- 2 years ago
Dr. Isaac Kohane, Professor and Chair, Department of Biomedical Informatics, Harvard Medical School; Editor-in-Chief, NEJM AI Dr. Peter Lee, President, Microsoft Research, Microsoft Julie Yoo, General Partner, Andreessen Horowitz Moderator: Chrissy Farr, Owner and Editorial Lead, Second Opinion Media; Co-chair, Fortune Brainstorm Health
Category
🤖
TechTranscript
00:00We have so much to talk about and not a lot of time to do it, and I feel like we have just gone from
00:07Spermageddon today to climate change and now we're talking about AI which I think in some ways is incredibly
00:15Amazing just thinking about the applications and in others deeply terrifying at the same time
00:19So I wanted to focus what time we have today on not talking about the thing where all the money is going which is mostly
00:27Revenue cycle management and payment and anything touching the back office. I want to talk instead about diagnosis
00:34So I'm going to start with with Isaac. I just went to your conference
00:38We talked a lot about the potential for diagnoses using generative AI
00:43Tell us a little bit more about why you've come around on this because you used to be skeptical
00:47And I think now you're increasingly saying I think doctors everywhere should be using this tool in their own practice
00:53So I'm not even sure that doctors everywhere should be using it I do think they should but more importantly
01:00patients are using it now and
01:03Here's where
01:04well, I got my PhD in computer science back in the 1980s, which was also a heyday of AI and
01:10It we were disappointed afterwards
01:12Why am I not worried about this happening now because we released these tools right to the public just as we
01:19released them to doctors
01:21Doctors not have privileged access. This has been game-changing in the context where we cannot even get primary care
01:29for my own faculty in Boston
01:33Patients on diagnostic odysseys
01:36We just have at the meeting that we were together
01:38We heard about a mom who had a child has a child who for three years had bad gait
01:47trouble walking trouble chewing
01:50Paralyzing
01:51headaches all the studies for three years
01:54Nobody could diagnose she then types everything into GPT-4 is told
02:01Most likely this is a tethered cord syndrome. Your spinal cord is trapped by the bones around it
02:08She goes to a neurosurgeon and there's a neurosurgeon looks at the MRI says. Yep. That's what it is goes to surgery
02:14Game-changing, this is happening again and again. And so with all the warts that we see in
02:22AI
02:23Today, it's changing and moving so fast, but even as it is today
02:28It's changing the lives of patients. I think the real question for medicine is is
02:34medicine going to keep up with the rate of change of
02:38Knowledge that AI is going to bring to patients and do so
02:42Symmetrically in medicine and medical care so that brings me to a question I had for you Peter
02:48Which is that it feels like whenever I'm on stage at these conferences
02:52It's a big talking point for the technology companies to say we are not replacing humans. We merely augment them
02:58This is a term I have had many many times in my career
03:02But what you're talking about is a case of actual replacement
03:06This patient could not get a diagnosis had to turn the patient's mother had to turn to GPT for it was essentially a
03:14Physician of sorts in the loop that had to be brought in to make this ultimate assessment that proved to be correct
03:21So when you hear about case studies like this, but you must do all the time. Where do you land in terms of?
03:27Augment versus replace. Well, you know first off just listening to Zack. It is an amazing
03:34Journey that we've been on because Zack was the first outside person that I looped in
03:43Brought into the tent so to speak on what we now know is GPT for and I think that first phone conversation
03:49You were pretty skeptical. I think you were tell us that what was your exact?
03:55I don't know. I don't know what the words I used but I was just thinking what is a cell job? I'm getting
04:01He was very bad at suppressing his eye rolling, yes
04:05But you know, but then I think the first things you tried were some diagnostic
04:10Clinical problem-solving puzzles and it was shocking how well it does allow me to quickly jump in
04:16so there's something called the undiagnosed disease network where you do genetic diagnosis on people who have not been diagnosed for a long time and
04:21I took one of our hardest cases that we had not yet diagnosed
04:25Where there were five mutations that could be it and I just gave it to GPT for as it was much more interesting back then
04:32And it got it
04:33And how do I know it got the right answer because we didn't did the bench research afterwards and it was the mutation it picked
04:39That was the causative one. Yeah, and of course now there have been over a hundred generative AI
04:45Research articles published in major medical journals, including the new one that you started the New England Journal of Medicine
04:52And so it's very well established that there's capability there. And so now the question is what about this augment?
04:59versus
05:01replace
05:02thing and
05:04First off just as a practical matter
05:07Until we get clear accountabilities or what the bioethicists would call recourse straightened out in our society
05:13Under no circumstances in my opinion should the machine be replacing a human doctor
05:18It's only with human doctors that we as a society have
05:22Clear accountabilities in place having said that I do believe very soon sooner than we expect
05:30Patients might demand that a human doctor
05:34Double check his or her work with the use of AI in the same way that if you are a pregnant woman and you go
05:40to an obstetrician and
05:42That doctor says, ah, no, I'm gonna you
05:45Do manual palpation because I don't believe in ultrasound
05:48I think you'd be pretty uncomfortable today, even though it was just the opposite a little while ago
05:54so I think there's going to be some
05:57flipping of
05:59Patient expectations in addition to patients self-serving themselves with AI. So I want to disagree, but I want to hear instead from your question with
06:07Julie, yeah, Julie
06:09I mean
06:09so many questions for you because you're actually one of the few folks who's really investing right now in AI and
06:16One of the conversations you and I have had is really about the sheer amount of money that needs to go into some of these
06:22companies just because of the cost of compute and
06:25Adresin's been posting about this just taking some incredible content going out of just like how much does it cost to spin up some of
06:33These models that need to be running, you know
06:35A lot of that money going towards companies like NVIDIA and and others
06:40So how much is VC now just a game of like?
06:44Tons of money going to essentially, you know propping up at video
06:48I think Nvidia has been on a great streak
06:50And how do you how do you think about that as you make these sorts of investments where you're looking to make a return?
06:55Yeah, I mean just to tag on to this past conversation about diagnosis. I mean when you look at
07:00Number one like our industry is one of one of the only industries if not only the only industry that actually has
07:06regulatory rails for how to approve
07:09Diagnostic AI products, right?
07:10And so I think it would be a shame if we weren't to take advantage of that and actually lean into that opportunity set as
07:15the technology evolves
07:17Such that it's appropriate for those use cases
07:19The second thing is that all the rage is I'm sure everyone most folks in this room at this point have seen
07:24Some kind of demo about the ambient scribe technologies. I think you had done like a live one
07:29Okay, great
07:30But when you look at what that tool what those tools are actually doing
07:34It's actually creating a proposal for diagnoses, right? And so that kind of technology is already out there
07:40It's already in the hands and it's going viral across, you know physicians
07:44I have never been told by any clinician ever
07:46That they were excited about anything let alone a piece of technology and how many calls have I received from you know clinicians going out
07:52Of their way to tell me how magical some of these tools have been to just transform their careers and their life
07:57And actually make it, you know, pleasurable to be a doctor again. I mean all these things are quite incredible
08:03so I think it's we're already there and
08:05You know again we have we actually have the regulatory rails within our industry to actually take advantage of this
08:10From an investment standpoint, you know, our belief is that number one. We do require specialist models within our domain, right?
08:17So yes chat GBT and all of the open models are remarkable in many ways
08:22But at the end of the day our industry especially when it comes to clinical decision support and and those kinds of applications
08:28require a mix of
08:30Probabilistic and deterministic intelligence right and you're not going to get the latter with just off-the-shelf
08:37generative AI tools today
08:38And so there does need to be investment in novel architectures to take advantage of both
08:44You know more traditional and more next-gen
08:47AI capabilities I think in one on under one roof to be able to do this effectively and then to as we also all know
08:55The data that is necessary to inform these models for the most part lives below sea level
09:01Right are in proprietary assets within our health healthcare system are not yet optimized
09:06Necessarily for training use cases within the AI landscape
09:10And so we have a tremendous amount of work to do it to like really unlock
09:13All of the investment that we've now put into EHRs into other data systems
09:18All this voice data that's out there and all of these
09:21You know call center recordings are just a tremendous treasure trove that we have yet to scratch to scratch the surface of and so
09:27That's really the work to be done and therefore you need to invest
09:30And so that's really the way we think about a lot of this is that you do need to build proprietary models
09:35within our space that require
09:37Investment that said we believe that there will be a very finite number of companies that do so, right?
09:44So there will be a small just like in the rest of the industry
09:46There's a small number of companies that are actually building these foundational models at the scale that is being used by consumers
09:53We think the same thing is going to play out in our space and there will be a long tail of companies that then take
09:58Advantage of those foundational models to you know
10:00Do additional tuning and rag and whatnot to build applications on top of but that's sort of how we see the industry playing out
10:07At this at this vantage point just a reminder that if anybody has questions, feel free to to ask any point
10:14So I'm just going to dig a little bit deeper into this diagnosis question that I think we've all kind of
10:19Soft-circled as being one of the major areas that we will see AI disrupting health care
10:24What are you what do you think of the main?
10:27Concerns related to you know doctors clinicians getting a little lazy with these tools just kind of you know, we say oh, well, it's fine
10:35Well, you know, we'll audit we'll double check. Yes hallucinations happen, but we'll make sure that there is another set of eyes
10:41But we we know from looking at areas like
10:45Airlines that you know pilots can get lazy and get used to the autopilot mode
10:50So how how do we?
10:52Avoid that or prevent that from happening in the way that we design some of these tools and I'll throw that to you first
10:57So I have direct experience with this. I'm a bad driver and I have a Tesla. It's true. Very bad. I
11:06Am a bad driver and in fact my girlfriend who's somewhere in the audience
11:11When she gets in the car with me wants me to switch on autopilot, but here's the thing
11:16Even with the autopilot. I'm a bad driver. And so I have my hand the wheel. I know enough to like jiggle the wheel and
11:24then
11:26I
11:27Do something really bad. I pick up my iPhone and I start looking at and then the Tesla
11:34Paraphrase says to me Zach. Don't do that
11:38Keep on looking at this. Okay
11:41I'm switching off autopilot
11:43so fine
11:45All right, so I drive home and it's not not a great experience. I do it again
11:51We're driving. I look at it and says
11:56Stop it and I get home and it says Zach do that three more times. I'm switching off autopilot
12:04for
12:05Until the next update Tesla sounds like your mom
12:09That's how I talk to my
12:11so
12:12You know what? We're like kids many a times when when people are not looking at us and here's the point
12:16I think that if we're serious about this
12:19You have to find out is the doctor awake at the wheel and say something like oh
12:24Just drop into the summary some of the doctor didn't say and when they're checking off and say oh
12:31He'd really didn't say that there was an ugly hairy mole on their nose. There was no ugly. I think well
12:36It's going to be required checking in so either we're gonna
12:41buy into the importance of a set of human eyes in
12:46which case we have to check that those human eyes are on the patient and on the
12:51decision-making or
12:53we
12:54We let go of that illusion and drop the doctor out of the loop
12:57But if we think that the doctor has to be in loop
13:00We have to make sure it stays in I could tell you long
13:03From what we know from informatics the last 20 years doctors will argue endlessly about certain things
13:08But then when you give them an order entry system
13:10They'll just use the standard defaults without any argument because they're under time pressure and that's it
13:17We have time for a question from the audience
13:22Hi
13:24Thank Barthen with nano where we do a
13:27remote diagnostics at home
13:28We recently got an FDA approval for an AI enabled hypertensive diagnostic and self-administered continuous blood pressure at home
13:35I bring that up not only for the shameless plug for everybody, but in general that is not a generative AI tool
13:42It's fixed and part of the reason we got the approval is because we offered to give post
13:48Postmarket surveillance data every six months or at a certain
13:51Period of patients that we do with FDA and this brings me to my question on gen AI
13:57We often I feel like don't talk about the inputs being static versus dynamic
14:03Blood pressure, for example bounces all over the place. That's why we have to give PSM data every six months and then
14:09With regards to health care while we think of gen AI with language models
14:14we also don't talk about closed source versus open source and
14:18We were able to create these LLM s because the internet is completely open source and all the dialogue of all the postings that we've done
14:26We don't share data across organizations and across systems in health care in that regard
14:31I'm just really curious on everyone's panel static versus dynamic inputs
14:34And then how do we get to actual gen AI for diagnostics if we're really just gonna be a closed source industry
14:44Yeah on the second question
14:47Part of the question which has about to do with data
14:50I think the unfortunate thing that I predict we'll see for a few years and I don't know how long it'll be
14:55I don't know if it'll be one year three years ten years, but some period of time
15:00Health data will be more severely protected not less
15:05organizations will be more selfish and
15:08Proprietary with their data and the reason for that is the generative AI helps quote-unquote unlock the value of data
15:16Forget about the chat interface chat interface is going to be important
15:19But the real value generative AI is in its ability to rummage through large amounts of unstructured data and derive structure
15:27that creates value and
15:29That is going to make all of those health care organizations that have invested millions of dollars
15:34Aggregating their data more protective because they want to realize that value and it's going to take some sort of intervention or some sort of tipping
15:42point
15:43sort of like what we saw in the early days of
15:45Internetworking before people realize that if we all work together, we'll all be better off on the
15:53Maybe I should give someone else a chance
15:56Yeah, no, I think I mean the dynamic input piece, too
15:58I mean this gets back to my point earlier about there's entire
16:02Datasets that need to be created to even be able to train these models to be representative true by a lot biology true
16:09you know health care status and so
16:11You know, I think we are today constrained like I think if you're building a model that's entirely
16:16Reliant on just EHR data, which you know to some feels like a holy grail
16:19I think that's actually completely missing the point, you know that the information that's stored in the EHR data is highly biased highly sporadic highly
16:26You know centralized etc. And is a very poor represented represented
16:31Representation of an actual, you know longitudinal care journey
16:34And so I think that that's what I meant earlier about, you know
16:37So much work that needs to be done to really lay down additional data rails that don't exist today in 20 seconds
16:42I want to highlight an alternative model which is out there
16:45Right now Apple Health gives patients access to their own data
16:50For over 800 hospitals, that's 40 million Americans at least if you exercise that you will have in one space
16:58all your health data labs diagnoses medications
17:03procedures and all the
17:05lifestyle data from Fitbit, etc
17:07in one place I
17:11Really am waiting for the companies that are going to take that
17:16Population based view taking the patients who have greater far greater
17:22Incentive to share than the institutions. I'll stop there. Sounds like a stunning idea. It's not a biting. I love it. Thanks everybody
Comments