# Sarah Urbut: Predicting Your Health Arc

An episode of [Ground Truths](https://invisiblenetwork.ai/podcast/1728526108), 4 Aug 2026, 54 min.

Audio: https://api.substack.com/feed/podcast/207215981/d6a40f77ff94a7d2a75587636cac6cd7.mp3

Sarah Urbut is a physician-scientist at Harvard Medical School and the Massachusetts General Research Institute. She and her colleagues published a landmark paper in Nature on patient health journeys for 3 different cohorts—the Mass General Brigham patients, the UK Biobank, and the All of Us research program—with a cumulative 683,000 individuals, 348 diseases, and up to 52 years of follow-up. Integrating electronic medical records and polygenic risk scores (N=36). An AI model was developed and validated called ALADYNOULLI, with each individual represented latent disease signatures (N=21), mathematically and biologically driven. An example of a patient’s dynamic life journey (health arc) is provided in the Figure below (adapted from the paper, made with NotebookLM). The longitudinal story for each individual is key here. Using a Gaussian process, the risk continually shifts like a GPS navigation when you make a wrong turn and you are re-routed. Considerable effort was dedicated to addressing selection bias with inverse probability weighting. Here are 20 temporal pattern signatures for the UK Biobank cohort, which help to provide the biologic mechanism of a person’s disease, as shown for multiple conditions like heart attack, breast cancer, and depression. That is, the same disease phenotype can link to different genomic pathway underpinnings. An example for breast cancer could be related to signature 8 or signature 7; the latter tied to inflammation and metabolic abnormalities. This biological pattern could also be used to detect likely medication failure (such as SSRI treatment for depression) and predict rare diseases. In fact, the model was used to discover genomic signatures for different diseases, as shown below. This model predicted 1 and 10-year risk of diseases at the individual level and far surpassed routinely used clinical prediction calculators such as the pooled cohort equation (PCE) for coronary artery disease (0.89 vs 0.68, respectively, 1-year risk ) and the GAIL model for breast cancer (0.783 vs 0.54, respectively). ALADYNOULLI also surpassed accuracy for precision when compared with the recent Delphi-2M model that I wrote about here The ALADYNOULLI paper is packed with information and equations and is not easy to get through without a sophisticated math background. That’s why the summary infographic below (made by Notebook LM) may be helpful to grasp its main findings Prediction of Your Health Trajectory in Context The new model substantively adds to many other recent reports that include Delphi 2M, APOLLO, and AURORA that predict future health events up to 20 years out. Each of these models were different with respect to inputs, how patient data were represented (tokens, embeddings, or latent states), and prediction outputs as briefly summarized in the Table I made below. Notably, the inputs are complementary and likely additive, such that the deeper and longer (longitudinally) the data re for each individual, the better predictive accuracy will be possible. Adding to these models are many specific disease prediction (e.g. cancer, asthma, diabetes, neurodegenerative diseases) and sources of data (such as sleep lab, wearables, or high-throughput proteomics). We are now seeing intense efforts at medical forecasting which will continue to evolve and something that was not previously possible in patient care. What I mean is that we could provide a rough sense of a patient’s risk for, let’s say Alzhiemer’s disease, but could not temporally place when it might occur. Age 65 or 95? It makes a huge difference. And so does the ability to apply counterfactuals, what-if scenarios for the individuals to reduce the risk, markedly defer a condition’s appearance, or even fully prevent it. That’s why ALADYNOULLI adds yet another dimension, different methodology, multiple patient cohorts, and novel findings that further add to a new prediction and prevention dimension of medical practice. It couldn’t have been done…

## Transcript

[0:00] SPEAKER_02: I think we have enough folks on board.
[0:02] SPEAKER_02: A lot more will join.
[0:04] SPEAKER_02: I want to welcome today to Ground Truths,
[0:07] SPEAKER_02: Sarah Erbud,
[0:08] SPEAKER_02: who is an instructor at Harvard Medical School,
[0:12] SPEAKER_02: the Broad Institute, Mass General Hospital.
[0:15] SPEAKER_02: I always call it Man's Greatest Hospital in Boston.
[0:19] SPEAKER_02: And she's an incredible researcher, MD,
[0:24] SPEAKER_02: PhD, who just published this week in Nature, A paper about predicting one's health.
[0:33] SPEAKER_02: It wasn't framed that way necessarily, but that's what we're going to get into.
[0:38] SPEAKER_02: The name of the model that she is working with and on the title of the paper is Aladdin Newley.
[0:46] SPEAKER_02: Aladdin Newley, yeah.
[0:47] SPEAKER_02: So maybe first you could give us a little bit about your background and tell us
[0:52] SPEAKER_02: about what is Aladdin Newley.
[0:57] SPEAKER_00: Absolutely.
[0:57] SPEAKER_00: Well, thank you so much for inviting me today.
[1:01] SPEAKER_00: I am, yeah, so just a bit about my background.
[1:03] SPEAKER_00: So I'm a physician scientist.
[1:05] SPEAKER_00: I see patients at Mass General and I do research at the Broad and at Harvard Medical School.
[1:13] SPEAKER_00: And so I did my MD-PhD in statistical genetics, actually, at University of Chicago.
[1:19] SPEAKER_00: And I was speaking with Dr. Topol, became interested in math.
[1:23] SPEAKER_00: Probably as a small child when I realized that you really could tell stories,
[1:28] SPEAKER_00: stories with math.
[1:28] SPEAKER_00: It was kind of a really objective,
[1:31] SPEAKER_00: quantitative language for describing natural phenomena and you can be creative
[1:35] SPEAKER_00: with it too.
[1:37] SPEAKER_00: So that was sort of my first attachment to math.
[1:42] SPEAKER_00: And so Aladdin Wooley, I'm sorry,
[1:44] SPEAKER_02: go ahead.
[1:45] SPEAKER_02: Sorry about when you did all the questions.
[1:48] SPEAKER_00: Yes.
[1:49] SPEAKER_00: Yeah.
[1:49] SPEAKER_00: So when I was in,
[1:50] SPEAKER_00: I was speaking with my father actually about,
[1:53] SPEAKER_00: you know,
[1:53] SPEAKER_00: I thought maybe I would get the question about when did I sort
[1:55] SPEAKER_00: of first become interested?
[1:56] SPEAKER_00: And he said for him,
[1:57] SPEAKER_00: his memory was there was a middle school problem of the week
[2:01] SPEAKER_00: and most kids were required to do one problem.
[2:04] SPEAKER_00: And I did all 36 or 37 times.
[2:07] SPEAKER_00: And the reason was, there were several reasons.
[2:09] SPEAKER_00: One, in the Midwest,
[2:10] SPEAKER_00: you were allowed to stay up past midnight because,
[2:12] SPEAKER_00: you know, it had like an 11 o'clock submission time or something in the Midwest.
[2:16] SPEAKER_00: Yeah.
[2:16] SPEAKER_00: Anyway, so that was one.
[2:18] SPEAKER_00: But the bigger thing was there were two parts of the question.
[2:20] SPEAKER_00: You had to answer the question.
[2:21] SPEAKER_00: And then you had to do a write-up.
[2:23] SPEAKER_00: And in the write-up,
[2:24] SPEAKER_00: it gave you an opportunity to be creative,
[2:26] SPEAKER_00: but it also gave you an opportunity to see where the math was telling a story,
[2:29] SPEAKER_00: I think.
[2:29] SPEAKER_00: And for me,
[2:30] SPEAKER_00: that was a decision point where I really discovered that math was a lot more
[2:35] SPEAKER_00: than sort of multiplication division that you kind of hammered home when you're
[2:39] SPEAKER_00: a child.
[2:40] SPEAKER_00: And that sort of followed me forward.
[2:41] SPEAKER_00: Yeah.
[2:43] SPEAKER_00: But to answer your direct question about Aladdin Uli, so there's several parts to the word.
[2:49] SPEAKER_00: So a latent, and latent is unobserved.
[2:53] SPEAKER_00: We talk about this a lot in math and computer science,
[2:57] SPEAKER_00: but they're sort of unobserved phenomena that we hope, well,
[3:01] SPEAKER_00: in this case,
[3:02] SPEAKER_00: for instance,
[3:02] SPEAKER_00: captures a lot of underlying processes that are reflected in the diseases
[3:06] SPEAKER_00: that we do observe.
[3:07] SPEAKER_00: So it's something in math and science that we're trying to get at this underlying phenomenon.
[3:11] SPEAKER_00: So that's the latent.
[3:12] SPEAKER_00: It's dynamic because things are only interesting when they're moving, right?
[3:17] SPEAKER_00: So there's this evolving process of both the individual and the population level.
[3:21] SPEAKER_00: An individual is walking through time and acquiring these new diagnoses.
[3:24] SPEAKER_00: And similarly,
[3:25] SPEAKER_00: in the population,
[3:26] SPEAKER_00: A particular latent signature,
[3:28] SPEAKER_00: that's a big deal in the paper,
[3:30] SPEAKER_00: but has these diseases that have characteristic incidence patterns
[3:33] SPEAKER_00: that have different sort of sequences.
[3:36] SPEAKER_00: And then the Uli piece, now that's the trick.
[3:38] SPEAKER_00: So Uli rises from an Italian statistician, Bernoulli, actually Daniel Bernoulli.
[3:45] SPEAKER_00: And one of the senior authors on this paper,
[3:48] SPEAKER_00: in addition to Brideep Natarajan and Dr. Alexander Gusev, is Giovanni Parmigiani.
[3:53] SPEAKER_00: And he's a wonderful statistician and colleague and And so,
[4:03] SPEAKER_00: you know,
[4:06] SPEAKER_00: we had this idea that not only did we want to describe these processes, but Dr.
[4:10] SPEAKER_00: Topol,
[4:10] SPEAKER_00: I think to your earlier point,
[4:11] SPEAKER_00: we also wanted to do some prediction, which is unusual.
[4:14] SPEAKER_00: Ordinarily,
[4:15] SPEAKER_00: these sort of descriptive processes,
[4:17] SPEAKER_00: these latent approaches,
[4:19] SPEAKER_00: they go backwards,
[4:20] SPEAKER_00: the retrospective, and they describe what's happening historically.
[4:24] SPEAKER_00: But we want to be predictive because I'm a physician.
[4:26] SPEAKER_00: I want to tell a patient what's going to happen tomorrow based
[4:28] SPEAKER_00: on what happened yesterday or what happened today.
[4:30] SPEAKER_00: And so a Bernoulli process is a coin flip.
[4:33] SPEAKER_00: It's the underlying probability distribution between heads and tails, for instance.
[4:38] SPEAKER_00: And so it was those three features,
[4:40] SPEAKER_00: this latent unobserved feature,
[4:41] SPEAKER_00: this fact that it's time evolving,
[4:43] SPEAKER_00: and the probabilistic coin flip nature of it that came together and made a lot
[4:47] SPEAKER_00: of noise.
[4:47] SPEAKER_00: So it's a long answer.
[4:49] SPEAKER_00: But I think it's the only thing you could call it.
[4:51] SPEAKER_02: Oh, it's kind of bringing together Aladdin and Uli, Bernoulli.
[4:57] SPEAKER_02: It's really quite a hybrid and explainer.
[5:00] SPEAKER_02: Now, for everybody that's recently watching, we have...
[5:06] SPEAKER_02: Really extraordinary data set that Sarah worked with 680,000 plus patients coming
[5:14] SPEAKER_02: from three sources,
[5:15] SPEAKER_02: the UK Biobank,
[5:17] SPEAKER_02: which is incredibly rich,
[5:19] SPEAKER_02: the All of Us, and also the Mass General Brigham health system data.
[5:24] SPEAKER_02: And that gave access to looking at almost 350 diseases and up to 50 plus years of follow-up.
[5:33] SPEAKER_02: Yeah.
[5:34] SPEAKER_02: And also,
[5:35] SPEAKER_02: not just electronic health records,
[5:38] SPEAKER_02: but what I love about this is that you got basically the polygenic risk scores,
[5:44] SPEAKER_02: 36 different disease polygenic risk scores,
[5:48] SPEAKER_02: identifying 150 loci in the genome that are of interest.
[5:53] SPEAKER_02: And then you got together and you were able to,
[5:56] SPEAKER_02: in this very remarkably three different cohort to make a huge progress,
[6:02] SPEAKER_02: you came up Latent signatures, like
[6:06] SPEAKER_00: 20 of them.
[6:07] SPEAKER_00: Beautiful.
[6:07] SPEAKER_00: Exactly.
[6:08] SPEAKER_00: Right.
[6:09] SPEAKER_02: That's in the extended figures.
[6:12] SPEAKER_02: I forgot which one.
[6:13] SPEAKER_02: But that's where I want to start is you found things that people haven't found before.
[6:20] SPEAKER_02: And so the word latent is, you know, is apropos.
[6:23] SPEAKER_02: So tell us about these signatures.
[6:26] SPEAKER_02: Yeah.
[6:27] SPEAKER_00: Yeah, absolutely.
[6:28] SPEAKER_00: No, right.
[6:28] SPEAKER_00: So, yeah.
[6:29] SPEAKER_00: So taking a step back, I think the 350 or 348 diseases is a really key feature.
[6:35] SPEAKER_00: And so when I was describing with Dr.
[6:37] SPEAKER_00: Topol before, you know, the idea was sort of a scaling up feature.
[6:41] SPEAKER_00: It was that as a physician or and as a Bayesian statistician,
[6:45] SPEAKER_00: I know that sort of more information is better, right?
[6:48] SPEAKER_00: I can learn about a pattern of disease from diseases that are perhaps
[6:55] SPEAKER_00: at face value not directly related to the disease I'm considering.
[6:59] SPEAKER_00: And then incorporating more information really adds to my understanding of the issue at hand.
[7:04] SPEAKER_00: But modeling sort of these patterns of disease is not directly clear
[7:08] SPEAKER_00: when you have such large numbers, right?
[7:11] SPEAKER_00: You could imagine there's these processes called Markov models
[7:14] SPEAKER_00: where you have transitions, right?
[7:16] SPEAKER_00: So an individual goes from one state to the next.
[7:19] SPEAKER_00: But when you have 348 different conditions,
[7:22] SPEAKER_00: modeling those potential transitions becomes impossible.
[7:25] SPEAKER_00: It's intractable.
[7:26] SPEAKER_00: And so you want to separate them into some patterns.
[7:29] SPEAKER_00: And you want these patterns to have the perfect balance of something that is both,
[7:34] SPEAKER_00: you know, so it's learned.
[7:37] SPEAKER_00: It's not something that we pre-specified.
[7:39] SPEAKER_00: We call that learned or unsupervised.
[7:42] SPEAKER_00: But you also want them to make sense at the end of the day.
[7:44] SPEAKER_00: And that's always the secret,
[7:45] SPEAKER_00: that once you throw whatever algorithm you've developed at something that's learning
[7:48] SPEAKER_00: in an unsupervised fashion,
[7:49] SPEAKER_00: that what you get at the end of the day,
[7:51] SPEAKER_00: you know, doesn't make the elephant weigh three pounds, so to speak.
[7:54] SPEAKER_00: Yeah.
[7:56] SPEAKER_00: And so it's this very critical piece.
[7:59] SPEAKER_00: And you also,
[7:59] SPEAKER_00: the secret,
[8:00] SPEAKER_00: the other fashion is that because you want it to come out of the other end
[8:03] SPEAKER_00: and doing some form of prediction,
[8:05] SPEAKER_00: You basically need everything to be in a prediction scale.
[8:08] SPEAKER_00: So everything to be in a probability scale,
[8:09] SPEAKER_00: which inherently sort of condenses the space of variation that you get to work with,
[8:13] SPEAKER_00: right?
[8:13] SPEAKER_00: Because on a yearly basis, prediction, the absolute rates are sort of small.
[8:18] SPEAKER_00: So there's some mathematical tricks, essentially, to make this something that you can estimate.
[8:22] SPEAKER_00: But one is transforming this to the logit space so that we model these patterns
[8:27] SPEAKER_00: of disease enrichment over time on sort of the log-odd space.
[8:31] SPEAKER_00: And then there's these two-state process of a form of clustering that determines
[8:37] SPEAKER_00: which patterns of co-occurrence happen.
[8:40] SPEAKER_00: And each signature,
[8:41] SPEAKER_00: for a given signature and a given disease,
[8:43] SPEAKER_00: it's centered around basically the population incidence for the disease,
[8:47] SPEAKER_00: plus an offset that recognizes how important that disease is to that signature.
[8:52] SPEAKER_00: So, for example...
[8:53] SPEAKER_00: If we have a cardiovascular disease and we did things right,
[8:56] SPEAKER_00: you know, it wouldn't be telling the story if it didn't work out.
[8:59] SPEAKER_00: So if we have an ischemic cardiovascular disease signature,
[9:02] SPEAKER_00: which the model learns, right,
[9:03] SPEAKER_00: that wasn't supervised just because I'm a cardiologist,
[9:06] SPEAKER_00: but you would hope that it would learn that the enrichment
[9:11] SPEAKER_00: of ischemic cardiovascular disease and precursors as well,
[9:15] SPEAKER_00: like hypercholesterolemia,
[9:16] SPEAKER_00: coronary artery disease,
[9:17] SPEAKER_00: myocardial infarction,
[9:19] SPEAKER_00: and even congestive heart failure are enriched in this signature and not in others.
[9:24] SPEAKER_00: And then there's a beautiful term that's added to that as well, which is this Gaussian process.
[9:29] SPEAKER_00: And that's the term that really allows these diseases to evolve smoothly over time.
[9:33] SPEAKER_00: The Gaussian process is really just a fancy form of a kernel.
[9:37] SPEAKER_00: A kernel is like a covariance matrix.
[9:39] SPEAKER_00: But it has this nice natural interpretation that says that If I look at a trajectory
[9:45] SPEAKER_00: of time between two years,
[9:46] SPEAKER_00: three years,
[9:47] SPEAKER_00: four years,
[9:48] SPEAKER_00: ten years,
[9:48] SPEAKER_00: for instance,
[9:49] SPEAKER_00: the value of the disease's importance to that signature is more close
[9:53] SPEAKER_00: to adjacent time periods than it is to far time periods.
[9:56] SPEAKER_00: It's intuitive.
[9:57] SPEAKER_00: It makes sense.
[9:58] SPEAKER_00: It's interpretable.
[9:59] SPEAKER_02: Yeah, the way I kind of comprehend this Gaussian is
[10:04] SPEAKER_02: that let's say you're driving your car,
[10:08] SPEAKER_02: you have your GPS on,
[10:10] SPEAKER_02: And then the GPS, you made a wrong turn and your GPS is rerouting.
[10:16] SPEAKER_02: And that is kind of what's happening here is you get rerouting to account for
[10:22] SPEAKER_02: the fact that risk of a disease change shift with age.
[10:27] SPEAKER_02: So it's constantly reassessing the risk as part of this process.
[10:31] SPEAKER_00: Is that a fair?
[10:34] SPEAKER_00: That's actually a perfect analogy.
[10:35] SPEAKER_00: I mean, it's like you're a perfect, or there's a perfect, I guess, plant.
[10:39] SPEAKER_00: Because,
[10:40] SPEAKER_00: you know,
[10:41] SPEAKER_00: in the optimization step,
[10:43] SPEAKER_00: which is kind of the perspiration aspect of this,
[10:45] SPEAKER_00: but in the optimization framework,
[10:47] SPEAKER_00: the way that this model is solved, is using a loss term.
[10:50] SPEAKER_00: And the loss term effectively penalizes deviations that are too big.
[10:54] SPEAKER_00: And so if you're moving a lot,
[10:55] SPEAKER_00: if the GPS is detecting that you made a wrong term, and I do that a lot.
[10:58] SPEAKER_00: In Boston, I actually try not to drive a car at all.
[11:00] SPEAKER_00: In Boston, it's bad for my health and the health of others.
[11:03] SPEAKER_00: But so if you move too much in a short period of time, it's going to penalize you.
[11:08] SPEAKER_00: But there's a little bit more forgiveness if you're over time that
[11:13] SPEAKER_00: the risk is changing substantially.
[11:15] SPEAKER_00: But in short periods of time, you don't want it to change very much.
[11:19] SPEAKER_00: But needless to say,
[11:21] SPEAKER_00: so these signatures that end up being sort of learned really suggest these patterns
[11:26] SPEAKER_00: of disease that are consistent with what we'd expect.
[11:30] SPEAKER_00: And within each pattern,
[11:31] SPEAKER_00: within each signature,
[11:33] SPEAKER_00: is another really critical feature that is sort of unique to our approach,
[11:37] SPEAKER_00: I think,
[11:37] SPEAKER_00: that has not been used in some of these other kind of clustering approaches.
[11:41] SPEAKER_00: And it is that The diseases that are enriched in that particular signature have
[11:46] SPEAKER_00: a characteristic pattern of incidence and timing that is neither the same
[11:50] SPEAKER_00: for all diseases, nor is it the same sort of among the signatures.
[11:55] SPEAKER_00: So for instance, and they have this natural biological sequence.
[11:58] SPEAKER_00: So in this ischemic cardiovascular signature, which I was describing, there's a nice...
[12:02] SPEAKER_00: Pattern in which hypercholesterolemia,
[12:04] SPEAKER_00: mild cardio infarction,
[12:05] SPEAKER_00: and congestive heart failure,
[12:06] SPEAKER_00: they're all enriched in the ischemic cardiovascular signature.
[12:10] SPEAKER_00: But the sort of characteristic time peak incidence is different between all three.
[12:14] SPEAKER_00: So hypercholesterolemia tends to precede mild cardio infarction,
[12:17] SPEAKER_00: tends to precede congestive heart failure.
[12:19] SPEAKER_00: Of course, that's at a population level.
[12:20] SPEAKER_00: These signatures are at the population level.
[12:22] SPEAKER_00: They're denoted by KD,
[12:23] SPEAKER_00: which is for the population,
[12:25] SPEAKER_00: which signature, which disease, not the individual level.
[12:28] SPEAKER_00: But that's kind of the biological sequence that we think about too, right?
[12:31] SPEAKER_00: And similarly,
[12:32] SPEAKER_00: there's a malignancy signature in which sort
[12:34] SPEAKER_00: of neutropenia precedes primary malignancies precede sort of secondary malignancies.
[12:38] SPEAKER_02: That makes sense.
[12:40] SPEAKER_02: Now, again, those of you who've joined, got over a few hundred already, please don't hesitate.
[12:46] SPEAKER_02: If you have questions, pop them in the message.
[12:49] SPEAKER_02: What this sets the stage for, what you just described, is something that's unique.
[12:54] SPEAKER_02: I know it took years to do all this work,
[12:57] SPEAKER_02: but what you were able to do is like a movie script for a person's health.
[13:04] SPEAKER_02: You basically,
[13:05] SPEAKER_02: instead of getting the snapshot when a doctor sees a patient and says, oh, well,
[13:09] SPEAKER_02: you know,
[13:10] SPEAKER_02: you've got this problem,
[13:11] SPEAKER_02: that problem, now you have a script for what's going to happen in the years ahead.
[13:16] SPEAKER_02: And this is quite extraordinary because I know like in clinic,
[13:19] SPEAKER_02: you're in prevention clinic seeing patients and we've not really had,
[13:23] SPEAKER_02: and we're going to talk about the broader topic here in a minute,
[13:28] SPEAKER_02: but we've not had the ability to predict accurately the Diseases in
[13:33] SPEAKER_02: a patient's what's in store.
[13:34] SPEAKER_02: I mean, we could say if, oh,
[13:38] SPEAKER_02: you have a polygenic risk or it says you have a high risk of heart disease,
[13:42] SPEAKER_02: but we don't know if it's going to show up when you're 50 or 99.
[13:45] SPEAKER_02: You know, that's kind of where we are.
[13:47] SPEAKER_00: Right.
[13:48] SPEAKER_02: It's a whole different level.
[13:50] SPEAKER_02: Exactly.
[13:50] SPEAKER_02: Yeah.
[13:52] SPEAKER_02: Let's just take,
[13:52] SPEAKER_02: you know,
[13:53] SPEAKER_02: I think people will understand the power of this better because if we start
[13:58] SPEAKER_02: to drill down on some individual conditions,
[14:02] SPEAKER_02: What you've done is not only being able to predict these latent signatures and all
[14:06] SPEAKER_02: the things you just described,
[14:08] SPEAKER_02: but you're also able to say,
[14:09] SPEAKER_02: you know what,
[14:11] SPEAKER_02: this woman has breast cancer,
[14:12] SPEAKER_02: this woman has breast cancer,
[14:14] SPEAKER_02: but the mechanism of their breast cancer,
[14:17] SPEAKER_02: even though they have the same phenotype, is different by the latent signature.
[14:22] SPEAKER_02: So can you expound on that?
[14:24] SPEAKER_00: Sure.
[14:24] SPEAKER_00: Yes, yes.
[14:26] SPEAKER_00: That's a great point.
[14:28] SPEAKER_00: So I think to the first point about sort of walking through
[14:32] SPEAKER_00: and making these predictions and to the polygenic risk piece,
[14:35] SPEAKER_00: you know,
[14:36] SPEAKER_00: I think about people beginning this and they have this underlying sort
[14:39] SPEAKER_00: of predisposition that's grounded in genetics and genetics.
[14:42] SPEAKER_00: Some of the more time-fixed features that perhaps you're born with
[14:47] SPEAKER_00: and have not necessarily been revealed by environment.
[14:50] SPEAKER_00: So you begin with that before we've acquired any diagnostic codes.
[14:53] SPEAKER_00: And so an individual is walking through and they're acquiring these new diagnostic codes.
[14:58] SPEAKER_00: And the Bayesian machinery is essentially saying, I've seen these new diagnostic codes.
[15:02] SPEAKER_00: Now I'm going to change my signature loading to reflect that.
[15:05] SPEAKER_00: So perhaps I had a predisposition for this type of disease, for coronary disease.
[15:10] SPEAKER_00: But in fact,
[15:10] SPEAKER_00: I'm actually acquiring some You know,
[15:13] SPEAKER_00: conditions, I'm acquiring some gastrointestinal type conditions.
[15:16] SPEAKER_00: And so actually at this point in time, my loading on that kind of signature is increasing.
[15:21] SPEAKER_00: So that will change throughout time.
[15:23] SPEAKER_00: So that's on the genetic background.
[15:24] SPEAKER_00: Genetics sort of lays the foundation and then the EHR codes are continuously updated
[15:30] SPEAKER_00: and incorporated into our practice.
[15:32] SPEAKER_00: So to your point about the heterogeneity behind a patient who had one condition,
[15:37] SPEAKER_00: so you take three individuals who had the same condition.
[15:39] SPEAKER_00: Breast cancer, I think, was the example that you gave.
[15:42] SPEAKER_00: And because they may have all had sort of the same, well,
[15:46] SPEAKER_00: they may not have had the same underlying genetic predisposition,
[15:50] SPEAKER_00: but they arrived at the same condition.
[15:52] SPEAKER_00: And if you look retrospectively,
[15:54] SPEAKER_00: you can then take their signature distribution over time,
[15:57] SPEAKER_00: meaning we had those 20 signatures,
[15:59] SPEAKER_00: and at each point in time, we evaluated which signature was most likely for them.
[16:02] SPEAKER_00: What was their distribution of signatures?
[16:05] SPEAKER_00: And we can say there are groups of people.
[16:08] SPEAKER_00: So there are people who had sort of more of a cardiovascular profile
[16:11] SPEAKER_00: that preceded their breast cancer signature.
[16:13] SPEAKER_00: There are people who had more of a metabolic profile
[16:16] SPEAKER_00: that preceded their breast cancer signature.
[16:17] SPEAKER_00: There are people who had,
[16:19] SPEAKER_00: you know, no really risk factors at all and just emerged into that type signature.
[16:24] SPEAKER_00: And so there are these groupings.
[16:25] SPEAKER_00: I think we have a figure actually that shows the heterogeneity between individuals
[16:29] SPEAKER_00: who had condition on having the same signature and You
[16:31] SPEAKER_00: then decompose their time-bearing,
[16:33] SPEAKER_00: or having the same disease, you decompose their time-bearing signature profile.
[16:38] SPEAKER_00: And we did it for CAD as well.
[16:39] SPEAKER_00: And it turns out that those groups of people track very well,
[16:43] SPEAKER_00: actually, with underlying differences in polygenic risk scores.
[16:47] SPEAKER_00: And that was where the 36 PRS really helped us.
[16:49] SPEAKER_00: Yeah.
[16:50] SPEAKER_02: Yeah, and there's a lot of things.
[16:52] SPEAKER_02: This is so loaded with rich insights.
[16:56] SPEAKER_02: By the way, I just want to mention, Michelle Malcolmson, thanks for your points.
[17:01] SPEAKER_02: We're going to get to a little later here the point that our system is only set up
[17:05] SPEAKER_02: for sick care and we don't do prevention,
[17:08] SPEAKER_02: particularly primary prevention across most conditions and diseases.
[17:13] SPEAKER_02: So we'll get into that.
[17:15] SPEAKER_02: It's sick, okay?
[17:16] SPEAKER_02: It shouldn't be happening.
[17:18] SPEAKER_01: Yeah.
[17:19] SPEAKER_02: And Anne Kenny just asked,
[17:21] SPEAKER_02: very good point,
[17:22] SPEAKER_02: isn't a cardiovascular signature very close to a metabolic signature?
[17:28] SPEAKER_00: Yes, yes.
[17:29] SPEAKER_00: So to both questions.
[17:31] SPEAKER_00: So the cardiovascular signature,
[17:32] SPEAKER_00: it is true,
[17:33] SPEAKER_00: but there are other pathways to get to CAD without necessarily being in
[17:36] SPEAKER_00: the metabolic signature, right?
[17:37] SPEAKER_00: And so we actually,
[17:38] SPEAKER_00: I think that was one of the examples that we showed in the In the paper
[17:42] SPEAKER_00: that some individuals get there sort of through a classic cardiometabolic type
[17:45] SPEAKER_00: of pathway.
[17:47] SPEAKER_00: Some individuals may have had precursors that didn't necessarily appear in
[17:50] SPEAKER_00: the metabolic pathway, for instance.
[17:51] SPEAKER_00: You know, they had an inflammatory predisposition.
[17:53] SPEAKER_00: Some people had nothing at all, right?
[17:55] SPEAKER_00: The risk factors that I was describing,
[17:56] SPEAKER_00: they just had high genetic risk and we really didn't observe anything at all.
[18:01] SPEAKER_00: And we do observe that in clinic as well,
[18:02] SPEAKER_00: individuals who just sort of draw the unfortunate lottery.
[18:07] SPEAKER_00: Yeah.
[18:07] SPEAKER_00: And then within the signatures,
[18:09] SPEAKER_00: there is a signature that actually had patterns of co-occurrence
[18:12] SPEAKER_00: that really were more like type 2 diabetes,
[18:15] SPEAKER_00: for instance,
[18:17] SPEAKER_00: was enriched in a signature that was distinct from
[18:19] SPEAKER_00: the ischemic cardiovascular signature.
[18:21] SPEAKER_00: So there was enough separation in the individual,
[18:24] SPEAKER_00: the patterns of co-occurrence, to separate those two signatures.
[18:27] SPEAKER_02: Oh, it's great.
[18:28] SPEAKER_02: So each of the conditions that you drill down on for these examples,
[18:32] SPEAKER_02: like depressions,
[18:33] SPEAKER_02: Like coronary disease,
[18:35] SPEAKER_02: breast cancer,
[18:36] SPEAKER_02: you're able to show that they have the same ICD code,
[18:39] SPEAKER_02: but their polygenic risk scores tell a different story about how they got there.
[18:46] SPEAKER_02: So what you contributed here was not just the ability to predict conditions,
[18:50] SPEAKER_02: but also to understand,
[18:51] SPEAKER_02: dissect, you know, how people got into this in the first place.
[18:55] SPEAKER_02: And you also,
[18:56] SPEAKER_02: you mentioned the word co-occurrence,
[18:58] SPEAKER_02: and this is also Point, which is people don't just have one disease.
[19:03] SPEAKER_02: They don't just present that way.
[19:06] SPEAKER_02: They have lots of things going on at any given time.
[19:09] SPEAKER_02: And so that's another, I think, a big feature here is that you're able to factor all that in.
[19:15] SPEAKER_00: That's exactly it.
[19:16] SPEAKER_00: Yeah, I view these signatures as a dimension of,
[19:19] SPEAKER_00: you know, when I say latent, right, it's unobserved.
[19:22] SPEAKER_00: So I view these signatures as a new tool to describe the heterogeneity in patients.
[19:26] SPEAKER_00: And we can describe it not only across the comorbidities,
[19:29] SPEAKER_00: across diseases that we may not have considered previously, but also across time.
[19:34] SPEAKER_00: So it's really expanding it in two dimensions,
[19:36] SPEAKER_00: temporally and across the multiple disease domains.
[19:39] SPEAKER_00: And so that's what that sort of signature clustering picture is trying
[19:42] SPEAKER_00: to demonstrate,
[19:43] SPEAKER_00: that three patients with the same condition can be further separated based
[19:48] SPEAKER_00: on their temporal associations with a variety of signatures over time.
[19:53] SPEAKER_02: So now James McNamee has already anticipated my next question, which is great.
[19:59] SPEAKER_02: He asked,
[19:59] SPEAKER_02: how much more predictive certainty does this new approach add to the patient
[20:03] SPEAKER_02: and physician?
[20:04] SPEAKER_02: And before you answer,
[20:05] SPEAKER_02: just to point out,
[20:06] SPEAKER_02: because you're modest about what this model,
[20:10] SPEAKER_02: Aladdin newly contributed, but you crushed it.
[20:13] SPEAKER_02: I mean, for breast cancer, the most accepted model, it's called PCE for pooled cohort.
[20:23] SPEAKER_02: You got it.
[20:24] SPEAKER_02: The prediction, 89% accurate, that model at 68.
[20:29] SPEAKER_02: We're talking about a huge gap.
[20:31] SPEAKER_00: Well, I should, yeah.
[20:32] SPEAKER_00: So PC is for coronary disease.
[20:35] SPEAKER_00: That's okay.
[20:36] SPEAKER_02: I was thinking about the Gale model for breast cancer.
[20:38] SPEAKER_00: That's okay, yeah.
[20:40] SPEAKER_02: You had it at 78 plus percent accurate versus 54%.
[20:44] SPEAKER_02: So anywhere you looked at it,
[20:46] SPEAKER_02: you predicted better than any of the models that are used today for prediction.
[20:50] SPEAKER_02: Is that fair?
[20:51] SPEAKER_00: That is fair.
[20:53] SPEAKER_00: It is fair.
[20:53] SPEAKER_00: I would say the other caveat that we enable that a lot of these models don't, too.
[20:57] SPEAKER_00: So one, we do it in a dynamic fashion.
[21:01] SPEAKER_00: So the argument with our approach is that because it ideally would be integrated in
[21:06] SPEAKER_00: the EHR and is capable of incorporating new information in a continuous fashion,
[21:11] SPEAKER_00: that you would make a prediction for what will happen to the patient in
[21:14] SPEAKER_00: the next year.
[21:15] SPEAKER_00: And then you would reevaluate because patients are constantly changing.
[21:19] SPEAKER_00: So they perhaps acquired new phi codes, which are sort of collapsing of the ICD codes.
[21:23] SPEAKER_00: We know the interaction between that phi code
[21:25] SPEAKER_00: and their underlying genetic comorbidities as well.
[21:29] SPEAKER_00: And we'd make an update.
[21:30] SPEAKER_00: And so the predictions that we quote, we actually, we do a variety in the paper.
[21:34] SPEAKER_00: We show what the 10-year sort of static prediction would be,
[21:36] SPEAKER_00: as if nothing else changed,
[21:38] SPEAKER_00: what would be the prediction based on the EHR codes
[21:40] SPEAKER_00: and their signature distribution today for the next 10 years.
[21:43] SPEAKER_00: And then we also do it on a dynamic fashion where For the one-year prediction
[21:47] SPEAKER_00: at enrollment and the median of the one-year predictions for
[21:50] SPEAKER_00: that same 10-year horizon.
[21:54] SPEAKER_00: And this is where we make the comparisons to some of the existing clinical risk tools.
[21:58] SPEAKER_00: And it's really interesting.
[22:00] SPEAKER_00: So that is one piece.
[22:05] SPEAKER_00: And then the other,
[22:05] SPEAKER_00: I think, piece to note is that we didn't require any expensive, really, cohorts.
[22:10] SPEAKER_00: I mean, of course,
[22:10] SPEAKER_00: the UK Biobank and the All of Us and Mass General Brigham are tremendous cohorts,
[22:14] SPEAKER_00: and we are so thankful for the people who participated in these cohorts.
[22:19] SPEAKER_00: But a lot of these scores to develop have required disease-specific information
[22:25] SPEAKER_00: and disease-specific collaboration behind institutions,
[22:28] SPEAKER_00: and they are the opposite of unsupervised.
[22:31] SPEAKER_00: You collect the variables that you think are interesting,
[22:34] SPEAKER_00: and then you build your algorithm on that.
[22:36] SPEAKER_00: And in a way that's,
[22:37] SPEAKER_00: I'm not going to say counter to the process of discovery,
[22:40] SPEAKER_00: but in a way, I mean,
[22:41] SPEAKER_00: it's not,
[22:42] SPEAKER_00: it allows you,
[22:43] SPEAKER_00: it prevents you from making these discoveries between the connections
[22:46] SPEAKER_00: between seemingly unrelated conditions, right?
[22:49] SPEAKER_00: It prevents you from learning.
[22:51] SPEAKER_02: Yeah, I mean, it's letting the data talk by itself rather than hiding everything.
[22:57] SPEAKER_02: And this is what I hear is that,
[22:59] SPEAKER_02: you know,
[22:59] SPEAKER_02: now you can have massive amounts of data,
[23:03] SPEAKER_02: you can Learn from AI analytics that we couldn't do and stop with the guiding stuff.
[23:09] SPEAKER_02: And in fact,
[23:10] SPEAKER_02: you took on a very,
[23:12] SPEAKER_02: before we,
[23:12] SPEAKER_02: this is another technical point,
[23:13] SPEAKER_02: which is quite extraordinary because you anticipated selection bias and
[23:19] SPEAKER_02: a big procedure of inverse probability weighting to try to really You know,
[23:26] SPEAKER_02: factor that in and granular,
[23:28] SPEAKER_02: you know, to really mitigate that because obviously that can influence the results.
[23:32] SPEAKER_02: Maybe you can tell us a bit about that.
[23:33] SPEAKER_00: Yeah, absolutely.
[23:35] SPEAKER_00: Yeah.
[23:35] SPEAKER_00: So, you know, one of the big concerns is that the UK Biobank, again, tremendous resource.
[23:41] SPEAKER_00: And I have made much of my really academic career has been,
[23:45] SPEAKER_00: you know, looking very closely at this.
[23:46] SPEAKER_00: So very appreciative for everyone in the UK Biobank.
[23:49] SPEAKER_00: But it is well known that it is a healthier biobank.
[23:52] SPEAKER_00: It tends to be older individuals, very well educated.
[23:55] SPEAKER_00: And so one of the concerns you might make is that conclusions that you draw in
[23:58] SPEAKER_00: the UK biobank wouldn't apply to other populations.
[24:02] SPEAKER_00: So the first sort of analysis that we did to address that concern is we looked at
[24:07] SPEAKER_00: the overlap of these signatures between UK biobank,
[24:10] SPEAKER_00: all of us, and Mass General Brigham.
[24:12] SPEAKER_00: Because we wanted to make sure that the biological relationship we would find was
[24:16] SPEAKER_00: Between the diseases and the signatures that we identified was similar.
[24:20] SPEAKER_00: And it was over 85% similar across the board of the signatures.
[24:23] SPEAKER_00: So that's the first piece.
[24:24] SPEAKER_00: To your question about the inverse probability weighting, though,
[24:28] SPEAKER_00: one of the really beautiful things about having this model-based approach,
[24:31] SPEAKER_00: and I use the word beautiful because truly the math is really beautiful here.
[24:35] SPEAKER_00: It's elegant.
[24:36] SPEAKER_00: It's, I mean, it's lit.
[24:38] SPEAKER_00: It's like a visceral reaction.
[24:39] SPEAKER_00: I really like it.
[24:41] SPEAKER_00: So there's a likelihood, right?
[24:42] SPEAKER_00: And so a likelihood is the probability of the data given the parameters of interest.
[24:46] SPEAKER_00: And so you write down the likelihood,
[24:47] SPEAKER_00: and then the inverse probability weight is basically the probability
[24:50] SPEAKER_00: of participating or being in the UK biobank versus the general population,
[24:56] SPEAKER_00: relatively.
[24:57] SPEAKER_00: Essentially, that is what the inverse probability weight captures.
[24:59] SPEAKER_00: So if there's a population that is biased,
[25:01] SPEAKER_00: that would be captured in that inverse probability weight.
[25:04] SPEAKER_00: And so now,
[25:05] SPEAKER_00: for every person that you observe,
[25:07] SPEAKER_00: instead of having each person contribute equally,
[25:10] SPEAKER_00: you have them just with a simple multiplier
[25:12] SPEAKER_00: that reflects how likely they would be appearing in this population if we had it in
[25:17] SPEAKER_00: a more general population sense, not the UK Biobank population sense.
[25:22] SPEAKER_00: And so you can just recalculate that likelihood and recalculate the model fit using
[25:26] SPEAKER_00: that simple inverse probability weighting approach.
[25:28] SPEAKER_00: And it turned out it really did not change things at all.
[25:32] SPEAKER_00: And we even, in one of our supplementary analyses, we really sort of pushed it to the max.
[25:37] SPEAKER_00: And this was a supplementary analysis that was actually Dr. Alexander Gusev suggested.
[25:42] SPEAKER_00: So to his credit, he really pushed it.
[25:44] SPEAKER_00: And I was like, oh boy, this is going to fail.
[25:45] SPEAKER_00: It's not going to work.
[25:46] SPEAKER_00: But it worked.
[25:47] SPEAKER_00: So, and that was,
[25:50] SPEAKER_00: suppose you have a population that you've now contrived, you said is 90% female.
[25:56] SPEAKER_00: Can you, with these inverse probability weights, re-weight?
[25:59] SPEAKER_00: So essentially now the probability of breast cancer should be higher
[26:02] SPEAKER_00: than you would have in your general population.
[26:04] SPEAKER_00: And if you use the inverse probability weights to correct this, can you get it back to your old?
[26:08] SPEAKER_00: And it turns out you could do both.
[26:10] SPEAKER_02: That's terrific.
[26:12] SPEAKER_02: Now there's one thing that I've especially been interested in for years,
[26:16] SPEAKER_02: and actually a lot of work has come from your place,
[26:19] SPEAKER_02: and Pradeep, Nanarajan, and colleagues.
[26:22] SPEAKER_02: And that's just clonal hematopoiesis of indeterminate potential chip.
[26:27] SPEAKER_02: Now, if I lose anybody, what basically that is is the blood stem cells have a mutation, okay?
[26:34] SPEAKER_02: The clones of mutation.
[26:36] SPEAKER_02: And we've known for years that the I for indeterminate is a pretty stupid term
[26:42] SPEAKER_02: because it isn't a good thing to have.
[26:44] SPEAKER_02: Certain clones are really bad for your health, okay?
[26:48] SPEAKER_02: Yeah.
[26:49] SPEAKER_02: And they could predict blood cancers,
[26:52] SPEAKER_02: cardiovascular disease, big-time inflammation, other cancers.
[26:57] SPEAKER_02: You don't want to have CHIP of certain types.
[27:00] SPEAKER_02: And we don't assess it.
[27:02] SPEAKER_02: In fact, I wrote a ground truth about this.
[27:04] SPEAKER_02: We don't even measure it in the clinic, which drives me nuts because it is so helpful.
[27:09] SPEAKER_02: But you studied CHIP in this out loud newly.
[27:13] SPEAKER_02: What do we know about CHIP?
[27:15] SPEAKER_00: Yeah, great.
[27:15] SPEAKER_00: So we used it as a confirmatory assay as well.
[27:18] SPEAKER_00: So, you know,
[27:19] SPEAKER_00: one of the concerns is that these signatures that we are identifying,
[27:24] SPEAKER_00: I wouldn't say concern,
[27:24] SPEAKER_00: but one of the questions was, are they identifying true biological phenomena?
[27:28] SPEAKER_00: How do you show that, right?
[27:29] SPEAKER_00: Because a lot of these black box sort of deep learning based approaches,
[27:33] SPEAKER_00: they make no claims about biological phenomena.
[27:35] SPEAKER_00: There's not really anything interpretable about them.
[27:38] SPEAKER_00: Yeah.
[27:39] SPEAKER_00: And we have this observation or this hope,
[27:41] SPEAKER_00: this hypothesis that individuals who are enriched in these signatures,
[27:46] SPEAKER_00: if they are truly biologically interesting,
[27:48] SPEAKER_00: that individuals who have an event and are enriched in the signatures and also have
[27:53] SPEAKER_00: a biological condition like CHIP
[27:56] SPEAKER_00: or like familial hypercholesterolemia would be enriched in the signatures above and
[28:01] SPEAKER_00: beyond those who have the condition but were not.
[28:04] SPEAKER_01: Mm-hmm.
[28:05] SPEAKER_00: And that actually ended up being the case.
[28:06] SPEAKER_00: So for some of the critical illnesses,
[28:08] SPEAKER_00: for instance,
[28:09] SPEAKER_00: and the malignancies,
[28:10] SPEAKER_00: and I think it was signature 12 and signature 16,
[28:13] SPEAKER_00: I think we show also in the paper,
[28:15] SPEAKER_00: that individuals who had CHIP were enriched above and beyond other individuals
[28:19] SPEAKER_00: who had the same condition.
[28:20] SPEAKER_00: So having a condition already makes you sort of enriched in that signature.
[28:25] SPEAKER_00: So that's not enough of a threshold.
[28:26] SPEAKER_00: You have to show that it's above and beyond the condition.
[28:29] SPEAKER_00: Cases did not have CHIP.
[28:31] SPEAKER_00: And so we showed that for both CHIP and we showed that for familial hypercholesterolemia.
[28:36] SPEAKER_02: Yeah, no, it's so valuable, new insights here to really bring it home.
[28:41] SPEAKER_02: Now, the broader field, let's talk about that.
[28:44] SPEAKER_02: And by the way,
[28:45] SPEAKER_02: again, if you have questions for Sarah, don't hesitate to put them in the message.
[28:49] SPEAKER_02: We'll try to get to them.
[28:50] SPEAKER_02: But the broader question is,
[28:52] SPEAKER_02: we're in this new era that I think began,
[28:55] SPEAKER_02: I don't know if you agree,
[28:57] SPEAKER_02: Back last September in Nature,
[28:59] SPEAKER_02: the same journal you just published in,
[29:02] SPEAKER_02: Moritz Gerstam from Germany's colleagues published the so-called Delphi 2M paper.
[29:09] SPEAKER_02: And what they did was, again, the UK Biobank, it took 400,000 people.
[29:14] SPEAKER_02: And then they went to 1.9 million people in Denmark.
[29:18] SPEAKER_02: And they basically showed that you could predict diseases in people for the next 20 years.
[29:23] SPEAKER_02: And I talked to Moritz about this.
[29:25] SPEAKER_02: And he said, we were shocked.
[29:27] SPEAKER_02: We never thought you'd be able to do that from the electronic health record.
[29:31] SPEAKER_02: And they did have polygenic risk score, although much more limited than what you have.
[29:37] SPEAKER_02: And in their supplement, they showed that would help the accuracy of the prediction.
[29:41] SPEAKER_02: But again, we're not talking about, you know, do you have a risk?
[29:45] SPEAKER_02: We're talking about when.
[29:47] SPEAKER_02: We're talking about when the person would show up.
[29:50] SPEAKER_02: Which we'd never had before, and Delphi 2M ushered that in.
[29:54] SPEAKER_02: This tokenization of health data,
[29:56] SPEAKER_02: which, so people know LLMs, that term is now embedded, large language models.
[30:02] SPEAKER_02: These are LHMs, large health models, where the tokenization is not words or portion of word.
[30:09] SPEAKER_02: It's not pixels.
[30:11] SPEAKER_02: It's health data.
[30:13] SPEAKER_02: And so what you did here took the Delphi 2M to another level.
[30:20] SPEAKER_02: You had the electronic health records.
[30:21] SPEAKER_02: You didn't just have UK Biobank and two other big cohorts,
[30:25] SPEAKER_02: but you also had much more rich genomic data, the polygenic risk scores.
[30:32] SPEAKER_02: And you didn't just go for,
[30:33] SPEAKER_02: oh, we can predict this when this person's going to get lung disease.
[30:37] SPEAKER_02: You went for why?
[30:39] SPEAKER_02: What was the mechanism?
[30:41] SPEAKER_02: And using,
[30:42] SPEAKER_02: a lot of people don't appreciate polygenic risk scores because they say,
[30:45] SPEAKER_02: oh, well, you know, what does it really mean?
[30:47] SPEAKER_02: But you use this in a different way.
[30:49] SPEAKER_02: You use this to understand pathways.
[30:53] SPEAKER_02: You took the polygenic risk scores to a level that in many ways,
[30:56] SPEAKER_02: you know, hadn't really been utilized before, it seems.
[31:01] SPEAKER_02: So can you compare, which by the way, you did predict better than Delphi 2M.
[31:07] SPEAKER_02: Delphi 2M, Apollo,
[31:09] SPEAKER_02: which was recently a preprint from one of your colleagues at Harvard,
[31:13] SPEAKER_02: Faisal Mahmood.
[31:15] SPEAKER_02: And then other of these, what I call tokenization of healthcare, large health models.
[31:20] SPEAKER_02: What are your thoughts about this big new thing?
[31:24] SPEAKER_00: Yeah.
[31:25] SPEAKER_00: I mean, tremendous work, both papers.
[31:28] SPEAKER_00: So I think there's a couple of distinctions between ours.
[31:33] SPEAKER_00: So I would say ours truly is a generative model-based approach.
[31:37] SPEAKER_00: I mean, you can write it down.
[31:38] SPEAKER_00: And maybe that seems like just a technical detail,
[31:43] SPEAKER_00: but I don't think it is because,
[31:46] SPEAKER_00: for instance,
[31:47] SPEAKER_00: in your question about the inverse probability weights,
[31:49] SPEAKER_00: if you wanted to reweight your population to reflect the biases that you know,
[31:53] SPEAKER_00: you can do that if you have a likelihood you write down, right?
[31:55] SPEAKER_00: You can interpret the parameters that you've developed these signatures.
[31:58] SPEAKER_00: If I have tokenization in an LLM or an LHM,
[32:01] SPEAKER_00: It's hard to exactly know what's producing the predictions that I get out.
[32:04] SPEAKER_00: I get out these predictions, but I don't really know why.
[32:06] SPEAKER_00: I just am satisfied with the predictions.
[32:07] SPEAKER_00: And there's nothing wrong with that.
[32:09] SPEAKER_00: I mean, it's fine.
[32:10] SPEAKER_00: It serves a purpose.
[32:12] SPEAKER_00: But it doesn't really allow you to sort of do upstream science of that.
[32:17] SPEAKER_00: So with the genetic scores, we actually use genetics in two ways.
[32:21] SPEAKER_00: We also remove the polygenic risk score background.
[32:26] SPEAKER_00: And we refit an individual on their signature loading.
[32:29] SPEAKER_00: And then we said, now that becomes a new phenotype.
[32:32] SPEAKER_00: So you have an individual's loading on a signature over the history that they've been observed.
[32:36] SPEAKER_00: And you can say,
[32:37] SPEAKER_00: let's associate everyone's lifetime signature exposure,
[32:40] SPEAKER_00: if you will, with an underlying genetic variant.
[32:43] SPEAKER_00: And so you can actually do for each individual or for each, I mean,
[32:46] SPEAKER_00: it's the population,
[32:47] SPEAKER_00: but you do 20 different GWAS because you have a set of genetic variants
[32:52] SPEAKER_00: that are associated with each of the signatures.
[32:54] SPEAKER_00: And it was incredible.
[32:56] SPEAKER_00: I mean, again,
[32:57] SPEAKER_00: I guess to the equal incredulity,
[33:01] SPEAKER_00: I guess of the Delphi authors, I didn't know what we'd get at the end.
[33:04] SPEAKER_00: You know,
[33:04] SPEAKER_00: we have these signatures,
[33:05] SPEAKER_00: maybe they're not biologically interpretable,
[33:07] SPEAKER_00: but it turned out That they actually reproduced all of
[33:10] SPEAKER_00: the GWAS findings we would have suspected.
[33:12] SPEAKER_00: And they did another thing which you would hope for,
[33:15] SPEAKER_00: which was because they're learning about a variety of diseases within one signature,
[33:19] SPEAKER_00: right?
[33:19] SPEAKER_00: So in the ischemic cardiovascular signature,
[33:21] SPEAKER_00: there's hypercholesterolemia, coronary artery disease, and heart failure.
[33:27] SPEAKER_00: If there are genetic variants that are associated with each of them,
[33:30] SPEAKER_00: but are not associated with All of them within one trait, for instance.
[33:35] SPEAKER_00: You would hope that a signature that brings those things together actually identifies more.
[33:39] SPEAKER_00: And in fact, it does.
[33:40] SPEAKER_00: So it borrows power by sort of bundling those diseases together and actually learns
[33:45] SPEAKER_00: about variants that are sort of just below the power threshold
[33:48] SPEAKER_00: in each individual trait.
[33:51] SPEAKER_00: And we did a rare variant association to a similar effect,
[33:54] SPEAKER_00: you know, with each phenotype and found, again, very sort of...
[33:59] SPEAKER_00: Reassuring variants associated there.
[34:02] SPEAKER_02: Oh, that's great.
[34:03] SPEAKER_02: Now, Andy Cosgrove is asking about, okay, you've got these pristine data sets, all of us.
[34:09] SPEAKER_01: Yeah.
[34:10] SPEAKER_02: Biobank and Mass General Brigham is,
[34:12] SPEAKER_02: you know, probably a cut above a lot of other health system data sets.
[34:15] SPEAKER_01: Yeah.
[34:16] SPEAKER_02: And in the real world,
[34:18] SPEAKER_02: if you're trying to use this idea,
[34:21] SPEAKER_02: you've got issues about data cleaning and incomplete missing data.
[34:26] SPEAKER_02: Yeah.
[34:27] SPEAKER_02: How do you foresee dealing with that stuff?
[34:30] SPEAKER_00: It's a great question.
[34:31] SPEAKER_00: I mean, it's one of the reasons that,
[34:33] SPEAKER_00: you know,
[34:33] SPEAKER_00: working with diagnostic codes,
[34:35] SPEAKER_00: and the other thing I should add is the Delphi paper is great.
[34:38] SPEAKER_00: It's looking at ICD-10 codes, which are billing codes.
[34:41] SPEAKER_00: And we showed in actually one of the analyses,
[34:44] SPEAKER_00: sub-analyses of our paper,
[34:45] SPEAKER_00: that we use fee codes,
[34:46] SPEAKER_00: and even we did further sort of phenotype curation after the fact,
[34:51] SPEAKER_00: not in an unsupervised,
[34:52] SPEAKER_00: not in the unsupervised way, but when we were evaluating prediction.
[34:55] SPEAKER_00: To say, here,
[34:56] SPEAKER_00: this is physician,
[34:57] SPEAKER_00: turns out it's convenient to be a physician sometimes
[34:58] SPEAKER_00: because you can have these physician-supervised collections of fee codes that I say,
[35:02] SPEAKER_00: this is actually a meaningful,
[35:03] SPEAKER_00: distinct disease category,
[35:05] SPEAKER_00: as opposed to three fee codes or three ICD codes that are kind of referring to
[35:09] SPEAKER_00: the same thing, right?
[35:11] SPEAKER_00: And that's distinct because you could imagine the Delphi model, you're predicting...
[35:16] SPEAKER_00: So that's a little bit dangerous.
[35:30] SPEAKER_00: But okay,
[35:31] SPEAKER_00: so to the point about the missing data,
[35:32] SPEAKER_00: the reason I brought that up is because certainly in the UK Biobank and
[35:35] SPEAKER_00: to some extent in any large health system where a patient is,
[35:40] SPEAKER_00: you know,
[35:41] SPEAKER_00: the ICD codes that collapse into fee codes and ultimately diseases are collected by
[35:46] SPEAKER_00: the entire health system, you can run but you can't hide.
[35:50] SPEAKER_00: Right.
[35:50] SPEAKER_00: Right.
[35:51] SPEAKER_00: Right.
[36:12] SPEAKER_00: Right.
[36:32] SPEAKER_00: Regardless of how many times you have follow-up or don't have follow-up, it's already occurred.
[36:37] SPEAKER_00: You're considered no longer at risk.
[36:38] SPEAKER_00: And that is helpful for chronic disease.
[36:40] SPEAKER_00: So coronary artery disease,
[36:41] SPEAKER_00: for instance,
[36:42] SPEAKER_00: you might have your index hospitalization for a myocardial infarction,
[36:45] SPEAKER_00: and then you follow up with your primary cardiologist every six months,
[36:48] SPEAKER_00: every year, etc.
[36:49] SPEAKER_00: It's not a new diagnosis.
[36:51] SPEAKER_00: You had one diagnosis that you're getting subsequent follow-up for.
[36:54] SPEAKER_00: So you're contributing one measure point.
[36:55] SPEAKER_00: And that kind of guards against this concern for missing data because if
[36:59] SPEAKER_00: a patient is lost to follow-up,
[37:01] SPEAKER_00: Or a patient is seen at fewer or less intervals because we have health inequities,
[37:05] SPEAKER_00: which are tremendous in any health system, right?
[37:08] SPEAKER_00: It is less of a problem, I think, with diagnostic codes.
[37:12] SPEAKER_00: I do worry about that a lot for imaging data and laboratory data.
[37:15] SPEAKER_00: I worry if we're just learning patterns of prediction among patients who are willing
[37:18] SPEAKER_00: to participate.
[37:19] SPEAKER_02: Yeah, really good point.
[37:21] SPEAKER_02: Now, of these different...
[37:24] SPEAKER_02: Large health models, tokenization, predicting events.
[37:27] SPEAKER_02: You know, some have predicted like then the Gerstung Delphi 2M out to 20 years.
[37:33] SPEAKER_02: I think it was three to five years.
[37:37] SPEAKER_02: As you go further and further out, your accuracy has to suffer to some extent, right?
[37:41] SPEAKER_02: Yeah.
[37:42] SPEAKER_00: It does.
[37:43] SPEAKER_00: Yeah.
[37:43] SPEAKER_00: Most of the analyses in the Delphi paper,
[37:45] SPEAKER_00: they actually look at these zero,
[37:47] SPEAKER_00: one, two-year horizons, and they have washout intervals, which I think is important.
[37:51] SPEAKER_00: When you're looking in an EHR,
[37:52] SPEAKER_00: you know,
[37:53] SPEAKER_00: there might be a basis,
[37:54] SPEAKER_00: a problem in which you have,
[37:57] SPEAKER_00: there was like just a sort of delay,
[37:59] SPEAKER_00: a lag until something was actually recorded,
[38:01] SPEAKER_00: but the patient was coming in for presentations.
[38:03] SPEAKER_00: So you'll argue that you're making A prediction on information
[38:06] SPEAKER_00: that you actually kind of leaked into the data set from just a few months before.
[38:10] SPEAKER_00: And so we did do that censoring analysis where we excluded information three months,
[38:13] SPEAKER_00: six months, one month, one year before.
[38:15] SPEAKER_00: Certainly the prediction suffers the further you will go out.
[38:18] SPEAKER_00: But I don't think that's what any of these models nor are models should be designed for.
[38:23] SPEAKER_00: I think this model is designed to be dynamically reassessed.
[38:26] SPEAKER_00: Hey, you've come in,
[38:27] SPEAKER_00: you've got these new EHR diagnoses,
[38:28] SPEAKER_00: we're going to relook at your signature distribution and And it's also,
[38:32] SPEAKER_00: you know, it's not only a negative sort of pessimistic outlook.
[38:36] SPEAKER_00: Maybe you made changes that show that you were on this trajectory.
[38:40] SPEAKER_00: And if you had stayed heavily loaded on these signature,
[38:42] SPEAKER_00: these were the things that would happen to you.
[38:44] SPEAKER_00: But for some reason, you had better health behaviors, better exercise, better diet.
[38:48] SPEAKER_00: Maybe you started taking a medication.
[38:50] SPEAKER_00: And so you actually start deviating from that trajectory, which is a really positive thing.
[38:55] SPEAKER_00: You just brought up the other big thing here, which is the what if.
[38:57] SPEAKER_00: So basically you've got,
[38:58] SPEAKER_00: you know,
[38:58] SPEAKER_00: this prediction accuracy capability we've never had in medicine before,
[39:00] SPEAKER_00: really, and it's just going to keep getting better.
[39:23] SPEAKER_02: You know, we're not done here yet.
[39:25] SPEAKER_02: You're not done
[39:25] SPEAKER_01: here yet.
[39:26] SPEAKER_02: Never done, no.
[39:28] SPEAKER_02: And, you know,
[39:29] SPEAKER_02: short term, a year,
[39:30] SPEAKER_02: two years, and it's really sharp, and it may not be so bad even out many years.
[39:35] SPEAKER_02: But then the point is what you're just bringing up,
[39:38] SPEAKER_02: which is, okay,
[39:39] SPEAKER_02: this person,
[39:40] SPEAKER_02: risk for heart disease,
[39:43] SPEAKER_02: and we're going to see what happens in the model that
[39:46] SPEAKER_02: If this person takes medication to lower their LDL, you know, to 20.
[39:52] SPEAKER_02: Exactly.
[39:53] SPEAKER_02: The what if thing is another part of this.
[39:56] SPEAKER_02: Actually, one of the Chinese large health models did a what if scenarios.
[40:01] SPEAKER_02: But how do you see that playing out?
[40:03] SPEAKER_02: I mean, are we going to go to the doctor and the doctor is going to say,
[40:07] SPEAKER_02: if you do this, we're going to do that.
[40:10] SPEAKER_02: You're going to see, you know, this.
[40:12] SPEAKER_02: Is that in the works, you think?
[40:13] SPEAKER_00: Such a good question.
[40:14] SPEAKER_00: It is, I mean, it's in our works for sure.
[40:16] SPEAKER_00: You know, there's this huge issue of confounding by indication, right?
[40:19] SPEAKER_00: So the patients,
[40:20] SPEAKER_00: I had a great example actually when I was initially getting motivated for some
[40:23] SPEAKER_00: of this work years ago,
[40:24] SPEAKER_00: which is if you observe laboratory values,
[40:26] SPEAKER_00: kind of the example I was giving before,
[40:28] SPEAKER_00: so you observe LDL cholesterol values,
[40:30] SPEAKER_00: for instance,
[40:31] SPEAKER_00: retrospectively,
[40:32] SPEAKER_00: the years prior to someone having a heart attack,
[40:33] SPEAKER_00: and you might actually observe that they drop for the general population for
[40:36] SPEAKER_00: a particular patient that you chose.
[40:38] SPEAKER_00: And you might just draw a silly conclusion that,
[40:40] SPEAKER_00: oh my gosh, LDL lowering is associated with heart disease.
[40:44] SPEAKER_00: I want to keep some of these LDL nice and high, you know.
[40:46] SPEAKER_00: But of course you'd be making the erroneous—I mean, no, I would lose my job for that, right?
[40:51] SPEAKER_00: But you would— You would be making the erroneous conclusion because you were missing
[40:58] SPEAKER_00: the fact that there was an astute clinician who said,
[41:00] SPEAKER_00: wait a second,
[41:00] SPEAKER_00: I see Mr.
[41:01] SPEAKER_00: Jones is at risk here, and so I'm going to start treating him, keeping his LDL low.
[41:04] SPEAKER_00: And maybe you actually staved off heart disease for years.
[41:07] SPEAKER_00: He would have had it earlier, but he actually had it later.
[41:09] SPEAKER_00: So it's a big issue, this confounding by indication.
[41:12] SPEAKER_00: And if you just take individuals who are treated in the biobank
[41:15] SPEAKER_00: and untreated individuals and estimate what is the risk of developing disease,
[41:19] SPEAKER_00: the treated individuals are almost inevitably going to have a higher risk
[41:21] SPEAKER_00: of developing disease because they were sicker to begin with.
[41:24] SPEAKER_02: And so I see our approach, it's a very hard problem because of this confounding.
[41:29] SPEAKER_00: And if you just control for static covariates,
[41:32] SPEAKER_00: which are covariates at the baseline,
[41:34] SPEAKER_00: so individuals,
[41:35] SPEAKER_00: you match them on blood pressure,
[41:36] SPEAKER_00: LDL cholesterol,
[41:37] SPEAKER_00: sex, age even, for instance, you still can't get rid of that confounding.
[41:42] SPEAKER_00: Because there's things that have been in the mix for years before, like that example.
[41:45] SPEAKER_00: And so I see our approach as a way to de-confound These observational data by using
[41:52] SPEAKER_00: an individual's contribution to the signature for years previously.
[41:57] SPEAKER_00: So that incorporates temporal information in a multivariate fashion,
[42:00] SPEAKER_00: perhaps in diseases that put them at risk in ways
[42:03] SPEAKER_00: that we may not have measured previously.
[42:05] SPEAKER_00: So that's one direction in the post hoc calibration.
[42:08] SPEAKER_00: In the other direction,
[42:09] SPEAKER_00: you could imagine that you actually include treatment as a variable in one
[42:13] SPEAKER_00: of your phenotypes.
[42:15] SPEAKER_00: And so then you're sort of estimating these deconfounded signatures de novo.
[42:19] SPEAKER_00: So receiving a treatment now puts you into a signature of
[42:21] SPEAKER_00: a low-risk coronary phenotype,
[42:24] SPEAKER_00: whereas not puts you into a signature of high-risk coronary phenotype, for instance.
[42:28] SPEAKER_00: So I certainly see treatment contributing,
[42:30] SPEAKER_00: but I think that also falls into the camp of missing data, for instance.
[42:34] SPEAKER_00: You have to be very careful about who's being treated and who's not being treated.
[42:38] SPEAKER_00: And the propensity weighting, right, that we've described.
[42:40] SPEAKER_02: Yeah, well, I'll add another one to that kind of example.
[42:45] SPEAKER_02: Later this month, we're starting a trial to prevent Alzheimer's disease.
[42:49] SPEAKER_02: And these are high-risk people.
[42:53] SPEAKER_02: Confirmation of high-risk is that they have a high PTAO-217 level,
[42:57] SPEAKER_02: which we've learned is a little bit like LDL because if you really go
[43:01] SPEAKER_02: through lifestyle improvements,
[43:02] SPEAKER_02: you can bring down your PTAO-217 substantially, 50%, 80%.
[43:06] SPEAKER_02: So we have...
[43:11] SPEAKER_02: A 1,200 participant high-risk trial
[43:14] SPEAKER_02: where half get this aggressive lifestyle improvement program,
[43:19] SPEAKER_02: the others just get some education material.
[43:22] SPEAKER_02: The whole idea here is for us to learn,
[43:25] SPEAKER_02: because you probably know there's been studies that say if you have P-tau
[43:28] SPEAKER_02: of this level and you're of this age,
[43:31] SPEAKER_02: you're going to have mild cognitive impairment in 6.1 years.
[43:35] SPEAKER_02: So we're just starting to see.
[43:37] SPEAKER_02: Now, the problem,
[43:38] SPEAKER_02: of course,
[43:38] SPEAKER_02: is in your, all of us, UK Biobank, Mass General Brigham, we don't have P-tau to 17.
[43:44] SPEAKER_02: No, right.
[43:45] SPEAKER_00: Of course, right.
[43:45] SPEAKER_02: We create a data set that eventually...
[43:49] SPEAKER_02: But this is such an exciting time because Alzheimer's is like the most dreaded disease of man.
[43:55] SPEAKER_02: And this is another one of these what if.
[43:57] SPEAKER_02: If we can get Aggressive lifestyle coaching to work and see what levels we affect,
[44:03] SPEAKER_02: you know, what magnitude, you know, maybe we can help prevent or put off a disease.
[44:08] SPEAKER_02: Now, this whole,
[44:10] SPEAKER_02: here's Benjamin Lewis, he has,
[44:12] SPEAKER_02: how does this work with the prevent calculator and the CT assessment of inflammation
[44:19] SPEAKER_02: in the arteries called fat attenuation index, FAI?
[44:23] SPEAKER_02: How do you deal with that?
[44:25] SPEAKER_00: Yeah, so we did compare it to the Prevent calculator.
[44:29] SPEAKER_00: And on the 10-year risk, we're slightly better, actually, than the Prevent calculator.
[44:32] SPEAKER_00: And certainly on the one-year basis where we're continuing to reevaluate,
[44:37] SPEAKER_00: it sort of blows it out of the park.
[44:39] SPEAKER_00: Prevent doesn't have a one-year comparison,
[44:44] SPEAKER_00: but they don't You imagine that these,
[44:47] SPEAKER_00: for instance,
[44:47] SPEAKER_00: in Gale,
[44:48] SPEAKER_00: the one-year and the 10-year horizons,
[44:49] SPEAKER_00: which they do offer,
[44:50] SPEAKER_00: are very strongly correlated because they're not allowing for the dynamic assessment
[44:55] SPEAKER_00: of these variables, how their importance changes over time.
[44:58] SPEAKER_00: And so I suspect that on a one-year basis,
[45:00] SPEAKER_00: we would also be really significantly outperforming prevent.
[45:05] SPEAKER_00: And we haven't compared it to the fat attenuation CT. Yeah,
[45:08] SPEAKER_02: that's a hot thing right now that has to score and give a quantification of
[45:13] SPEAKER_02: the Inflammation in one to three coronary arteries and been validated
[45:19] SPEAKER_02: in some 40,000 consecutive patients.
[45:23] SPEAKER_02: Now, we've covered a lot here, Sarah.
[45:27] SPEAKER_02: The work is extraordinary.
[45:28] SPEAKER_02: I get a landmark paper, I really do.
[45:32] SPEAKER_02: And not only that,
[45:33] SPEAKER_02: but you have reinforced to me the love of math,
[45:37] SPEAKER_02: how important that is because this is what was manifest when you were
[45:41] SPEAKER_02: in middle school.
[45:46] SPEAKER_02: And you're modest because some of these predictions truly trust our current things
[45:51] SPEAKER_02: that we use clinically.
[45:53] SPEAKER_02: So it really works.
[45:55] SPEAKER_02: And as we said, it's going to get better.
[45:57] SPEAKER_02: So what have we missed in this conversation?
[46:00] SPEAKER_02: What didn't I get at which you think is important?
[46:04] SPEAKER_00: I mean, I think it was remarkably efficient, I guess, in the discussion.
[46:09] SPEAKER_00: I think...
[46:11] SPEAKER_00: So the big novelty of this paper for me,
[46:15] SPEAKER_00: methodologically,
[46:15] SPEAKER_00: I think it's really novel,
[46:16] SPEAKER_00: actually,
[46:16] SPEAKER_00: that we're able to do something that is both discovering these latent features
[46:21] SPEAKER_00: but also doing some prediction.
[46:23] SPEAKER_00: And yeah,
[46:24] SPEAKER_00: I guess I would say that is the methodological novelty,
[46:27] SPEAKER_00: that ordinarily we view these as kind of dueling objectives,
[46:29] SPEAKER_00: that you have something that either describes things retrospectively
[46:33] SPEAKER_00: and makes some discovery,
[46:34] SPEAKER_00: some underlying association,
[46:35] SPEAKER_00: or Or something that just really improves prediction and we kind of don't care about
[46:39] SPEAKER_00: the interpretability.
[46:40] SPEAKER_00: This does both and it does it in a framework that's actually a generative approach,
[46:44] SPEAKER_00: a generative model that I can write down in really three equations on a blackboard
[46:47] SPEAKER_00: and teach other middle schoolers how to do it.
[46:50] SPEAKER_00: So, I mean, you know, now it had to be scaled up, right?
[46:54] SPEAKER_00: It had to be scaled up.
[46:54] SPEAKER_00: So Thomas Bayes probably could not have this calculated this in
[46:57] SPEAKER_00: the 17th century Monte Carlo simulation framework because
[47:00] SPEAKER_00: The simulation would probably have paused or the computer would have died or
[47:03] SPEAKER_00: the computer didn't exist.
[47:06] SPEAKER_00: So there had to be some new tools that were used for this approach.
[47:10] SPEAKER_00: But at its essence, it's both discovery, it's prediction, and it's a generative model.
[47:17] SPEAKER_00: I think the new directions that we're working
[47:20] SPEAKER_00: with are precisely this causal inference direction.
[47:23] SPEAKER_00: How do we model treatment effectively, both retrospectively and in a counterfactual way, right?
[47:28] SPEAKER_00: I want to tell my patients what can they do tomorrow to prevent the trajectory
[47:31] SPEAKER_00: that I've observed in patients yesterday.
[47:35] SPEAKER_00: And then I think also the incorporation of new data types and An understanding of
[47:40] SPEAKER_00: the underlying biology.
[47:42] SPEAKER_00: I think that we just scratched the surface with these GWAS,
[47:45] SPEAKER_00: and I think that there's a lot of insights to be gained here about some of
[47:48] SPEAKER_00: the other sort of genetic associations.
[47:52] SPEAKER_02: Like, for example, you can go back to UK Biobank and all of us with sequence data.
[47:58] SPEAKER_02: But what I think people may not get here yet is that we're talking about enabling preventions.
[48:06] SPEAKER_02: We've never been good in medicine to prevent primary diseases.
[48:10] SPEAKER_02: And in fact,
[48:11] SPEAKER_02: as Michelle made that point,
[48:14] SPEAKER_02: Malcolmson,
[48:15] SPEAKER_02: earlier that we get our systems all geared to,
[48:19] SPEAKER_02: you know, treatment, treatment, treatment, six.
[48:22] SPEAKER_02: This is a different world here that's starting to emerge.
[48:26] SPEAKER_02: Now, in cardiology,
[48:27] SPEAKER_02: the area you and I work,
[48:29] SPEAKER_02: we have been trying to prevent heart disease,
[48:31] SPEAKER_02: you know, with statins and lipid lowering.
[48:33] SPEAKER_02: But, you know,
[48:34] SPEAKER_02: there are lots of people out there who have risk of heart disease
[48:37] SPEAKER_02: and they're not getting in the mix.
[48:39] SPEAKER_02: But there aren't that many preventive, primary prevention strategies out there.
[48:43] SPEAKER_02: Like women with breast cancer might take tamoxifen if they're a certain group.
[48:48] SPEAKER_02: But, you know, we're talking about a very small list.
[48:50] SPEAKER_02: This is the big game changer.
[48:53] SPEAKER_02: What I see here is this is the segue to primary prevention of diseases,
[48:59] SPEAKER_02: of the important diseases.
[49:00] SPEAKER_02: And the way I look at it is this will be the singular biggest contribution of AI in medicine.
[49:08] SPEAKER_02: Now, I don't know if you would agree with that, but maybe I'm a little too, you know.
[49:13] SPEAKER_00: Oh, boy.
[49:14] SPEAKER_00: I don't know.
[49:15] SPEAKER_00: I'm a modest Midwesterner.
[49:16] SPEAKER_00: I am.
[49:19] SPEAKER_00: I'm a White Sox fan,
[49:20] SPEAKER_00: so I'm always a little pessimistic,
[49:21] SPEAKER_00: but
[49:27] SPEAKER_00: I do think this opens a lot of doors, actually.
[49:32] SPEAKER_00: I don't know if I'd say,
[49:33] SPEAKER_00: I would never say single biggest,
[49:34] SPEAKER_00: I guess,
[49:35] SPEAKER_00: about anything,
[49:37] SPEAKER_00: but it's a revolution,
[49:38] SPEAKER_00: and I think that we are changing the way that we think about incorporating time,
[49:43] SPEAKER_00: about incorporating multiple conditions.
[49:45] SPEAKER_00: I think we're well on our way, so.
[49:48] SPEAKER_02: Yeah, I mean, look,
[49:49] SPEAKER_02: I'm hoping,
[49:50] SPEAKER_02: of course,
[49:51] SPEAKER_02: as I've written a lot about,
[49:52] SPEAKER_02: that AI is going to improve the patient-doctor relationship and improve accuracy
[49:56] SPEAKER_02: and, you know, all sorts of things that I think it can and will do.
[50:01] SPEAKER_02: But I think this is the one we shouldn't lose sight of.
[50:04] SPEAKER_02: This is a new capability that's emerging now with multiple large health models.
[50:09] SPEAKER_02: This is the latest.
[50:10] SPEAKER_02: This one is very unique.
[50:12] SPEAKER_02: I mean, I have to say.
[50:13] SPEAKER_02: Thank you.
[50:15] SPEAKER_02: So before we wrap up, are there any other questions from the group?
[50:18] SPEAKER_02: We've had some great ones, and thank you for that.
[50:21] SPEAKER_02: But there's a chance to ask the wizard.
[50:24] SPEAKER_02: I mean, she is a wizard, I have to say.
[50:28] SPEAKER_02: I haven't met Sarah before, so, you know, I'm just saying, you know, she is a star here.
[50:35] SPEAKER_02: I'm just so blown away by this paper.
[50:37] SPEAKER_02: And by the way, the word of the day of the week is Aladdin Uli, okay?
[50:42] SPEAKER_02: Yeah, yeah.
[50:44] SPEAKER_02: William Bauer has a question.
[50:45] SPEAKER_02: This added to a snapshot of a patient, tremendous chance for prevention and treatment.
[50:50] SPEAKER_00: Yeah.
[50:51] SPEAKER_02: Yeah, we agree.
[50:52] SPEAKER_02: I think so.
[50:54] SPEAKER_00: I think so.
[50:54] SPEAKER_00: You know, I only schedule the longest I'll go out sort of with a patient is one year, right?
[51:00] SPEAKER_00: I want to know what's going to happen to them and have the opportunity to make
[51:02] SPEAKER_00: an intervention dynamically.
[51:05] SPEAKER_02: What's the one thing today to lower risk of heart disease?
[51:09] SPEAKER_00: Go for a run.
[51:10] SPEAKER_00: Go for a bike ride.
[51:11] SPEAKER_00: Go out and do some exercise.
[51:13] SPEAKER_00: It's good for your mind, too.
[51:15] SPEAKER_02: And thanks for the comment by Merrick about thanks for bringing it to us.
[51:18] SPEAKER_02: What I want to just say in closing is in Ground Truths,
[51:22] SPEAKER_02: we like to get into things that people don't know about,
[51:25] SPEAKER_02: things that are not in the media.
[51:27] SPEAKER_02: I didn't see this in the New York Times.
[51:29] SPEAKER_02: It actually belonged in the New York Times or other science magazine, whatever.
[51:34] SPEAKER_02: It deserves to be there.
[51:35] SPEAKER_02: But What we try to do is find,
[51:38] SPEAKER_02: you know,
[51:38] SPEAKER_02: really interesting stuff,
[51:40] SPEAKER_02: great researchers and scientists we're talking to today and bring it,
[51:45] SPEAKER_02: you know,
[51:45] SPEAKER_02: break it down so that everybody can understand because this is a big deal,
[51:50] SPEAKER_02: also known as BFD.
[51:53] SPEAKER_02: All right.
[51:54] SPEAKER_02: So I think that's about it.
[51:56] SPEAKER_02: If there's any other comments or questions before we wrap up,
[51:59] SPEAKER_02: but I can't thank you enough for what you're doing, Sarah.
[52:03] SPEAKER_02: It's exciting.
[52:04] SPEAKER_02: If you ever want to relocate to script research, we'd love to have you.
[52:09] SPEAKER_00: Thank you so much.
[52:10] SPEAKER_00: Thank you.
[52:11] SPEAKER_00: Thank you so much for inviting me.
[52:13] SPEAKER_02: And Jennifer, exercise, yes.
[52:16] SPEAKER_02: Number one thing.
[52:17] SPEAKER_02: The number one thing on the list you can't get enough of.
[52:20] SPEAKER_02: So thanks a lot.
[52:22] SPEAKER_02: We'll get this posted in the days ahead and we'll get everybody else,
[52:26] SPEAKER_02: you know,
[52:26] SPEAKER_02: the annotations and, you know, more things so that everyone can understand this.
[52:30] SPEAKER_02: But I think what I like about it is,
[52:32] SPEAKER_02: since you are this math wizard,
[52:35] SPEAKER_02: that you can discuss this so that everybody understands it.
[52:39] SPEAKER_02: You know,
[52:39] SPEAKER_02: you didn't have to use any As far as I can tell,
[52:41] SPEAKER_02: yeah, there were some fancy terms, but you explained well.
[52:45] SPEAKER_02: So besides all these other qualities,
[52:47] SPEAKER_02: you have the great explainer attribute, which is really good.
[52:51] SPEAKER_00: Thank you.
[52:52] SPEAKER_00: Yeah.
[52:52] SPEAKER_00: I'm from the Midwest.
[52:53] SPEAKER_00: You got to explain yourself.
[52:54] SPEAKER_00: You're always in trouble and you always got to explain yourself.
[52:59] SPEAKER_02: Thanks everybody for joining.
[53:01] SPEAKER_01: Thank you.
[53:02] SPEAKER_01: Thank you so much for inviting me.
[53:04] SPEAKER_02: Thank you.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.