Sarah Urbut: Predicting Your Health Arc

**Eric Topol** (0:00)
I think we have enough folks on board. A lot more will join. I want to welcome today to Ground Truths Sarah Urbut, who is an instructor at Harvard Medical School, the Broad Institute, Mass General Hospital. I always call it man's greatest hospital in Boston. And she's an incredible researcher, MD. Ph.D., who just published this week in Nature a paper about predicting one's health.
It wasn't framed that way necessarily, but that's what we're going to get into. The name of the model that she is working with and on the title of the paper is Aladynoulli.
Aladynoulli, yeah. So maybe first you could give us a little bit about your background and tell us about what is Aladynoulli.

**Sarah Urbut** (0:57)
Absolutely. Well, thank you so much for inviting me today.
Just a bit about my background. I'm a physician scientist. I see patients at Mass General and I do research at the Broad and at Harvard Medical School. And so I did my MD PhD in Statistical Genetics actually at University of Chicago. And I was speaking with Dr. Topol, became interested in math probably as a small child. When I realized that you really could tell stories, stories with math. It was kind of a really objective quantitative language for describing natural phenomena and you can be creative with it too. So that was sort of my first attachment to math. And so Aladynoulli, I'm sorry, go ahead.

**Eric Topol** (1:45)
Story about when you did all the questions.

**Sarah Urbut** (1:48)
Yes. Yeah. So when I was in, I was speaking with my father actually about, I thought maybe I would get the question about when did I sort of first become interested? And he said for him, his memory was, there was a middle school problem of the week. And most kids were required to do one problem. And I did all 36 or 37 And the reason was, there were several reasons. In the Midwest, you were allowed to stay up past midnight because it had like an 11 o'clock submission time or something in the Midwest. Yeah. Anyway, so that was one. But the bigger thing was, there were two parts of the question. You had to answer the question and then you had to do a write-up. And in the write-up, it gave you an opportunity to be creative, but it also gave you an opportunity to see where the math was telling a story, I think. And for me, that was a decision point where I really discovered that math was a lot more than the multiplication division that you hammered home when you're a child. And that followed me forward.
But to answer your direct question about Aladynoulli, so there are several parts to the word, so a latent, and latent is unobserved. We talk about this a lot in math and computer science, but there's unobserved phenomena that we hope, well, in this case, for instance, captures a lot of underlying processes that are reflected in the diseases that we do observe. So it's something in math and science that we're trying to get at this underlying phenomena. So that's the latent. It's dynamic, because things are only interesting when they're moving, right? So there's this evolving process of both the individual and the population level. An individual is walking through time and acquiring these new diagnoses. And similarly, in the population, a particular latent signature, that's a big deal in the paper, but has these diseases that have characteristic incidence patterns that have different sort of sequences. And then the Uli piece, now that's the trick. So Uli rises from an Italian statistician, Bernoulli, actually Daniel Bernoulli, and one of the senior author on this paper, in addition to Pradeep Natarajan and Dr. Alexander Gusev, is Giovanni Parmigiani.
And he's a wonderful statistician and colleague and friend also at Dana-Farber. And he is, Giovanni Parmigiani, as you might imagine, is Italian.
And so we had this idea that not only do we want to describe these processes, but Dr. Tobol, I think to your earlier point, we also wanted to do some prediction, which is unusual. Ordinarily, these sort of descriptive processes, these latent approaches, they go backwards, the retrospective, and they describe what's happening historically. But we want to be predictive because I'm a physician. I want to tell a patient what's going to happen tomorrow based on what happened yesterday or what happened today. And so a Bernoulli process is a coin flip. It's the underlying probability distribution between heads and tails, for instance. And so it was those three features, this latent unobserved feature, this fact that it's time-evolving, and the probabilistic coin flip nature of it that came together and made Aladynoulli. So it's a long answer, but I think it's the only thing you could call it. Yeah.