Latest / AI for Founders with Ryan Estes / Inside the $50M mission to fix clinical evidence

Transcript

Inside the $50M mission to fix clinical evidence · AI for Founders with Ryan Estes

0:00
Welcome back to AI for Founders. I'm Ryan Estes, and today we're
0:03
diving into one of the biggest opportunities in healthcare and
0:06
artificial intelligence, closing the evidence gap.
0:10
My guest is Doctor Brigham Hyde, Co founder and CEO of Atropos
0:14
Health, the company behind Chat RWD, a generative AI platform
0:19
that turns real world healthcare data into clinical evidence in
0:23
minutes instead of months. Atropos has raised more than 50
0:27
million partners with giants like Merck and McKesson and is
0:30
tackling the mind blowing fact that only 14% of medical
0:34
decisions today are backed by high quality evidence.
0:38
This conversation is for founders, investors, and
0:40
builders who want to understand how AI can save lives, not just
0:43
improve margins. You'll learn how Brigham's team
0:45
built an AI engine trusted by doctors, how they avoid
0:49
hallucinations in high stakes environments, and what it really
0:52
takes to scale an AI company that blends science, speed, and
0:56
ethics. If you're curious about the
0:58
intersection of medicine, machine learning, and market
1:00
opportunity, this one's going to light you up.
1:03
And if you love this podcast, you're going to love the
1:05
newsletter. Go to aiforfounders.co, sign up
1:08
for that today. This episode is brought to you
1:11
by Kitcaster, the easiest way to get founders booked on podcasts.
1:14
If you ever wanted to Share your story on shows just like this
1:17
one to connect with new customers, investors, or talent,
1:21
Kitcaster makes it happen. Their team of expert agents
1:24
pitch your message to curated podcasts, get you booked, and
1:27
even prep you for each appearance so you sound like the
1:30
pro you already are. Over 800 clients have used
1:33
Kitcaster to become industry influencers, generate millions
1:36
in revenue, and build authority that lasts.
1:39
So if you're ready to turn conversations into conversions,
1:42
head over to kitcaster.com and apply to join today.
1:46
Kitcaster podcasts are the new stage.
1:48
Let's get you on it. That's Kitcaster by MO Burst.
1:52
All right, all right. Coffee is great, but coffee
1:54
won't prep your meetings for you.
1:55
It'll just keep you awake while you panic, hastily scrolling
1:59
through LinkedIn at 6:59 AM. Ambient, however, will prep your
2:04
meetings. Each morning you get a founder
2:06
tuned briefing, sourced overnight from your calendar,
2:09
your meeting history, and the web.
2:11
It's like walking to every call with a CHEAT SHEET that you
2:13
didn't have to write. No more scrambling, no more
2:16
guessing. Just context, confidence, and
2:19
focus. And if you're raising money,
2:21
ambient investor dossiers are an absolute must have.
2:25
You'll know the investors portfolio, their interests, and
2:28
their latest moves before you even step into the room.
2:31
You'll even get a heads up on red flags.
2:34
Think of it as the caffeine upgrade that your calendar
2:36
actually needs. Prep while you sleep, perform
2:40
while you're awake. Sign up for free at ambient dot
2:43
US. Go check it out at ambient dot
2:45
US. That's ambient dot US.
3:05
Bring and welcome to the show, I sure.
3:07
Appreciate you being here. Thanks for having me, Ryan.
3:11
I want to talk about Atropos Health, your company and I want
3:15
you to introduce it with this amazing statistic right on the
3:18
front page of your website. So it says only 14% of daily
3:23
medical decisions are backed by high quality evidence.
3:28
That's. Actually one of the higher
3:29
estimates, believe it or not. And when I sort of say this,
3:33
that to people who aren't in healthcare, they're like, what
3:35
do you mean? The sort of nice way to say it
3:39
would be this is the art of medicine.
3:41
You know, doctors, amazing individuals rely on their
3:44
training and experience to make decisions, but the reality is
3:48
we're just not producing enough evidence to help inform them.
3:51
And I, I trust a physician, if you give them a, a well done
3:55
study, a publication peer reviewed, they can read it, they
3:59
know what to do with it. They can use it to make
4:01
decisions, but we just don't have enough.
4:02
And the reasons why maybe not surprising, but you know, we
4:06
just don't run enough clinical trials, right?
4:08
They're expensive and most of them you are funded by drug
4:12
companies to focus on very specific populations.
4:15
They also tend to include them the healthiest people because
4:18
they want to make sure the effects aren't confounded.
4:22
And you know, the stat I sort of always include after we say that
4:25
stat is, you know, if you look at the trials we do have, you
4:29
know, the 14% we do have evidence for, 70% of those
4:33
trials exclude patients with comorbidities and that includes
4:36
things like diabetes, obesity, heart disease.
4:40
Guess what percent of US patients, that is about 6070% of
4:43
the patients. So even the evidence we do have
4:46
was done on, I like to say white male triathletes and Memorial
4:51
Song Kettering, right, not on your average patient.
4:54
So all this is to say the purpose of our company is to
4:58
fill, we call the evidence gap. We need more evidence.
5:01
If there is good high quality evidence out there, doctors know
5:03
what to do with it and we built our company around solving that
5:07
problem. Incredible.
5:08
So how are you aggregating this, this evidence is are these like
5:12
are you funding studies? Are you bringing studies
5:14
together into one centralized location?
5:16
And then who is the audience ultimately?
5:18
Is it patients? Is it doctors?
5:19
Is it both? Yeah, to tell the origin story,
5:22
all this started at Stanford University and my Co founder is
5:25
Doctor Nigam Shah, who's the chief data scientist at
5:28
Stanford. And he began something within
5:30
Stanford called the green button, OK, green button, blue
5:33
button. They have meaning in healthcare.
5:35
And the idea was you press the green button and through
5:39
technology he built, they would run a brand new study in real
5:43
time or close to it at that point that would run on patient
5:48
data. So all this data that's been
5:49
collected about us and the HRS and things like that, it's been
5:53
being built up for the last 20 years or so.
5:55
But you know, doing something with it to inform a decision,
5:58
converting it from just data to evidence was the part that was
6:03
hard and needed to be automated. And that's the core of what we
6:07
do, is that conversion from data into evidence.
6:10
And what do I mean about that? What I mean is you have to do
6:14
the analysis the right way. There are techniques, methods,
6:17
statistical approaches, some AI related, some even more basic
6:21
that are just hard to do. Typically, even if you have the
6:25
data. And if you do them wrong, it
6:28
won't pass peer review. It wouldn't meet the standards
6:30
of of medical evidence. So doing it well and do it
6:33
quickly is what we enabled. And that ran within Stanford for
6:36
about 18 months. They had hundreds of physicians
6:39
use it. The ratings were extremely high.
6:42
And there were amazing stories in there.
6:43
I'll give you just one. There was a case for the
6:47
neurology department. They had a young boy come in.
6:49
It was about 12. He had distal nerve pain,
6:52
meaning tingling and pain in his limbs.
6:55
The immediate concern from the neurology team was, man, this
6:59
might be early onset Ms. which is rare in children.
7:02
But like the the signature sort of looked like that.
7:05
And the action to take would have been to do a Spinal Tap,
7:09
would have been to do a, a big imaging work up, stick them in a
7:12
tube. Parents of the child obviously
7:15
concerned about this. These are big procedures, scary,
7:17
dangerous. And one of the neurology team
7:20
members had access to our tool and said, wait a minute.
7:22
Before we do that, let's see if we can ask Atropos what happened
7:27
to patients with this presentation, what typically
7:30
occurred with them, maybe what treatments, what diagnostics
7:33
worked. We ran that study in a matter of
7:35
four hours of the time. We now do these in minutes.
7:38
This is five years ago. And they came back.
7:41
They found that across the data at Stanford, we now have much
7:45
more data at other institutions. But at that time in Stanford,
7:47
there's about 300 kids who looked really similar, and about
7:51
85% of them had had a latent viral infection within the last
7:55
two weeks. They had not confirmed that with
7:57
the patient and the family yet. They asked him and said, was he
8:00
sick, you know, over the last couple weeks?
8:02
And he said yes, actually, he had a, you know, a viral
8:05
respiratory illness. They gave him corticosteroids,
8:09
which would be sort of a normal treatment for that.
8:13
Within 24 hours, the distal nerve pain went away.
8:15
There was no work up, no Spinal Tap, no need for a $30,000
8:19
procedure to be done. And the, the family was
8:24
incredibly happy, right? Because want to relieve their
8:27
concerns. They had an idea what it might
8:29
be and they avoided something that might have been more
8:32
difficult for their child. They monitored for two days.
8:34
The symptoms went away. And it was a great success
8:37
story, by the way, success story for the patient, for the
8:40
physician team, and for the cost of healthcare.
8:43
All of those were big wins. Just the one example.
8:46
It's incredible. And you know, the, the siloed
8:48
information and the, the kind of medical industry is kind of a
8:52
known issue and always has been. So was the problem really the
8:56
the data being an aggregate or is it just so much information
8:59
it's hard to parse prior to what you guys are doing?
9:03
Yeah, I, I started my career as a, a bench researcher, got my
9:06
PhD at Tufts University in Boston in the early days of the
9:10
GM revolution. And I think we all felt in the
9:12
2000s, maybe showing my age there, that, you know, all this
9:15
data would be flowing magically. You also had the High Tech Act,
9:19
which, you know, created electronic medical record
9:23
mandates and we thought we're off to the races.
9:25
Well, you're right, it actually didn't go quite that way for the
9:28
1st 10 or so years. But the truth is now we have a
9:32
lot of data of a lot of different types, imaging claims,
9:37
EHR registries, patient reported data.
9:40
And over the last, I'd say 5 to 8 years, the silo problem has
9:45
started to come down. And that's because of two
9:47
things. One, cloud computing is is
9:51
everywhere and healthcare data is no longer locked in the
9:53
server in the basement of a hospital.
9:55
It's on the cloud, you know, private clouds, hybrid clouds
9:59
and that means it's available to compute.
10:01
The other thing is that we've done a good job handling
10:03
identity. So things like de identifying
10:07
patients, companies like Datavant have built a lot of
10:10
infrastructure to do this safely.
10:12
And that brings us to companies in our generation which are
10:14
really more focused on the application layer.
10:16
One of the things that we do is we work in a Federated way,
10:20
which I in healthcare, I think is the way to go.
10:22
Meaning I don't buy and sell data, I don't move data around.
10:26
The data never leaves the healthcare system where I think
10:29
it's appropriately stewarded. Instead, we're bringing
10:32
technology through the cloud directly to that data and that
10:35
opens up the silo problem. As I sit here right now, we have
10:39
over 20 different core data sets on what we call our evidence
10:43
network. These include health systems,
10:44
individual health systems and other data partners.
10:47
And I have over 894,000,000 patient timelines.
10:51
That's a data. And when you're in the
10:53
predicting game and you're in the AI game, that's great
10:56
substrate to be able to produce really high quality, high
10:59
performance models that enable these use cases.
11:02
We're talking about 100%. So one of your kind of marquee
11:06
products is it's called Chat RWD and it's a generative AI chat to
11:14
database interface. So it's basically, I'm, I'm
11:16
assuming kind of like a chat function for the data that
11:19
patients or doctors can use explain what the product does.
11:22
Yeah, back to the Stanford days, right?
11:25
We made some good choices then this before LMS had come out in
11:28
the way that, you know, the release of Open AI originally
11:31
that really took the world by storm.
11:33
But even then, we realized the user experience for clinicians,
11:36
if you made it any more difficult than a couple of
11:39
sentences like they're emailing or texting appear, they weren't
11:42
going to use it right? This is about user experience.
11:44
So we always had a text based input to our system.
11:48
When LLMS came out, as many companies at the early stage
11:52
did, we sort of said, hey, does ChatGPT do my company now we
11:56
tested that and for a variety of reasons both in the early days
12:00
and now, the hallucination rates associated with LMS and some of
12:04
the things that we do that don't allow for that level of error
12:07
meant that we felt pretty good, that we were OK.
12:09
But we did say, hey, look, these are great tools.
12:11
Why don't we build some infrastructure atop this to add
12:14
to user experience. So Chatter WD became the copilot
12:17
on the front end of our tools where a user could talk back and
12:22
forth to doing that. We don't have our own LLM there.
12:25
We're using a rag framework, which is something many people
12:29
have done that leverage the infrastructure of all these big
12:32
LLMS out there. So you know, we train and sub
12:34
train on our content, the open AI or CLOD, and you know, these
12:38
things get better and worse every month on the scoreboard.
12:41
So we built it to be flexible. As these tools improve, our
12:44
thing improves and we focused on using them for our user
12:48
experience and on our users needs.
12:50
And so now that's that Ms. example of the child with those
12:54
symptoms that would take approximately 20 seconds to get
12:57
that answer. And the doctor could do it right
13:00
in the ER, you know, just typing right into their phone.
13:03
It's incredible. Can can, Is it available for
13:05
patients as well? So we've not made it available
13:07
to patients yet. I, I think we work really
13:10
closely with the clinical community, like this is a big
13:13
change happening in how doctors practice medicine, how they
13:17
discover evidence and insights. We have had patients reach out.
13:21
What we say to them is, you know, please add your doctor to
13:25
the chat. We'd be happy to talk to them.
13:28
You know, at the same time, Brian, I think you're right to
13:30
ask that because I think we're in a stage of consumer
13:34
healthcare, be it longevity or other things like Epic has
13:38
announced a patient facing agent now that the patient can talk to
13:42
Ali Baba also has one Tempest in the oncology realm has talked
13:45
about it and I think the gut has been fired on the patient facing
13:48
application. I am certain that Open AI, Meta,
13:53
Claude, they're all receiving a ton of patient healthcare
13:56
questions. Our view is let's just make sure
13:58
we do it the right way. Let's get the clinical community
14:01
bought in, then we can think about how that user experience
14:04
could involve the patients. 100% and this kind of taps into, I
14:07
think a lot of frustrations maybe folks have with the
14:09
healthcare system. And I won't say a lot of people,
14:12
I'll just say from my experience, you know, and, and
14:16
that's to say that I understand and support doctors and they
14:19
have a hard task at hand, which is everybody's got their Google
14:24
PhD, you know, and so they're coming in.
14:27
Not only that, but they have pharmaceutical drugs being
14:30
advertised 24 hours a day on television, telling them you got
14:33
to ask your doctor about, hey, is your big toe hurt?
14:36
You should ask him about tolexra or whatever it might be.
14:39
You know, So with me personally, you know, I, there's all these
14:43
options now for preventative healthcare, for performance
14:47
enhancing, you know, I'm 47. So you start wondering like, OK,
14:51
what about SRT? What about GOP ones?
14:54
What about these things that I don't necessarily have any
14:57
pathology, but you get into kind of this Gray area with doctors.
15:00
So like you start coming to the doctor, like what about this,
15:03
what about that? Will you need to come in?
15:05
Well, for me personally, I prefer telehealth.
15:07
I don't want to go down and spend 1/2 day in the doctor's
15:09
office. So I guess all that's to say is
15:12
that on one side you have medicine is going to be
15:16
radically disrupted and changed by these new technologies and
15:19
new therapies and the smarter AI it's in finding new therapies
15:23
for us, it's going to change even more dramatically on the
15:27
other side. Some might say that we live in a
15:30
post truth society where people can take data and people do take
15:35
data and use it to fit certain agendas or ideologies or certain
15:41
forms. You know, now I don't know if
15:44
Tylenol is OK anymore. I don't know if the you know, I
15:49
don't know if the, you know, any of the, if the polio vaccine is
15:53
bad, like it's, it's very hard to bridge that gap from data to
16:01
what was it. It wasn't outcome.
16:03
It was data to evidence evidence.
16:06
So like this is almost kind of a story that needs to be told.
16:10
And collectively we have to have some kind of foundation of trust
16:13
and expectation. Is that part of the reason why
16:16
you're kind of limiting it to clinicians as opposed to, to,
16:21
to, you know, patients? So that basically you, you know
16:24
that there's some some. It's a really important, yeah,
16:27
really important question and something I'm I'm passionate
16:30
about. And let me, let me just tell
16:31
you, like I think it goes back to the first thing we talked
16:33
about, which is the evidence gap.
16:35
OK, If there was evidence, OK, the doctors could administer it
16:40
and maybe even we can build ways for patients to consume it in an
16:44
effective way. And I'm just giving you a dumb
16:46
example. Let's say we're in our 40s,
16:49
Ryan. Let's say we Instagram tells us
16:53
that an ashwagandha supplement is going to increase our
16:57
testosterone and prolong our, you know, healthy adulthood.
17:02
Well, if you go to your doctor and ask him about that, what is
17:04
your doctor going to say? Your doctor's going to say there
17:07
is no study out there that shows that this improves those
17:10
metrics. Now let's talk about why.
17:13
OK, it's not that maybe that's not interesting.
17:16
It's just there's been no funding to run a clinical trial
17:19
on ashwagandha supplements that could produce a peer reviewed
17:22
paper that your doctor could then review.
17:24
And when you get into the funding of what gets studied and
17:27
doesn't get studied, be that NIH and government down or be that
17:31
pharmaceutical and profit interest up, right?
17:34
It's just a reality that there's nothing for your doctor to go
17:37
reference. And the, the part about this
17:40
that I find as a conundrum and why we focus so heavily on
17:43
quality is because here's reality.
17:46
I have a PhD in clinical pharmacology.
17:48
I was trained in statistics and clinical trial methodologies.
17:52
The problem for the consumer is, let's say I gave you 2 studies,
17:55
one that said ashwaganda worked and one that said it didn't.
17:59
Do you as a consumer have the ability to thoughtfully review
18:02
the statistical methodology in a way to understand where bias may
18:06
come in? And you know, I, I think the
18:08
problem number one is there is no studies.
18:10
So let's get the studies. But once we get them, how do you
18:13
create that discerning? And I and I think right now the
18:15
reality is that's where the role of experts like positions still
18:19
lives. But if we had those studies,
18:23
could we build a user experience that translates that in a way
18:26
members can understand? I can tell you in some
18:29
partnerships we have looked at creating member facing
18:32
deliverables, being the patient could actually get one of my
18:35
reports. And when we do that, it's really
18:37
about UI and user experience and explaining complex statistics.
18:41
You know, it goes from being a 30 page report to A2 page report
18:46
with two key graphics on it, right?
18:48
That then citations so you can go back and review like those
18:52
changes have yet to totally be figured out.
18:54
And here's the risk, OK, in the world that I live in, which is
18:58
around real evidence and observational research, the
19:01
approach and method are super important.
19:03
If you do a study in an incorrect manner might show that
19:07
you know that the Sofani urea added to your diabetes regimen
19:10
would help lower A1C. But the potential bias of this
19:14
data is maybe that's just because everybody who happened
19:17
to be honest, Sofani urea in the data was healthier, right?
19:21
And you don't want to make a wrong conclusion like it's the
19:23
drug causing this that could confound things.
19:27
And you know, the the Tylenol debate has only started, but I'm
19:30
sure there'll be lots of good studies.
19:31
They'll peel apart those elements, right, to try and
19:34
determine causation, because that is the importance of these
19:38
statistical techniques in observational research.
19:41
You have to hold everything else the same and only change one
19:44
variable, the treatment, right? And you have to make sure if you
19:48
show an outcome change, like they get better or worse, that
19:52
it's because of the treatment, not because of some bias in the
19:55
data. So there's a hard problem.
19:57
OK, we're right on the front line that I think it's start
19:59
with creating more studies so that they're out there so the
20:02
experts can figure out how to deal with them.
20:04
And then it's about how do you change that user experience and
20:07
get it down to patients so they can factor into their own
20:09
consumer choices which are coming at them left, right and
20:12
center? Yeah.
20:13
You know, it's interesting kind of talking about bias because,
20:15
yeah, there's like confirmation bias you have that you want to
20:18
be right. There's the bias of like the
20:20
test itself. To your point, it's like when
20:22
they they studied, they're like, wow, people who smoke tobacco
20:26
have really high testosterone and they're like, yeah, well,
20:29
those people are more social generally.
20:30
And so it's not it's not causal. Yeah, When you get the causal
20:35
inference, which are the methods we use, you've got to know the
20:38
cause is the one variable in a therapy you might have changed.
20:42
And all the statistical techniques we use are designed
20:45
around optimizing to ensure that, which is really critical.
20:49
And how do we? How do we?
20:52
How do we got in a big statistical discussion here?
20:55
I love it, but it may be a little deep fervor.
20:58
The audience of my point is that method and quality matter when
21:01
you're making visions about somebody's life or death, and
21:04
we're trying to ensure that that layer flows in as all this great
21:08
automation and scale come with it.
21:12
OK, everyone loves to talk about vision, the pitch, the funding
21:15
round. But you know what gets 0 love?
21:17
The actual product build, that magical, terrifying phase where
21:21
nothing works, the docs are out of date, and you've asked your
21:24
dev team how's it going? For the 47th time today on
21:27
Slack, That's where Code Story comes in.
21:30
It's a podcast for startup founders who know that building
21:33
The thing is the thing. Hosted by NOAA Lab Heart, Code
21:36
Story goes deep with tech leaders who've been through it,
21:39
shipping real products, surviving messy launches, and
21:42
figuring out what actually works.
21:44
And when. You're on the edge of chaos.
21:47
If you're building now or ever plan to, you'll want these
21:49
stories in your ear Balls. Search code story wherever you
21:53
get your podcasts. Or visit codestory.co.
21:56
Because founders talk vision. But builders?
21:59
Builders tell code stories. OK, I've got a confession.
22:04
As a podcaster and founder, I'm constantly meeting incredible
22:07
people, guests, listeners, or other entrepreneurs.
22:10
But here's my little secret. I'm terrible at following up.
22:13
My inbox is basically a graveyard of missed
22:15
opportunities. Does that sound like you?
22:18
You meet someone amazing, exchange emails, have great
22:20
intentions, and then life happens.
22:22
Six months later, you're kicking yourself thinking I should have
22:25
stayed in touch with that person.
22:27
This is exactly why I started using Warm Start dot AI and
22:30
honestly, it's transformed how I think about my network.
22:33
Instead of letting relationships go cold, Warm Start analyzes my
22:37
existing contacts and helps me identify who I should be
22:40
reaching out to and why. It even writes the outreach
22:43
messages for me. But here's the key.
22:45
I review everything before it goes out.
22:48
I'm not just talking about this because they're a sponsor.
22:51
I'm talking about it because watching a healthcare executive
22:53
use this to build his entire Advisory Board in one weekend
22:56
before CEO sign ups made me realize I was leaving money on
23:00
the table every single day. Your network is your net worth,
23:04
but only if you actually nurture it.
23:06
Check out warm start. dot AI. Use code AI IFF to get 50 bucks
23:12
off your first month. Stop letting your relationships
23:14
turn into miss revenue 100%. You know and I but there is some
23:20
responsibility, some personal responsibility to be able to
23:22
check your own biases when you're viewing any datas you're
23:25
making any health decisions. Problem is, is it's we seem to
23:28
be like celebrating our ignorance of, of bias culturally
23:33
right now for some reason. What do you think is a, a, a
23:36
stab for that? Is it just basic scientific
23:38
understanding, the ability to admit when you might be wrong?
23:41
You know, because if you, if you did show me two studies about
23:46
ashwagandha and I'm already like, I don't like ashwagandha.
23:49
It's got a scary name. And then I'm going to point
23:52
directly to like, well, this one said no, you know, so right,
23:55
right. Look, I, I think you're
23:59
referencing it culturally and I think you're probably have your,
24:02
your finger on the pulse of something there.
24:04
I, I see a little bit differently as a technology
24:06
person, I think we have an explosion happening in
24:10
information, right? And it, we're dealing with it in
24:13
different parts of our lives. Media, for instance, like God
24:16
knows you don't need another Marvel movie, right?
24:18
Like there's just so much to process.
24:20
So now what's happening is we're figuring out how to sort through
24:23
and process that. So while it's expressing itself
24:26
culturally, I think you're on to something there.
24:28
I do think it's more like in the information step we're in, in
24:31
the information age, this is just a hard period to sort
24:35
through things. I mind the idea, though, that
24:38
we're blowing up old systems. You know, you're talking about
24:40
somebody who's trained in the old system, OK.
24:43
And I don't like the idea that a small number of people, experts,
24:48
corporate entities control what evidence gets generated.
24:51
Why can't we run an Ashwagandha study?
24:53
And the reason 10 years ago, 15 years ago have been it's too
24:56
expensive and there's no profit motive.
24:58
Well, if we can lower the cost of running these studies, then
25:01
we can create good ones. Then we'll figure out a process
25:04
through it. So I I'm trying to tackle that
25:05
problem first, then we'll tackle the how do we sort through it
25:08
all and and make sense of a problem?
25:11
Absolutely and you've got like your double-blind placebo based
25:13
studies and all these kind of things that are very formal.
25:15
But what about just this flood of data you're talking about?
25:18
You mentioned like longevity specialist, which is really kind
25:21
of like just a keyword for people doing really cool kind of
25:26
like therapies where you could maybe you go get an MRI and you
25:29
send all the data through some like LLM so we can scan
25:32
everything for you. And that's it's basing that
25:34
against different data. Is there is there a pool of data
25:37
that you guys are collecting that's not necessary clinical
25:40
trials, rather it's just like all of this girth of testing
25:44
information. Now obviously that's going to be
25:46
protected for people's privacy, but is there an opt in function
25:50
where you can collectively maybe take the pulse of 5 billion
25:54
people at some point? Yeah.
25:56
I mean we're at hundreds of millions now.
25:59
We have and we have that depth. It is de identified and we don't
26:03
identify individual patients, but you know, it's powerful.
26:07
And I I sort of having spent my life doing this, I have sort of
26:11
a recurring nightmare that like the cure to cancers sitting in
26:13
some data I have access to. I just didn't look right.
26:17
So like I know we've only begun to tap the power of that and the
26:20
computational power of compute and LMS added to that, it's
26:25
scaling incredibly fast. We're going to discover things
26:28
at a pace never seen before. I'll give you a a sort of less
26:32
consumer example of where I think this is starting to apply.
26:36
You know, if you look at, you know, drugs that have been
26:39
approved for a while and we know that they're safe, relatively
26:43
speaking, you know, an antibiotic or something that's
26:46
been taken millions of times, and the data I have, I can see
26:49
it's not causing a significant number of adverse events.
26:52
Well, if I can apply an algorithm to that data and I can
26:54
figure out that that drug might be really useful for another
26:59
disease, why are we waiting 15 years to approve it, right?
27:04
We, well, it's safe, it's out there.
27:06
And if we can create evidence and give that context to
27:09
clinicians who are already prescribing it, you know, maybe
27:12
that's something to think about. And I tend to think we actually
27:14
have maybe a lot of the drugs we need.
27:17
We just haven't identified their use case.
27:18
So I'll just tell one story about this.
27:20
At Stanford, there's a researcher named Dylan Dodd.
27:23
And Dylan researchers at a basic science level, the gut
27:26
microbacterium, he does all the sequencing, all the basic
27:29
science papers, And he was looking at this in animal models
27:33
of inflammatory conditions. And one of the models was gout.
27:37
And he basically looked at it. He said, well, if I can shift
27:39
the gut bacteria mix this way, the gout goes away.
27:43
And I shifted this way, it gets worse.
27:44
And he tried a couple of drugs that have long been approved.
27:47
Clindamycin was one of them. There's another antibiotic he
27:50
used, and he basically found that one antibiotic shifted the
27:53
gut bacteria, made the gout go away, and then the other one
27:56
made it worse. And he's about to publish his
27:58
findings. He ends up publishing in Nature.
28:01
He has access to us. So he called and says, hey, he
28:03
goes, you know, surely in your data, somebody with gout must
28:08
have taken these antibiotics, right?
28:11
And just by chance, maybe they were took it for something else.
28:14
And he's like, can we see if they got better?
28:17
And he's like, on the one that made it worse, see if they got
28:19
worse. And he's like, by the way, let's
28:21
look at everybody who's ever taken this antibiotic and see if
28:24
there's less gout. And we were in that study in a
28:27
day, brought it back. We're talking millions of
28:30
patients. And it was clear as day, like
28:34
these antibiotics were both causing it.
28:36
Some are causing it, and some are causing it to go away.
28:39
He even understood the basic science mechanism in the gut
28:41
bacteria. He added our data to his
28:44
publication and got, you know, a top tier tier journal for it.
28:48
And it's now in the literature. And think of that as opposed to
28:51
the 15 years it would take to develop a drug for gout.
28:55
And we could just find out in a couple of days, oh, wow, this
28:58
actually looks like this works like with high statistical
29:00
surety, like maybe somebody should be taking that forgot,
29:04
right, instead of needing all this time.
29:06
And in my hope in the regulatory space is that the FDA starts to
29:10
respond to this. You know, we have stuff again.
29:13
I think, you know, the FDA is a bit controversial right now, but
29:15
they do seem to want to improve and speed up drug development.
29:18
And you take what they're doing in orphan drugs where they have
29:21
a fast track with the right Rd. evidence on the right drugs that
29:25
are proven safe already in in humans.
29:28
Shouldn't we find a way to fast track these new therapies?
29:30
And we work closely with a group called every cure guy named
29:33
David Feigenbaum. Very famously, he's a pen Med
29:36
student, got a rare disease, use this type of data approach to
29:39
find a cure to his own rare disease.
29:41
He's a great guy. And he's advocating out now for
29:44
this type of thing, which is called drug repositioning.
29:47
You know, if there's therapies out there that we know are safe,
29:50
but they haven't been proven to work in your disease yet, but we
29:52
could prove it quickly with great data and and sort of
29:55
method, Why are we not doing that faster?
29:57
We might have it already. Why are we, that's a great
30:01
question. I mean, I'm I'm looking at this,
30:02
it's like data to evidence seems like the most clear no brainer
30:06
medical product of all time. When I look at kind of the the
30:09
history of the company, it looks like you guys raised 54,000,000
30:13
launched in 2020. Why wasn't this launched in
30:16
1980? Good question.
30:19
I would say the the the roots of this idea, OK, we call the
30:24
report we give to doctors, we call it a prognostogram.
30:28
It's APDF report. The reason we call it that is
30:31
because this all started at the Duke group back in the 70s.
30:34
So like they actually the original paper was called, they
30:37
spelled it differently than we do.
30:38
But about producing a prognostogram on data, I'd say
30:41
we spent 25 years between the 70s and the 2000s getting the
30:46
data. We didn't have data.
30:48
Well, guess what, we have data now.
30:49
OK. Then we had to build was the
30:52
automation and that's what our company represents because
30:55
typically speaking, let's say without us, if you had data, it
30:58
usually takes between two to five months to produce one
31:02
study. And we work with life science
31:04
pharmaceutical companies. That's been the standard for
31:06
them. It takes them big teams of
31:07
people with our tool, they can do it in a day.
31:10
So you know that automation as business value to them helps
31:14
accelerate their process of developing cures just on their
31:16
internal process. But it was that conversion step.
31:19
First it was we didn't have the data, Then the data was bad, we
31:22
had to make it better. Now we have good data and a lot
31:25
of it. Now we got to convert it to
31:26
evidence with automation. That's step we're at.
31:29
Amazing. Yeah.
31:29
Because I think before I'm, I'm pretty sure this is right before
31:32
the 1970's, the kind of data they were doing was like
31:35
basically standing around with a Scotch saying because I said so
31:39
I think that's. New group.
31:40
The new group had punch card IBM machines.
31:43
Yeah. The start of all this, right?
31:46
Like, literally, you can go back and we have the paper somewhere.
31:48
That's how this all began. They, you know, the idea of a
31:51
learning health system started there.
31:53
And we're, you know, this what we're really talking about here
31:55
is personalized medicine. And it's funny because that term
31:58
gets thrown around all the time. It's like, this is actually it,
32:02
right? Like here's a study on 50,000
32:05
people just like you. All your genomics, all your
32:08
history have matched you. And here's the evidence that
32:12
this will work better or this will work better for you.
32:14
Yeah, listen. As a, as a patient, that's very
32:18
exciting to me because again, like I've always kind of
32:21
straddled multiple worlds. I'm a bit of a hippie, let's
32:23
say. So I like, I'm interested in
32:25
nutrition and you go to your doctor and you're like, hey,
32:28
should I be getting more vitamin C&D?
32:30
And like, you know what, what about echinacea?
32:32
And they're like, eat what tastes good.
32:33
Take these two pills. You're like, that's
32:35
unsatisfying. So you go, you go to your
32:37
naturopathic Dr. and you're like, well, I think I sprained
32:40
my ankle. What should I do?
32:41
They're like rub it with a crystal.
32:42
You're like, oh, that's not satisfying.
32:45
Like we. Haven't done any crystal studies
32:46
yet. It could work.
32:49
I mean, it's a little cooler, but but it, it just seems like
32:52
there's a lot of disparate people that have like these
32:55
different ideologies and finally it's starting to blend together.
32:57
I'm meeting a lot of physicians that know a lot about like
33:00
nutrition and sports medicine and a lot of openness.
33:04
Not going to be too one note about this, Ryan, but what are
33:06
they all grasping at? There's no study to tell.
33:10
Yeah, absolutely right. And we talked through why that
33:13
is the case. But what if it wasn't?
33:16
What if we could produce these things?
33:17
I mean, right now, in addition to being able to produce a study
33:20
on demand, we have over 100,000 novel studies that we produced
33:25
that exists in the library for us to reference.
33:28
And I would tell you that like back of the envelope math, if we
33:32
wanted to be larger than pub Med, all the publications in the
33:36
world, we could be there within a couple years.
33:38
So imagine the entire basis of medical literature being
33:42
replaced in three years or or exceeded.
33:46
So like that's the if we solve that problem and find the
33:49
content that your natural math needs to say, hey, look at the P
33:52
value of my crystal, you know, like who they need, right?
33:56
So like if we can produce that, that's what I think the first
33:59
solve is. That that that kind of overhaul
34:02
of the entire system sounds intimidating to me if I'm a
34:04
doctor. Like how are doctors and boomers
34:07
kind of aligning with with the with the product?
34:10
What? What is their reaction?
34:12
You know, it's a great question. So this comes back to where
34:15
chatter, WD and LLMS play a really important role.
34:18
It's not so much that an LLM by itself can make a new study.
34:21
You know, you need technology like ours to do that.
34:23
But what it's great at is summarizing a lot of
34:27
information. OK.
34:29
And one of the things we do, we published a paper on this
34:32
recently took 3000 questions and we ran it through ChatGPT,
34:36
Claude, Gemini, perplexity and our service.
34:39
And we measured it on a evaluation system based on
34:42
position review called answered with evidence.
34:45
Like when you summarize the answer back off all the
34:48
publications in the world and all the stuff in our library,
34:51
all the news sites. Did you directly answer the
34:54
question or not with evidence that would be backed up in a
34:57
study? And so every answer we put out,
34:59
we actually put a badge, you know, if it's green, that's the
35:02
answer, there's a study to cite. If it's yellow or red, we're
35:05
kind of saying, hey, maybe there's not a study yet, in
35:07
which case you can run a new one with us.
35:09
And that summarization role is the key for doctors.
35:12
And you're seeing this, they're using tools today, like open
35:15
evidence is out right now. You know, it's had big uptake
35:18
with users, the great user experience, like they're doing a
35:21
great job summarizing, right? And I think summarization is the
35:24
key to making it work for doctors and then eventually
35:26
patients. I'm trying to solve the problem
35:29
up. They can need a lot of stuff to
35:30
summarize if you're going to be able to answer every question.
35:33
If an LOL of Chatu BT, you ask it a question and there is no
35:36
study in the literature, there's never been a trial.
35:39
What's Chatu BT going to do? It can't make up a trial that
35:42
never happened, right? So it's a combination of the
35:44
need for new content, high quality, and great
35:47
summarization. Tools.
35:49
Amazing. If we're looking into the future
35:51
here, a little bit back to our crystal ball and we're thinking
35:55
about chat RWD. Do you see this as like a robot
36:00
that's walking around available for questions or making
36:03
recommendations? Do you see this as like a chat
36:07
interface? People are asking questions?
36:09
Do you see this as something it's noticing trends and then
36:13
kind of like creating studies specifically based on like data
36:17
that's coming in? What kind of future look like?
36:20
Yeah, So it's something we're really excited about.
36:22
And I want to talk about a gentic AI and what that I think
36:26
means and think of it this way and step out of healthcare for a
36:31
minute. You got you remember when the
36:32
iPhone came out? Yeah.
36:34
What happened? All of a sudden we went from
36:36
looking at our laptops all day long to looking at our phones,
36:40
right. We call that a shift in Glass.
36:42
Your eyeballs are looking at a different thing or different
36:46
interface, right? You could say the same thing
36:48
happened with the App Store to some extent.
36:50
We went from browsers to apps, right?
36:53
We think a Gentek AI represents another shift in Glass, perhaps
36:57
ever. I don't know about you in your
36:58
personal life. I, I use agents like ChatGPT to
37:01
book travel and stuff. Maybe 20% of the time, like I
37:04
haven't switched all the way over.
37:06
I'm still going to my apps, so I'm still going to things.
37:08
But what if all of a sudden I could just talk to an agent and
37:11
it could go hit all my apps? It could do all the stuff in the
37:14
background for me. Why would I look at a browser
37:17
ever again? Why would I look at an app?
37:19
I can just talk to my agent. My agent will go figure out how
37:22
to get stuff done. And the healthcare context for
37:24
this is really exciting because some think of the average Doctor
37:28
Who even three years ago, spent most of their time logged into
37:32
Epic typing information into the HR or maybe looking up papers on
37:37
a reference tool. That's what they did most of the
37:39
time. Yeah.
37:41
What if now, first of all, we have Ambien AI, which in a
37:44
physician setting means it'll record the visit while I'm
37:47
talking to the patient, write the note for me, and then I can
37:50
go ask my agent, hey, what should I prescribe for this
37:53
person? Is there any evidence for it?
37:55
And it can call an agent like ours in give you the answer and
37:58
you don't even have to log into the HR.
37:59
I mean, I'm married to a clinician.
38:01
The idea for not having to log into Epic ever again would make
38:04
my household a happier place. I promise.
38:07
I promise you I'm a fan of this. Has it happened all the way yet?
38:10
No, but what I'm seeing on the ground, we're involved at
38:14
Stanford, Sanford's, the deployment site for Microsoft's
38:17
agentic orchestration platform. And specifically what they've
38:20
done is they've created an agentic tumor board.
38:23
OK. And what it looks like is it's
38:25
actually a Teams chat. I can't believe it's Teams, but
38:29
it looks and they're doctors talking to each other in the
38:32
chat and saying, hey, you know, what do you think of this case?
38:35
What should we do? They can invite agents into the
38:38
chat, right, And then ask different agents to do different
38:41
things. There might be an imaging agents
38:43
that says, hey, can you go figure out the size of the tumor
38:45
on XYZ date? They'll go read the image.
38:48
They can call us in and say, hey, OK, for a patient like
38:50
this, you know, Opdivo or Keytruda, which drug would be
38:53
better? And we'll pull forward that
38:54
study without them having to do anything else.
38:57
So if this starts to shift, OK, all of a sudden you're
39:01
representing a whole different way of working for doctors, by
39:03
the way, much less burden. You know, they're not doing
39:05
documentation all the time. And then your agent and the way
39:08
it interacts in that, that interface could really be
39:11
orchestrating behind the scenes, sort of where it grabs key
39:14
information from and feeding you back in a new way.
39:16
Now, I'm not convinced we're 100% of the way there, but if
39:19
that starts to happen, imagine that shift.
39:22
You're going from reading your e-mail from logging to Epic,
39:25
from using an app to just talking to an agent to get your
39:28
work done and, you know, not having to do all that data
39:31
entry, not having to have 10 logins to everything.
39:35
So that's something we're watching really closely.
39:37
I expect us to play at the forefront of that with Chatter
39:39
WD. We think we have the premier
39:42
evidence agent in that setting. So when it comes time to decide
39:45
what treatment to give, we think we've got the best in breed
39:48
right now, but it's early days. Yeah, yeah, it's a data
39:52
analysis, but it then it would be data discovery and
39:56
aggregation, you know, So I love this idea of change of glass
40:00
too. I mean, imagine if you just took
40:02
the stethoscope that every doctor has around their neck at
40:06
every waking moment and put a camera and a recorder in there.
40:10
And so now it sees everything and everyone has it.
40:13
So not you're not, you don't have to do the data entry.
40:16
Rather, all the data's coming in in real time constantly, and
40:19
it's all getting centralized into a massive hub.
40:22
I mean. Maybe he knows before you even
40:24
need to ask. I mean, I, I would tell you at
40:26
Stanford without deployment in some settings now, you know,
40:30
while the doctor records the visits, talking to the patient
40:34
as they swivel their chair over to look at the note, we're
40:38
actually running in the background like they didn't type
40:40
us a question. We're absorbing from the visit
40:44
what the key decisions are, producing the evidence and
40:46
putting it right back in front of them.
40:47
So I imagine it going to a place where it is almost passive in
40:51
that way and you're getting that the best information you need,
40:54
sort of the insight you need before you even know it.
40:56
Yeah, the, the, the promise of agentic AI is really interesting
41:00
because then there's groups of agents that they're calling
41:03
swarms. So you could have swarms
41:05
gathering all this data from different angles for different
41:08
tasks, talking to other agents to help collaborate on different
41:12
projects. And we are heading to a place of
41:14
like highly personalized, highly higher percent, higher than 14%
41:20
of data informed evidence. It's a wonky area, but you just
41:23
dropped something I'm really interested in, which is one of
41:26
the things we're finding in this paper we just put out about two
41:29
weeks ago. People can look up Astro's
41:31
answer with evidence. One of the things we saw and
41:33
initially just the rates of who answered the best was how often
41:37
do all these LLMS agree, right? Like is there consensus formed
41:42
between these things? And to your swarm idea that
41:45
what's being called like in the technical circles is it's a jury
41:48
approach. So what if we should?
41:50
We should ask everybody, Ask ChatGPT, ask Gemini, ask us, ask
41:53
everybody. See how the consensus is formed
41:56
and the role in agentic. And this is an idea for your
41:58
founders out there. If you know a specific vertical
42:02
or workflow application, you know how it needs to work.
42:05
Being the one who can coordinate the swarm in a way that knows
42:09
when I should hit Gemini for this, when I should hit this
42:12
application, and how to parse back consensus forming amongst
42:16
those answers. That is the frontline of a
42:18
gentic AI right now. And it is to me the argument
42:22
that big tech doesn't swallow the world.
42:24
Because in order to do that well, you have to have domain
42:26
and vertical knowledge for your use case and workflow.
42:29
And I think that's a great opportunity for start-ups right
42:31
now. 100% And there there's got to be like a place for gut
42:36
instinct, right? I mean.
42:38
That's kind of what it represents.
42:40
It's like for a knowledge, doing that review of the swarm and you
42:44
were an expert, how would you review it?
42:46
And if you can represent that step, I think that that here's
42:49
why we do all this metric stuff, because I think that comes down
42:53
to trust, right? At the end of the day with the
42:55
gentigai or any of it, it's going to come down, do the
42:58
doctors or eventually maybe patients trust it or not and
43:02
trust equals adoption. So the ability to metricize and
43:05
measure that you're doing that step right is actually going to
43:07
be a differentiator. I think even a a way to market
43:10
you're offering. Yeah, 100%.
43:12
Just trying to think of a case study, you know, God forbid I
43:16
have to see a urologist for some invasive surgery type situation
43:20
and GPT telling me one thing and this guy, my doctors tell me
43:23
another thing. And I'd be like, wait, how many,
43:26
how how many years of experience did you say you've done this?
43:30
I'm like, buddy, I'm going with you.
43:31
I mean, GPT is fantastic, but at some point there's.
43:35
Like, what did you add an app that could take the input from
43:38
your doctor, the e-mail he sent you, hit five other LMS, run a
43:41
new study, compare and contrast differences and bring them back
43:45
in a way that you knew what the key decisions were and what the
43:48
evidence for each of the decisions was.
43:50
I mean, I think that we're not talking about agents of agents,
43:52
I guess, but it will be there faster than we think because I
43:55
think that comes back to user experience and trust.
43:58
Like you need to trust it or you're not going to use it,
44:01
right? So the the fight for trust,
44:04
particularly the medical realm, is going to be where the battle
44:06
is, we think. Cool and and speaking to the
44:09
listeners. If anybody wants to make the the
44:12
wearable stethoscope I'm in 100%.
44:15
I love the idea of taking something as far away from AI as
44:18
you can and bringing it as close as you possibly can to it.
44:21
Give Meta a call, Tell him to put the rifle on the top of the
44:24
stethoscope. The amazing Cool.
44:27
This has been awesome. Brigham, I really appreciate you
44:29
coming on the show really briefly, and I don't want to
44:33
pronounce it wrong. Atropos.
44:35
Yeah, you got it. Atropos.
44:37
Who's the Who's the customer for Atropos?
44:40
Yeah, today we work with health systems, academic medical
44:43
centers, community practices on the provider side.
44:47
I worked with some pay viters were kind of insurance and
44:50
health system like a Kaiser for example.
44:54
We also work with life science companies.
44:55
You know, this automation we build is great to accelerate
44:58
R&D, you know, and, and I think you know, are bridging that way
45:02
across the health system to where the need for evidence,
45:05
this evidence gap is so present and helping them fill it, you
45:09
know, for the long term, you know, we'll see.
45:11
I mean, the the patient facing version maybe someday.
45:14
But I think for now, solving that Evans gap, getting that
45:17
information out so people can consume and use it both at the
45:21
point of care and in research, is where we're focused.
45:23
Yeah, incredible, very intuitive product.
45:26
What's the next step? Are you guys going to raise more
45:29
capital? Are you looking for strategic
45:32
acquisitions? Are you looking to be acquired
45:34
and sell the company? What what makes most sense for
45:36
you guys right now? Yeah.
45:37
I mean we we're having great growth in the business and
45:40
sitting on cash, plenty of cash store from RB.
45:44
I think the question at the forefront for us is, you know
45:47
are these shifts happening, you know shifts in glass for
45:50
instance and a gentic and other things that may be worth
45:52
doubling down. So we'll we'll think about those
45:55
opportunities that come through and try and hit that window to
45:58
scale this everywhere. Tremendous people want to
46:02
connect with you online. What's the best place to do it?
46:04
LinkedIn I respond so please shoot me note talk people OK I.
46:08
Really appreciate you coming on the show.
46:10
I want to give a shout out to Alex from Solcoms as well for
46:13
making this happen. But super fun conversation.
46:15
Thank you. Yeah.
46:16
Thanks, Ryan. Have a great time.
46:19
Hey founders, let me ask you this.
46:21
Are you building with AI or getting left behind by it?
46:26
If you're serious about scaling smarter, faster and bolder, you
46:29
need to check out AI for founders.co.
46:33
It's your go to resource for cutting through the noise and
46:36
actually using AI to grow your startup.
46:39
Each week, AI for Founders drops a sharp, no fluff newsletter,
46:44
dives deep with expert guests on the podcast, and hosts monthly
46:48
workshops where you can get your hands dirty with the real tools
46:52
and strategies. Whether you're pre seed or
46:55
series B and beyond, launching something new or just staying
46:59
sharp, AI for Founders gives you the edge you need to win in 2025
47:03
and beyond. Hit AI for founders.co to
47:07
subscribe to the newsletter, tune into the podcast, and grab
47:11
your spot in the next workshop, and it's all free.
47:14
Build smarter, move faster, think bolder with AI for
47:18
Founders. Now back to the show.
47:23
This episode is brought to you by Moeburst, the AI first growth
47:26
engine trusted by iconic startups like Calm, Robin Hood,
47:30
Reddit, Uber and Google. Moeburst is a digital, full
47:34
service, mobile first marketing agency that helps companies to
47:37
scale and become category leaders.
47:40
Whether you're finding product market fit or ready to
47:43
accelerate growth, MO Burst delivers creative, data-driven
47:47
solutions that help with scale and efficiency and impact.
47:51
At MO Burst, we're not just guessing what works, we're
47:53
predicting it. As an AI first agency, MO Burst
47:57
uses cutting edge tools to analyze user behavior, adapt
48:01
creative in real time, and uncover hidden growth
48:04
opportunities. The result?
48:05
Marketing that's faster, smarter, and built to convert
48:09
from early stage start-ups to global brands.
48:11
We've helped hundreds of companies overcome complex
48:14
challenges, unlock new audiences, and optimize every
48:18
metric that matters. At MO Burst, we believe that
48:21
creative thinking, advanced technology, and data are the
48:24
core drivers of success at scale.
48:27
If you're serious about scaling your business with precision and
48:29
purpose, it's time to talk to MO Burst.
48:32
Visit moburst.com and Let's build something category
48:35
defining together.