the question. Three patients, three trajectories, distinct from the next patient who shares the same diagnosis. Same phenotype, different biology.

Post attachment
Sarah Urbut avatar
Sarah Urbut@tigerstatdoc·Jul 15

At its core are signatures: latent patterns of how disease co-occur and unfold over time, learned from data with no supervision. The model discovers, for instance, an ischemic cardiovascular signature which appropriately describing the sequence of events from hypercholesterolemia to CAD to ischemic heart failure. All learned in an unsupervised way from an interpretable generative model.

Post attachment
Sarah Urbut avatar
Sarah Urbut@tigerstatdoc·Jul 15

Risk isn't a number you're born with or a snapshot at a clinic visit. It lives in three dimensions at once — time, disease, and the individual. As a patient walks through life, her journey can take many forms. So we built Aladynoulli to model these patterns dynamically.

Post attachment