JUSTIN NORDEN: Welcome back to the Stanford Healthcare AI

podcast, and so thrilled to be joined by Dr. Jessica Mega,

who also is on faculty at Stanford, practicing

cardiologist, was chief medical officer at Google

and co-founder chief medical officer at Verily.

And most of you have already probably seen

her profile before.

We're just so thrilled to have her here today.

JESSICA MEGA: Thanks.

It's amazing to be here.

And the work you're doing is really

helpful for the community.

JUSTIN NORDEN: So maybe just to kick things off

and we're already starting to chat about this as we were

getting on, but adoption continues to rise,

and we continue to see more--

it feels like every month on these topics

where models are going.

And so Matt, help ground us here to a few of these data

points that we're looking at here around adoption, downloads,

usage, and just help ground us.

MATT LUNGREN: Yeah.

I mean, I know we have this thread going

as part of a longer conversation between the different episodes.

But, I guess, I'm continually amazed.

I think those of us who spend a lot of time thinking

about this maybe sometimes assume that, OK, everyone

is aware and everyone's already been using the models

and has a formed an opinion.

And when you see these continued spikes in the usage

keeps going up.

The number of downloads seems to be

sustained, at least for the last month at a million a day.

And then you're seeing this asymptotically increasing spike

in users.

I guess, it reminds me that, hey, maybe we are pretty early.

And I do feel like once in a while

I'll bump into somebody who is just really digging in as

opposed to just like the casual, hey,

oh, my kid used it for homework one day,

or I tried to write an email with it one time.

Now, it's actually a part of my daily routine

for a lot of different tasks.

I don't know what you all think.

I mean, I don't know if I'd call myself a power user per se,

but I would say that I spend a lot of time

trying out the models in different ways

and particularly testing them in various health care scenarios.

But I guess, I sometimes make the maybe

it's the wrong assumption that everyone is already doing this.

And I guess this kind of data shows me

that the classics were early still.

It feels like we're still early.

I don't know.

What do you guys think?

JESSICA MEGA: What I find interesting that you brought up

is there's a range of use cases, and there are a range

of groups who are using it.

And what I've found is when people start to use these tools

and it provides value, it's like this amazing flywheel.

And it's one of those things where I know personally, I

started around what I would call more traditional use cases,

let's say helping with a letter or an email.

And then you said, OK, can you translate that?

That gets you hooked on to another type of aspect.

And then I've gone from what I would call more standard tasks

to much more creative tools.

And I think, where the hook is.

And one of the observations I've personally made

is I think sometimes we're not always

thinking about getting the right tool for the right fit.

So if you want to add two numbers,

then a calculator is really good.

Grab that calculator and stick with it.

When you want more creative ideation,

you want to brainstorm around a concept.

I started to work with some of the models around a talk

that I was going to give, and it was

one of the best brainstorming tools that was out there.

And so I think we're going to get

better and better for the right tool at the right time.

And this is both in the world broadly

and particularly in medicine.

So that's what I'm seeing as these tools are being delivered.

JUSTIN NORDEN: We'll talk more about the medicine aspect.

I mean, we could talk about the consumer uses of AI.

I'm using translation now, speaking to my mother-in-law

in Vietnamese back and forth.

But let's ground it in medicine where

are you-- because you're still practicing, as we were just

talking about before.

Are you starting to use any tools

like this in your practice?

Actually, Stanford just launched their own with their EMR.

There's a few other companies in places

where that's starting to go.

Have any of these made it into your practice yet?

Have you started to see any of these tools or in your mind,

are they still a little bit too early?

JESSICA MEGA: Yeah.

So what I'll do is I'll give three examples,

sticking with the line of thought

of right tool for the right moment.

So when we are doing things where it's more rules based.

So something like drug-drug interactions

that we all know that are out there.

We do a pretty good job of trying

to remember our pharmacology, but those

are very basic rule-based algorithms that are

out there or, for example, risk scoring when we're

trying to decide whether someone needs to have

other testing before surgery.

We don't want anything that deviates from the rules

that we know.

So that's something I think we all hold hands and say,

this is tools that we've agreed on,

and we don't really need to iterate on those.

And then we have this second bucket.

And this is where you're seeing most of the tools that

are being approved by the FDA.

There were about 220 approved AI tools

in predominantly cardiology and radiology, where--

and I'll just give you an example,

something that I was able to work

on where we're looking at fundus images

to help patients with diabetes get triaged for care.

And as it turns out, these models

are really actually better than an individual reader.

And that shouldn't surprise us.

We're using deep learning.

We're doing multiple training sets for very discrete tasks.

And I think that there is pretty universal acceptance when

we feel like we understand if something is safe, effective,

and valuable, we'll take it.

The places where you're starting to see the most acceleration

with more of the generative tools

are actually on two very different ends of the spectrum.

And this gets back to my concept around creativity,

this idea of looking at patterns and trying

to predict new patterns.

And we may get into more detail around this,

but one is really thinking about biology broadly

and understanding, for example, how proteins work in the world,

thinking about new drug development.

And then, also, working with patients,

thinking about how do we deliver information

at the right time in the most interesting way?

So it's this spectrum between rules-based to the most

creative-based mechanisms that we need.

And so those are three different places which

today, these tools are being deployed.

MATT LUNGREN: Well, I'll push back a little bit on that

because I totally agree with you on each of those categories.

But what I'm starting to see is like as a user of all

those tools, do you have to have five different UIs?

Do you have to have a bunch of tools on your desktop?

Or are we now, again, to catch another buzzword

to the agentspace, where I have a model that

understands my intent well enough that I can just

talk to one model?

And then each of those applications

that you reference are now tools for the model

to then bring the relevant information

from those narrow models back to me.

And I can still have that.

I think this is the tension I'm starting to feel now where,

again, this accelerated pace, the general interest in,

hey, I've really used this conversational easy UI approach,

and I've gotten a lot of value out of it,

as you've said in your personal life.

Now, you're thinking about, well, maybe I

don't want to open up the calculator app,

or I don't want to open up a risk scoring app.

Just go find that for me.

And do you feel like that is a logical kind of progression

to this because those 220 now probably going

to continue to expand FDA cleared deep learning

solutions like, do I have to find a way to cobble them

all together to get the benefit of each of those innovations?

Or can I just abstract that away and continue to have a, quote

unquote, co-pilot or companion of some kind that's able to--

and then there's a lot of downstream challenges with that

is that really part of the way that they

were intended to be used or built, et cetera, et cetera.

But just in a general way, maybe I'm lazy.

I just want to work with one thing.

And I just want to have it do those other things for me

so I don't have to use my limited brainpower.

JESSICA MEGA: Well, OK, so no one's

going to call you lazy today.

But what I can say is the places where

you're seeing the most traction are places where these tools are

embedded into the workflow.

So we talk about scribe technology a lot.

And it's really for people who have used it

and have had positive experiences.

The acceptance and adoption is going up.

And the reason why is it's built into your day-to-day workflow.

The same thing is true for the number of approved tools

that I talked about.

The ones that are getting traction

are built into the workflow of a given cardiologist,

radiologist, and soon many other fields.

And so right now, we have to think

about how do we make it simple for not just

UMed, but for everyone.

Going forward, I think, you're right.

There are going to be platforms.

It's going to be more of the platform technologies

that allow this to be much more seamless.

No one's going to log in to 100 different portals.

It's just it's unrealistic.

So I think we look to see where the adoption is

and how do we make interoperability

more of a reality so that we can actually

take good care of our patients.

I would actually take it one step further

and say that if we're really ambitious

and we start to think about who we're actually caring for,

we're caring for real people.

We're caring for patients.

And as much as I may have a relationship with someone,

they may ultimately go to a number of different hospitals.

And they may-- not only may, but what

happens in the four walls of the hospital

is a very limited moment in time for that individual.

We have this idea that someone's healthy until they're sick,

but that's just not the way the world works.

So if you really wanted to plant a true North Star,

it's thinking about that individual.

And how do we create tools that follow them

throughout their longitudinal life?

So again, we know where we are now.

We know where we need to be.

But I just continue as we all go through this journey,

think about the people we're serving and how we bring down

the fragmentation in health care.

JUSTIN NORDEN: So there are four different things

you said that I just love one.

Totally agree on the platform nature of what's coming.

There's no way we get to the health system of tomorrow

with 40 different widgets or avenues to do that,

but I would also agree we're nowhere near that yet, right?

Even the most advanced ChatGPT, the most adopted thing,

you still have to pick which of the 10 different models

do I actually want to talk to.

It's a terrible user experience, but everyone else

has just copied it because that's

the best we've come up with so far.

So someday we'll get there on that.

But the consumer aspect, Jessica, that you talk about,

we're seeing echoes of this actually on the policy side

as well.

We're seeing the wearables push from RFK.

We're seeing interoperability and clear around data.

And I'm curious, as you think about this consumer push,

especially from your lens at Verily,

like you were pushing wearables, other technology

had the resources to think very beyond traditional health care.

So I guess, maybe what you can share,

take us through what's your vision

of that consumerization of what might be possible with AI

in this?

What did you get to work on before?

What did you feel like you had success in?

What was hard?

And how does that relate now with the current set

of AI tools?

JESSICA MEGA: Yeah, so I'll walk through an example that I think

is illustrative of where we could go as a system,

and it goes back to putting that person in the center

of their health care.

So one area that we were able to focus on

and there continues to be work on

is with cardiometabolic disease.

And we know that the number of people with diabetes.

If you look not only in the United States,

but you look internationally, has really been on the rise.

And so the question is, how do you

get the right care to patients?

Because we certainly don't have the number of endocrinologists

that we might need to support people, not only with diabetes,

but even pre-diabetes.

And so it started off with thinking about certain tools

that people might need.

And one of those tools could be a continuous glucose monitor.

Not that everyone has to wear it all the time,

but the insights around your own biology.

The fact that we call type 2 diabetes one thing

is, when you look at the data, some people

are having huge excursions.

They have hypoglycemic-- hypoglycemic,

they're up and down and some people ride.

There's clearly heterogeneity.

And so we were able to work in partnership with Dexcom

to create new continuous glucose monitors that

had a different form factor, were easier to use,

factory calibrated.

You then pull that information in and you show it.

So let's say the three of us go out to dinner

and we're playing afterwards.

We play cards, and we all have a cookie.

It turns out, we respond very differently to that food.

And so that information is surfaced,

yes, to a clinical support system,

but it's surfaced to an individual who's then empowered

to help through generative AI.

Think about what are some choices that people can use.

And then behind that, there's a clinical care team.

So this continuous understanding of information coming in

and action coming out is really a fundamental principle.

And I think about if you want to be really

simple with a lot of the work that we're

doing with AI and beyond, what information do we have

and how can we make a difference and make an action?

And so we were able to run clinical trials that

showed these tools actually reduced hemoglobin A1c.

And that's the metric we measure medicines based on that metric.

So that's just gives you an illustration of what

I think is possible more broadly across many different

conditions.

Again, it's the human machine interface,

I think what made it particularly useful.

MATT LUNGREN: I mean, that kind of ties

back to what you were saying before too, which

is that we see our patients in these episodic care instances,

and we don't have that visibility in a lot of cases,

even if they are being monitored of what's

going on in the meantime between the visits.

And further, going beyond glucose, which I think

is obviously an area where there's a ton of opportunity.

But outside of that, there's things like sleep,

and we know there's behavioral, and social issues.

All these other things that we know impact outcomes

and the effectiveness of the things

that we do for our patients.

It does make you wonder like, is there

a way to start to capture some of that longitudinal, all

those other what you would call data exhausted.

I think a lot of people call it.

But the things that happen in between,

did you find that kind of regular interaction

maybe is part of a clinical trial.

So it's somewhat artificial in the sense

that maybe that's not the routine.

But nonetheless, that must have also helped.

And then taking that a step further, Justin, you

have some graphics on this, but we're seeing patients already

kind of putting their medical data into the models

that the consumer facing models having these longer

conversations about their health care and some of the decisions

they're making.

Is there a place where these start to blend together?

I mean, we've learned from the data,

like, the work that you've done, that

having that other information that more granular,

we can take better care of our patients.

How do we bring these together?

And here's an example.

This is from a Reddit thread.

So take it for what you will, but it's someone

and you see many, many, many versions of this story.

I went to the hospital.

I felt a little weird about the diagnosis that they made.

I took my data.

I had a long, two hour long chat with the model,

and it came up with a completely different answer,

and I went back and then they confirmed it.

There's so many versions of these stories.

Now, obviously, there's sampling bias

and there's all kinds of challenges

with taking a lot of this and making big decisions about it.

But I do feel like it's a trend.

And if I'm being perfectly honest with you all,

I do have conversations with the model about my own health

decisions and data.

And honestly, I do just partially to push it,

and partially because I'm not an expert in a lot of aspects

of medicine for sure.

And so, is it helping me?

I think it is.

But then how do I translate that to my actual care team?

What's that future going to have to look like in order

to get the benefits of both?

JUSTIN NORDEN: Well, I can jump in with a few more data points.

And then, Jessica, you're up with the answer

for what the future looks like in the care team

and everything that happens.

But the 800 million active users from ChatGPT

from different discussions, 5% to 10% of those queries

are health-related.

That's a wild, wild number.

And I know we had Google searches before where people

were bringing in information.

But I think to all of us, it feels different.

I do it too all the time to understand more

about a condition or a symptom or something that comes in,

and it's way better.

And for all the studies we've looked at,

and we won't show you more charts today

of performance outperforming doctors or outperforming doctors

with AI.

But what is that care team look like?

So the easy questions for you, Jessica.

What does that look like?

JESSICA MEGA: Yeah, sure.

Softball.

So I have two thoughts here.

One is how do we handle this information back

to the information action?

And then we'll get to Matt's other question, which

is what does it mean to have multi-dimensional data

in the real-world that goes beyond maybe what we just think

about as the health layer.

OK, so going back to the first one around information.

You use the percent that it's about 10% of people

are using certain--

whether it's ChatGPT or otherwise,

they're using these tools for health information.

And the reality is that that's going to go up.

And we saw the same thing with search and other things.

So the piece where I--

I'll give you the optimistic view and then a few wrinkles.

So the optimistic view is let's-- so if the three of us

were sitting around in the 1600s,

and we had a patient and we were worried that they might be

anemic.

We might take maybe some of their blood

and put it in a tube, and we would shake it

and we would look to see, try to estimate a hematocrit.

And nowadays, I think it's great that we have cell counters,

and we don't question them at all.

And we're not sitting around trying

to estimate the hematocrit.

And we're not debating about the width of the cells.

We're not looking to see if the MCVs are different.

And so we trust these tools because it goes back

to them being safe, effective, and valuable.

And as long as the information that we

glean using different generative AI tools is of that caliber,

we should lean into that.

Because you know what that does?

That makes us action agents.

So the health care system, it turns into superheroes.

So we don't have to own all the information,

but we figure out what do we do with that information?

Who, for example, we trust troponins?

People come in with a heart attack and their troponin

is elevated, and I know they're going to benefit

from a cardiac catheterization.

We use that information--

again, whether it's hematocrit or troponin,

something that comes from an image because we know it's

actionable.

And so as long as the information

we're getting from generative AI provides

that level of information, I think we should embrace it.

The wrinkle that I talked about, and I was just

doing an assessment for myself looking at cardiac biomarkers

to see how accurate is the information.

It varies.

It varies.

And so we need to just get to a point

where we feel comfortable with that information.

But I have no concern in the same way.

The stethoscope was a big deal several hundred years ago.

People thought it would take people away from the patient.

But we work with MRI scanners.

Those are pretty talk about technology.

I mean, that's a big piece of technology.

So the fear shouldn't be in the technology.

It's really what we're doing with it and can we rely on it.

So that's how I see the human machine interface

and where I see action being so important.

OK, on the second one, I think we

do underestimate what health information really

means for a given individual.

And sure, I started a lot of my career in clinical trials

and working with genetics and pharmacogenomics

in this world of precision medicine.

So what biological features impact your health?

And we can talk more about that if that's of interest.

But over time, the aperture opened,

and I started to really think about precision health

and what really is important in someone's health journey.

And in some ways, whether someone has diabetes

or they have a cancer diagnosis.

There's a longitudinal journey that they are on.

And that's where Matt, to your question,

when you think about that health data,

that's where people estimate a 1,500 fold increase in health

information that can be very actionable on top

of the diagnostics and therapeutics.

So again, those would be roadmaps

for both of these tools.

Get information that we believe we can lean on in the way

that we lean on all of the assays that we do today,

and contextualize health not in a silo, but in the world

that a given person is living in.

So that's what I see as the true north.

MATT LUNGREN: I think that's a tremendous vision.

I think it does tie together the learnings that we've

had over the-- right, we've been chasing precision medicine

for so long, and we know it's like there's

a lot of pieces on the board, a lot of learnings tying this

together, maybe this is a way to do that.

And maybe the technology maybe is

part of the information asymmetry problem

that we faced in the history of medicine.

We had to spend 10,000 plus hours to learn,

and then we had to figure out how to translate that back

into a relationship with our patients to go on that care

journey, to get to the diagnosis or to the treatment.

And I feel like the information asymmetry

may start to be leveling off with the availability

of these tools to have a longer conversation.

I have a 15-minute slot.

I'm not going to be able to really get

to the end of every one of the questions of the concerns

that patients may have.

I may not be able to even surface them in that time.

But is this a kind of a companion technology

that can flatten that difference where we're on the journey

together, and frankly, some of the things that we know and we

want to do together are clear in both the patient

and the physician's mind.

That's the ultimate state is getting to the place

where you are literally side by side with your patient going

through the care process.

And there's always that feeling today,

at least in my experience, where I'm not

100% sure we're both on the same page at all times.

JESSICA MEGA: Yeah.

Well, it's so interesting.

If you think about even medical training,

we end up specializing because the only way we can really

wrap our head around the information is in many ways,

knowing more and more about less and less,

really being the world's expert in a given area.

And some of these tools may liberate us in that regard

and even get us to start thinking

across different disciplines, right?

It's interesting if you look at the field of inflammation.

Inflammation touches work that rheumatologists do.

So whether someone's seeing someone

with lupus or rheumatoid arthritis, it affects oncology.

It affects cardiology.

But I think there's actually a moment

for a much more expansive view of biology

as we maybe take off some of the constraints

that we needed before to really specialize.

And there's a book called Range by David Epstein,

and he talks about how potentially really knowing

an area, but having a generalist mindset can in some ways

create the most creative solutions.

So we'll see where that piece leads us.

But we have to keep an open mind about this idea of let's

continue to move health forward.

Because I think every person, probably all three of us

on the call today, every person listening to this,

knows someone who has a health condition where we don't have

a solution for it right now.

So there's a lot of work for us to.

I have no fear about being replaced or not

having enough work to do, but I want

to make sure we're focused on the best and highest problems.

JUSTIN NORDEN: Agreed on that.

And speaking of understanding the best and highest problems

places where we don't have treatments,

this is neither Matt nor I's background spending a ton

of time on drug development, new areas.

But this is something on across genomics

and personalized medicine that you've lived for years.

So I guess, tell us what excites you?

What parallels do you see with the current revolution

if you want to call it and these tools to perform with genomics?

What excites you about how we can now apply

these tools for new treatments?

JESSICA MEGA: Yeah.

So in my own personal journey in life

has always been about how do you get

new technologies to the hands of patients and clinicians?

And that started off first, obviously

as a clinician, but doing clinical trials to understand

how we get new medicines into the hands of those people I

talked about and studied a number

of antiplatelets and anticoagulants, which

at the time--

so now when I round with residents and fellows,

some of these things seem like no-brainers.

But we really were thinking about

do people need dual antiplatelet therapy

or how do we treat someone after a statin?

We didn't even really think that lipid lowering was going

to be as effective as it is.

But over time, learning about clinical trials

and then realizing if all three of us were in a clinical trial,

we probably don't respond to these medicines in the same way.

So built out capabilities with the TIMI Study Group

around studying genetics and genomics.

And I actually see a lot of parallels

between that early work, and what

I'm seeing with the deployment of AI and other decision support

tools.

And so the lessons that we learned

in the early days of genomics, and you probably recall this,

there was a lot of association studies.

So people were associating this particular snip

with this finding, and everyone was really excited.

And as it turns out, only about 25% of those studies

could be replicated up front.

So it was this-- it almost reminds

me a little bit of what we're talking

about with hallucinations.

It was different because it was a statistical problem.

It was essentially this multiple hypothesis issue.

But we had to wrangle with what is replication look like?

What is validation look like?

What do we trust?

How do we run a really good genome wide association study?

And then we also had compute go up.

I remember the days we had this external hard drive

and I'd bring it over to my colleague,

and then he'd work on it overnight,

then he'd bring it was this red.

I remember it was red maroon drive.

And then I just kept thinking, gosh, what if we lose this?

There's biological insights on this drive.

So, double rapid.

Anyway, so there was this combination

of understanding statistically how to evaluate the genetics

and then understanding and benefiting from compute

because we went from where we compute on-prem

to how we compute in the cloud.

And over time, we now are starting

to see, particularly in the world

of oncology, genetic-based therapies.

And we're really getting a much better sense

of targeting the right patient with the right therapy.

And so you might ask, well, OK, what

does that have to do with where we are today with AI?

Well, it's a very similar thing of what can we replicate?

What really stands the test of time?

How do we use compute to its best and highest value?

So nowadays, things are becoming really fast.

And so how do we make decisions that used to take a lot of time,

make them more real time?

And then genomics is a static data.

How do we integrate real-time information?

I gave the example of glucose, but that could be said for gait

if someone has Parkinson's.

There are so many examples.

And so I think many of the parallels

we learned years ago we can apply here.

And it's just an area of great interest to me.

And then you ask, so what's most exciting?

Well, the thing that's most exciting to me

is much in the way we use genomics

to continue to explore biology and get

the best care for patients.

I think AI is going to, as I already mentioned before,

give us just a better understanding beyond a given

discipline.

Biology is very experimental right now,

and if you talk to people who are

in physics and other disciplines, they have laws.

They have Newton's law and thermodynamics.

And biologists are really--

we're hoping we can get to that same place using

a lot of these tools.

And then on the personal end, getting the right information

to individuals.

So it's really, in some ways, technology

meets humanity on steroids.

So it's a daunting time.

But I almost want us to hold hands and say,

we have encountered things like this.

And let's learn from those experiences.

MATT LUNGREN: I love this vision.

I'm curious on as part of that as if all of the experience

that you've had--

again, I totally resonate, by the way,

with the days of the hard drives and trying to run to the lab

and, wait, is this the--

am I going to lose this data?

This is my whole career right here.

But what I feel like I'm seeing and, again, this is both Justin

and I believe are admittedly not as deep in the space.

But if we agree that there's opportunities for new drug

targeting, new drug discovery, small molecule development

at scale that can be at least more intelligently

driven to have a higher likelihood

of success in the next phases.

What I wonder is like, does that mean we have to change how we

derive the, quote unquote, "evidence"

in the phase II space?

Because what I assume is going to happen,

or it may already be happening, is

that there's a ton of now candidates that

are super promising, but they still

have to go through the same conveyor belt process for phase

II that have all those same problems

that we've always had finding the right patients,

filling the trials appropriately, dealing

with all the paperwork and the regulatory aspect.

Those feel like ossified, challenging, but necessary

processes.

Do you feel like we're going to have

to think about how we do clinical trials in order

to see the benefits of all the discovery happening today?

JESSICA MEGA: Yeah.

So as you said, to boil it down, there's the target,

there's the drug, and then there's the trial.

And so there are going to be examples

that people turn to are looking at certain diseases

where we thought there was going to be resistance

or antibiotic resistance, and now people

are actually finding particular compounds that may be effective.

What people said, oh, these certain areas are undruggable.

You can see that also in oncology.

And now people are coming up with new solutions.

So your very prescient question is--

OK, so potentially, we find new targets.

Ideally, we find better targets.

So we know that there's about a 10% success rate

as you go through that pipeline.

So even going from 10% to 20%--

and let's keep it, let's say, clinical trials say ossified,

which they won't, and I'll get back to that.

Even doubling the success rate by understanding the biology,

both the toxicity, as well as the efficacy upfront

is going to be helpful.

But then we want to marry that with a more effective

clinical trial system.

And I believe you're talking to someone who

trained as a clinical trial.

So clinical trials and randomization

really is important.

We have been burned by medicines that we just

knew we thought biologically they were going to work out,

and I can give many examples of those.

But we can do things more efficiently.

And I'll give you real-time examples.

So the first thing is just patient recruitment reaching out

to patients, let people know that there are trials available.

Once someone understands a trial, how do we--

it's like a fabric.

How do we connect them to the sites?

Right now, it's a pretty rudimentary process.

In fact, we usually think it's success

if it's one patient per site per month.

We are seeing examples where we're doing so much better.

Again, it's just connectivity.

Then in terms of understanding the data as

it's coming in, doing real-time data monitoring,

looking for anomaly detection that used

to be a very manual process.

So we would do source document verification.

So the three of us would sit there

and we would look at this data.

Is it going over here?

Now, you can start to look at the data coming in.

Pick up pockets of variance where you could go

and drill down.

And then finally, regulatory submissions.

It's a really great use case where

people are collating the information, things

that used to take months to even a year,

now taking a much shorter period of time.

It's a really nice use case of some of the generative AI tools.

So I think there's going to be progress on all of those fronts,

but I don't--

in fact, I wouldn't advocate getting rid

of clinical trials and real testing of biology.

But man, are we getting smarter?

And so that's the piece.

Again, we have to decide if we're a full or half empty.

But I think when you approach a problem

and you're able to see multiple pieces that

will line up to lead to efficiencies,

that's a good place.

The other thing is we're seeing real progress in all of these.

It's not, gosh, wouldn't that be nice?

I always look at how rapidly we're learning.

And the same thing can be true of models,

as we're starting to look at AI models.

Are they getting better?

Are they getting smarter?

It's almost like surrogate markers or surrogate endpoint.

So I can tell you in the space that we're talking about here,

things are getting better.

JUSTIN NORDEN: And it's not just the pilots

of even the FDA starting to rethink its own process.

They announced their own internal AI tool

a handful of weeks ago to start to take,

again, processes that used to take months down

to hours or days.

I think there is hope for optimism here.

And I agree with you.

It is very easy to get burned by associations.

And just the rigor of clinical trials, which

hopefully doesn't go away.

I know we're just about at time for what we have together here.

But you just have such a unique perspective

on having seen these changes come

across medicine and technology.

So like, what's your crystal ball?

What is three to five years look like?

And we'll just bracket it towards clinical medicine

and then we'll wrap.

JESSICA MEGA: Yeah.

So what I would say and, again, I hearken back

to some of the lessons we learned around the applications

of biomarkers and genetics and genomics

is we get a better handle on where

that information is most useful, information that is actionable.

We're going to see the same thing going on as we think

about AI and generative AI.

And as I look to the future, we're

going to have a better sense of what

we talked about at the very beginning, the right tool

for the right job.

And you could imagine moving forward

as we get better at the information,

as we link that to the ambient scribes,

are there clinical tasks that could be done in a way that

is more streamlined that gives us in the health care

professions more time to think about the new treatments,

spend time with our patients?

Absolutely.

But the call to action is in the reason

why people take new medicines and the reason

why people get percutaneous valves,

and the reason why people get genetic testing for triage

of their chemotherapy is because we believe in that data.

And so I think the next three to five years,

we're going to see a real acceleration.

It reminds me, again, you can turn to

and say, oh, there's hallucinations here or there,

but I think we're missing the bigger picture of where

we're seeing the biggest benefits and some

of what we talked about.

And let's think about the tools where much

like we talked about the cell counter and the complete blood

count.

Let's use the tools really well and let's free ourselves up

to better humanity and think about these broader issues.

So that's the big picture of what I see.

MATT LUNGREN: And I feel like this is exactly the North Star.

And this is part of the reason we do this podcast too

is, I think, we really want to try

to keep this conversation going and encourage folks to lean in.

There's never been a better time to--

and there's never been better access

to some of the most powerful tools on the planet.

Having an intuition of where they work, where they don't is

only going to benefit all of us on all sides of the table.

And I am glad that you didn't use the 1,600 analogy of how we

used to check for diabetes.

So I like the book.

[LAUGHTER]

JESSICA MEGA: We can do that.

We can do that next time.

But I think that point, you just got me super energized.

The people who are closest to the problems

should really feel empowered to figure out

how to use these tools.

And so it really is.

It's a moment to say, hey, as you walk around, whether you're

a researcher or a clinician, someone who is working in health

tech, look around at the problems you're trying to solve,

because that's where the real magic comes.

So it's a moment.

JUSTIN NORDEN: I think with that, Jessica, thank

you so much for coming on.

JESSICA MEGA: Thanks