Pathology Grand Rounds, November 7, 2024

Transcript

• 00:00Hi, everybody. Welcome to Grand
• 00:02Pathology Grand Rounds.
• 00:04I'm Rob Homer.
• 00:06I am exceptionally pleased to
• 00:08invite someone,
• 00:09to speak to.
• 00:11It's, Adam Rodeman's, has a
• 00:13dual degree in medicine public
• 00:14health from Tulane,
• 00:16trained in internal medicine at
• 00:17Oregon Health and Science University,
• 00:20the global health fellow at
• 00:21the Beth Israel Deaconess, where
• 00:23he actually wound up working
• 00:24in Botswana for a couple
• 00:25of years and wound up
• 00:26developing a curriculum for, interns
• 00:29in Botswana, which is great.
• 00:31He's currently assistant professor of
• 00:33medicine at Harvard Medical School
• 00:34and attending internist at Beth
• 00:36Israel Deaconess.
• 00:38So I became familiar with
• 00:39Adam from his,
• 00:41well known podcast
• 00:43on bedside rounds,
• 00:44which is phenomenal because it
• 00:46goes into the history of
• 00:47medicine. And I have to
• 00:48say that if you wanna
• 00:50understand, like, every time you
• 00:51look around and says, why
• 00:52are things like this? It
• 00:53turns out there frequently is
• 00:54a reason, and it's frequently
• 00:55found in history. And it
• 00:57was like, oh, that's why
• 00:58we do this. And so
• 00:59it really provides a lot
• 01:00of insight, and I surely
• 01:01recommend it. Also, I have
• 01:03to say, at dinner last
• 01:05night, I found Adam just
• 01:06enthusiasm. If we could just
• 01:08harness that, we don't need
• 01:09nuclear. We don't need solar.
• 01:11We just need Adam. It's
• 01:12phenomenal.
• 01:14He is,
• 01:15having his history. So he's
• 01:17trans he transformed from history
• 01:19to
• 01:20AI. How did that happen?
• 01:21Because there's a history of
• 01:22machine learning in medicine, which
• 01:23goes back a long way.
• 01:24When I was going to
• 01:25medical decided to go to
• 01:26medical school when I was
• 01:27in college, one of my
• 01:28classmates said, why are you
• 01:29doing this? Computers are gonna
• 01:30take this over any day
• 01:31now. And that was in
• 01:32nineteen seventy nine. So a
• 01:34little maybe it's true, but
• 01:36it just took a little
• 01:37while.
• 01:39His research founded, funded by
• 01:41the Macy Foundation,
• 01:42Gordon Betty Moore Foundation, National
• 01:44Academy of Medicine, explores integration
• 01:46of large language models and
• 01:47complex diagnostic challenges, which you're
• 01:49gonna hear today,
• 01:50associate editor for the New
• 01:52England Journal of Medicine AI,
• 01:53published work in JAMA Open
• 01:55Network, JAMA Internal Medicine, New
• 01:56England Journal of Medicine, among
• 01:57other journals.
• 01:59But
• 02:00and, actually, you know, the
• 02:01rationale for really being here,
• 02:03he's planning the innovation of
• 02:04digital in education and medicine,
• 02:06codirects at the EyeMed program,
• 02:08which is dedicated to exploring
• 02:09digital education.
• 02:10He is among other things,
• 02:12he has a digital education
• 02:13track for future educators to
• 02:15learn about,
• 02:16electronic education
• 02:18technology.
• 02:19Currently leader of a task
• 02:20force to integrate, into the
• 02:21Harvard Medical School undergraduate
• 02:23medical education.
• 02:25Received numerous teaching awards,
• 02:27and mentorship awards, including most
• 02:29recently the Herman Blumgard faculty
• 02:31teaching award.
• 02:32Join me in welcoming Adam.
• 02:34I'm really excited to hear
• 02:35this. Thank you. This is
• 02:36gonna be really cool. I'm
• 02:37a hi, everybody. I'm Adam.
• 02:41Okay. Can you guys hear
• 02:42me? Oh, good. I'm a
• 02:43pacer. Before I start
• 02:45oh, I'm so sorry. If
• 02:46you're an internal medicine physician,
• 02:48could you raise your hand?
• 02:51I've had a feeling I
• 02:52was gonna pick on you
• 02:52guys. I would like actually,
• 02:54so
• 02:55at
• 02:56about twenty minutes in, I'm
• 02:58going to demonstrate
• 02:59our actual clinical workflows.
• 03:01If one of the internists
• 03:02would be willing it's not
• 03:04a competition, but to actually
• 03:05use the AI to one
• 03:07of the real patients from
• 03:07our studies, would you be
• 03:08willing to?
• 03:11Yeah. Okay.
• 03:12So hi. My name is
• 03:13Adam. I, I'm not a
• 03:15pathologist,
• 03:16so I apologize in advance.
• 03:18I am I don't know.
• 03:19I don't know who's nerdier,
• 03:20but in general, internists are
• 03:21some of the nerdiest physicians
• 03:22alive.
• 03:23I got a nod over
• 03:24there. So
• 03:26I research how people think.
• 03:28And in pathology
• 03:29and when it comes to
• 03:30machine learning in particular, there's
• 03:32really been a focus
• 03:33unfairly, I think, on, like,
• 03:35classification models. Right? Just what
• 03:37is this image? Which I
• 03:38think, well, not I think.
• 03:40I know underplays the actual
• 03:41cognitive processes that go in
• 03:43pathology, which is why everyone
• 03:45still here has a job
• 03:46and is in no danger
• 03:47of losing a job anytime
• 03:48soon. Now what I'm going
• 03:50to talk about today is
• 03:51my research, but I'm gonna
• 03:53try to make it entertaining
• 03:54to talk about
• 03:56language models,
• 03:57reasoning, multimodal models, and what
• 04:00the data in here, I'll
• 04:01just have, you know, I
• 04:02was talking to my PhD
• 04:03student this morning. I have
• 04:04updated this slide with data
• 04:05that has not been published
• 04:06yet that was run yesterday.
• 04:07So this is literally
• 04:09the,
• 04:10latest breaking data here. But
• 04:13to give you an idea
• 04:13of where we've come already
• 04:15and where we're gonna
• 04:16go. So
• 04:17I wanna start with this
• 04:18idea that is relevant to
• 04:20basically every medical specialty, and
• 04:21it's what is diagnosis? What
• 04:23does it mean to make
• 04:25a diagnosis? This is what
• 04:27we all do. This is
• 04:27theoretically what internists do, though
• 04:29often we don't get a,
• 04:30we don't get a definitive
• 04:31diagnosis. This is what we
• 04:32do in pathology.
• 04:34Basically, all of medicine
• 04:36is not all of medicine,
• 04:38but it's focused on this
• 04:39idea of diagnosis. Even when
• 04:40we're talking downstream about
• 04:42management, it
• 04:44usually
• 04:45relies on the diagnostic process.
• 04:47But what does it mean
• 04:48to make a diagnosis? So
• 04:50there's a great paper. I
• 04:51always say it was just
• 04:52published. It was actually published
• 04:52about fifteen years ago. I'm
• 04:53just getting old and unstuck
• 04:55in time, but there's a
• 04:56great paper that looks at,
• 04:59like, what do we mean
• 04:59when we see diagnosis? What
• 05:00do we mean when we
• 05:01say clinical reasoning? And unfortunately,
• 05:04as I'm sure everybody here
• 05:06knows, there's not a standard
• 05:08definition and everybody is talking
• 05:09about something a little bit
• 05:11different, which means we have
• 05:12a tendency to talk past
• 05:13each other. So to give
• 05:15you a sense on where
• 05:16the science stands, it's complicated.
• 05:18Yeah. I'm gonna do some
• 05:19Simpson stuff in here. So
• 05:20I realized that as a
• 05:21geriatric millennial that that dates
• 05:23me quite considerably.
• 05:24But at its most basis,
• 05:26basic dating back to the
• 05:27ancients, diagnosis is a classification
• 05:30task. Like, what is nosology?
• 05:31Nosology is the way that
• 05:32we categorize
• 05:33diseases.
• 05:34And fundamentally, making a diagnosis
• 05:36means, okay. I have to
• 05:38come up with a single
• 05:39disease or multiple diseases in
• 05:41my differential out of this
• 05:42large classification
• 05:44schema. Well, that sounds easy.
• 05:45I, I think what's funny
• 05:47is if you go back
• 05:48to some of the early
• 05:49literature, like this is from
• 05:50the fifties on diagnosis, they
• 05:51literally thought it was this
• 05:52easy. Everybody here knows it's
• 05:54much more complicated than that.
• 05:57So modern researchers into the
• 05:58nature of diagnosis have focused
• 06:01more on human psychology.
• 06:03So, RIP Danny Kahneman,
• 06:05if anybody has read Thinking
• 06:07Fast and Slow lately, I'm
• 06:08I'm sorry. I'm gonna very
• 06:10quickly summarize it. I will
• 06:11also say I actually recently
• 06:12reread Thinking Fast and Slow
• 06:14just, like, two months ago.
• 06:15It is I don't recommend
• 06:16reading it. Danny Kahneman, brilliant
• 06:18person,
• 06:18very droll
• 06:20writer.
• 06:21Again, RIP Danny Kahneman. So
• 06:23our understanding of modern diagnosis
• 06:25is, you know, based on
• 06:26cognitive psychology.
• 06:28So this idea
• 06:30that
• 06:31most of the process is
• 06:32this very fast, very contextual
• 06:35automatic system one process. So
• 06:37system one, very classically, I
• 06:39I great example is you're
• 06:40driving on the highway going
• 06:41home. All of a sudden
• 06:43you stop paying attention. Fifteen
• 06:44minutes go by. You're still
• 06:45on the highway. Everything's going
• 06:47well. What happened? Your brain
• 06:49went into these very automatic
• 06:51thought processes and that's our
• 06:52classic system. One,
• 06:54we talk about heuristics or
• 06:56mental shortcuts often system one
• 06:58gets maligned, right? We spend
• 06:59a lot of time talking
• 07:00about how system one can
• 07:01lead us wrong. But the
• 07:02fact of the matter is
• 07:03system one exists. We evolved
• 07:05this way for a reason.
• 07:06It's fast, it's efficient, and
• 07:07it works pretty well.
• 07:09System two are the very
• 07:10slow, very contextual,
• 07:13very labor intensive, and theoretically
• 07:16less biased, not in reality,
• 07:17but theoretically less biased kind
• 07:19of formal thought processes.
• 07:21So what does this look
• 07:22like in process? Don't worry.
• 07:23I'm not gonna go over
• 07:24this. This is, from Pat
• 07:25Cross Gary who I've talked
• 07:26about a a bit over
• 07:27the last day. So this
• 07:28is, Pat Cross Gary's grand
• 07:30unified
• 07:31theory of clinical reasoning. Again,
• 07:33don't worry about it, but
• 07:34the gist is
• 07:35we cross over between system
• 07:36one and two. What does
• 07:37this look like in practice?
• 07:39So I think probably everybody
• 07:40has read or is aware
• 07:41of Judith Bowen's very famous
• 07:43paper from the early two
• 07:44thousands, which is based on
• 07:45the work of Bourdage about
• 07:46a decade before. But our
• 07:48current understanding,
• 07:50cognitive psychological understanding of the
• 07:52diagnostic process is something called
• 07:54script theory. And this is
• 07:55a knowledge encoding and knowledge
• 07:57activation theory.
• 07:59So
• 08:00script theory, how does it
• 08:01work? The idea is that
• 08:02we know things about our
• 08:03patients. We know things about
• 08:05path pathology from a variety
• 08:07of sources. Medical school, of
• 08:09course, reading journal articles, and
• 08:10more importantly, from seeing patients
• 08:12and from learning, from our
• 08:13practice, we get feedback. And
• 08:15all that knowledge gets encoded
• 08:17in our brains, but we
• 08:18are not a library. We
• 08:19are not the Dewey decimal
• 08:21system. I do not say,
• 08:23heart failure with risk reduction
• 08:24fraction
• 08:25is coded at e twenty
• 08:27three dot o five. That's
• 08:28not the Dewey decimal system.
• 08:29That's like another classification system
• 08:31that I forget. I spend
• 08:32too much time in academic
• 08:33libraries. But, no, how does
• 08:34that information get encoded? And
• 08:36it gets encoded
• 08:38in these things that in
• 08:39medical education and the psychology
• 08:41world we call scripts. Scripts
• 08:43are a psychological principle actually
• 08:44from the nineteen seventies. It
• 08:46comes from this idea of,
• 08:48there are stereotyped
• 08:49patterns of behavior. So for
• 08:50example, you walk into a
• 08:51restaurant,
• 08:52you have a script that
• 08:54you follow, and when you
• 08:55deviate from that script, it
• 08:56freaks people out. And if,
• 08:57for example, you were to
• 08:58go to your friend's house
• 08:59for dinner and you try
• 08:59to follow the restaurant script,
• 09:00it would be incredibly rude.
• 09:02So, you know, in the
• 09:03early nineties,
• 09:04psychologists who study clinical reasoning
• 09:06realized we're doing the same
• 09:07thing with diseases.
• 09:09So the idea here is
• 09:10that when we
• 09:12activate information about a disease,
• 09:14what we're really doing is
• 09:15telling a story to ourselves.
• 09:16And that story has, obviously,
• 09:18the presentation.
• 09:19That story has the pathological
• 09:21diagnosis, what that might look
• 09:23like. It has other, you
• 09:25know,
• 09:26the the treatments, all of
• 09:27that is organized together. And
• 09:29then more importantly, scripts do
• 09:30not exist
• 09:32in, isolation. It's not like
• 09:34a library. They exist in
• 09:35these parallel systems of networks
• 09:38called schema, where almost by
• 09:40definition, if you activate one
• 09:42script, you're excluding another. And
• 09:44in medical education, we spend
• 09:45a lot of time, you
• 09:46know, talking to our students
• 09:48about semantic qualifiers.
• 09:50You know, is this acute,
• 09:51subacute, chronic,
• 09:53polyarticular, monoarticular? And the reason,
• 09:55the psychological reason,
• 09:58this is right out of
• 09:58bordage, is that fundamentally
• 10:02those are the ways that
• 10:03we include or exclude different
• 10:05diagnoses. This goes back to
• 10:06a really interesting study by
• 10:07Arthur Elstein in the late
• 10:0870s, where at Harvard took
• 10:10a bunch of medical students,
• 10:11took a bunch of attendings,
• 10:12and then had them think
• 10:13out loud
• 10:15about what was going on.
• 10:16And as he expected, as
• 10:18the dominant theory was, everyone
• 10:20did the Sherlock Holmes thing.
• 10:22Everyone was like, okay, this
• 10:23is my theory. These are
• 10:24the next tests, the hypothetical
• 10:26deductive process. But what else
• 10:28he noted is, okay,
• 10:30well, the attendings only asked
• 10:32five questions and they got
• 10:34the diagnosis right. The medical
• 10:35students asked thirty questions and
• 10:37they were far less accurate.
• 10:39So there's something else going
• 10:41on than simply following a
• 10:42hypothetical deductive process. So what
• 10:45does this look like in
• 10:46practice?
• 10:47So I'm in clinic. A
• 10:48patient comes up to me.
• 10:49She says what did she
• 10:50say? I can't breathe. It
• 10:51started yesterday.
• 10:53Everyone has had this experience.
• 10:54I mean, it's a little
• 10:55bit different with pathology, but
• 10:56you look at something and
• 10:57then automatically your mind starts
• 10:59to sort that information. What
• 11:01is happening? You are activating
• 11:02one of these scripts. In
• 11:04medicine, we talk about a
• 11:05problem representation, the instantiation of
• 11:07a problem representation,
• 11:09which is in this case,
• 11:10acute dyspnea.
• 11:11This is a way of
• 11:12teaching, but it's activating that
• 11:13part of my mind. This
• 11:14idea that if somebody tells
• 11:15me all of a sudden
• 11:16I can't breathe versus six
• 11:18months slowly progressive shortness of
• 11:19breath fundamentally activates different ways
• 11:22that I think about information.
• 11:23And then
• 11:24and everyone has had this
• 11:25experience. Without thinking, you know
• 11:28what questions to ask next.
• 11:30Why? Because they're going into
• 11:32your script for acute dyspnea.
• 11:33You are going through that
• 11:34schema, and you're asking follow-up
• 11:36questions. So I hypothesis test
• 11:38COPD.
• 11:39Hypothesis test pneumonia,
• 11:41and based on that, further
• 11:42refine my problem representation
• 11:45until I get something that
• 11:47is reasonably close to the
• 11:48diagnosis or I find something
• 11:51that doesn't make sense and
• 11:52I cross over to other
• 11:54types of metacognitive processes. Now,
• 11:56I said that I'm going
• 11:57to go over some of
• 11:58the, like,
• 11:59different understandings of clinical reasoning
• 12:01because I will say that
• 12:02until,
• 12:03let's say, about twenty ten,
• 12:05we had basically agreed in
• 12:07the reasoning field that script
• 12:09theory was it.
• 12:10I think that
• 12:12it doesn't take much to
• 12:13think about the problems with
• 12:15script theory and that it's
• 12:16not a necessary or sufficient
• 12:18way to think about how
• 12:19we make decisions. So my
• 12:20classic example is right now
• 12:21I've had about a half
• 12:22gallon of coffee. So let's
• 12:23say I'm admitting patients, I'm
• 12:24well rested, well caffeinated, ten
• 12:26AM.
• 12:27My mind is gonna work
• 12:28differently than when I'm working
• 12:29at two AM. My pager
• 12:32goes off every five minutes,
• 12:33and my epic chat just
• 12:34keeps going red over and
• 12:35over and over again because
• 12:37our minds don't I mean,
• 12:39they don't work like that.
• 12:40There are a lot of
• 12:41ecological
• 12:42factors that affect our reasoning
• 12:44process. And this is ecological
• 12:46psychology is one of the
• 12:48frames now that we think,
• 12:49which is that reasoning is
• 12:50not something that just happens
• 12:51up here. Reasoning is something
• 12:53that happens in our bodies,
• 12:55happens in our environment, and
• 12:56these environmental factors play a
• 12:57large part. Again, I don't
• 12:58think any of this is
• 12:59controversial.
• 13:01I will put up this
• 13:02idea of situated cognition.
• 13:04This so to say that
• 13:05something is situated is to
• 13:06say that it's contingent. These
• 13:07are all psychological terms. But
• 13:09the idea here is that
• 13:10there's no such thing as
• 13:11neutral decisions. They're all shaped
• 13:13by our prejudices, by our
• 13:14life experiences.
• 13:16And because of that decision
• 13:17making, per situated cognition
• 13:20is individualized.
• 13:22I don't find this very
• 13:23helpful, like, from somebody who
• 13:24wants to, you know, make
• 13:25people better. I don't find
• 13:26this a very helpful psychological
• 13:27theory, but it is one
• 13:28that is out there that
• 13:29you should know about. And
• 13:30then one of the big
• 13:32things that I think is
• 13:32especially relevant in pathology is
• 13:34I think about how I
• 13:35make diagnosis in the real
• 13:36world. Also, you can tell
• 13:37this is an AI generated
• 13:38images because a lot of
• 13:39these, have, like, weird demon
• 13:41faces as the model gets
• 13:43confused and tries to average
• 13:44Simpsons characters and other things.
• 13:46But, you know, I might
• 13:47make a diagnosis, and I'll
• 13:48pat myself on the back
• 13:49that maybe I've diagnosed lupus
• 13:51nephritis and a new case
• 13:52of lupus. But you know
• 13:53what? What I'm doing is
• 13:54I'm looking at a creatinine
• 13:55trend that was identified by
• 13:57the patient's PCP, and nephrologist
• 13:58is an outpatient, and I
• 14:00might be looking at an
• 14:00ANA or a double stranded
• 14:02DNA that was sent much
• 14:03later, and all of this
• 14:04is mediated through the electronic
• 14:05health record. So the reality
• 14:07of diagnosis, it's not really
• 14:09collaborative, right? It's not in
• 14:11the sense that we're in
• 14:11the same room working together,
• 14:13but most
• 14:14diagnosis, I don't know most,
• 14:16I don't know the percentage,
• 14:17but a lot of diagnosis
• 14:18in the year twenty twenty
• 14:19four is mediated through the
• 14:20electronic health record happening with
• 14:22different people at different places
• 14:24in time.
• 14:25Okay. So why do we
• 14:27care so much about the
• 14:28cognitive processes of reasoning? Like,
• 14:31why why does anybody give
• 14:32me money money to study
• 14:33this, which I wonder myself
• 14:35also sometimes?
• 14:36And for the last, let's
• 14:37say, twenty five years, the
• 14:39focus has been on this.
• 14:40Ever since to is human
• 14:41has come out, there has
• 14:42been a huge focus on
• 14:44medical errors and the burden
• 14:45of medical errors. This is
• 14:46from David Newman Toker's, most
• 14:47recent paper. I actually think
• 14:49that this is what he
• 14:50pulled out. I think the
• 14:51more damning thing is that
• 14:53he says fifteen diseases are
• 14:54for half of serious harms.
• 14:55Four diseases accounted for thirty
• 14:57nine percent of harms in
• 14:58this study, and those diseases
• 14:59were such exotic things as
• 15:00stroke and heart attacks. So
• 15:03doctors make misdiagnoses that cause
• 15:04harm all the time. We're
• 15:06not missing lupus nephritis. Well,
• 15:08we are, but we're also
• 15:08missing, like, heart attack and
• 15:10stroke.
• 15:11This is from Andy Auerbach's
• 15:13group. Fantastic
• 15:14study that looked at
• 15:17very specific cognitive errors because,
• 15:19you know, one of the
• 15:19challenges always been, yeah, there's
• 15:21there's diagnostic errors. Of course,
• 15:22they are. But is that
• 15:23because of our human brains,
• 15:25or is that because of
• 15:26other systems factors? I mean,
• 15:27we all know that
• 15:29getting outside records can be
• 15:31difficult. There's all sorts of
• 15:32interruptions. So maybe it's just
• 15:34our fragmented health system that's
• 15:35doing this. And David Newman
• 15:37Joker, what what he sorry.
• 15:38Andy Auerbach, what he did
• 15:40is looked at fourteen different
• 15:41hospitals across the US, patients
• 15:43who were admitted to the
• 15:44hospital, appropriately triaged, meaning that
• 15:45they didn't decompensate within the
• 15:47first forty eight hours, and
• 15:48then either went to the
• 15:49ICU or died. And in
• 15:50those patients, a quarter of
• 15:52them had a diagnostic error,
• 15:54and a lot of these
• 15:56were severe. One fifth of
• 15:57those errors caused severe harm
• 15:58or death. And when we
• 16:00looked at though, he looked
• 16:01at the reasons, the most
• 16:02common cause was this, errors
• 16:04in human cognition by far,
• 16:05almost half, with the second
• 16:07biggest being in test interpretation
• 16:08and ordering, which is also
• 16:09a human cognitive error. So
• 16:11systems errors were actually much
• 16:12less common. Mind you, this
• 16:13is a very specific situation,
• 16:15so patients in the hospital
• 16:16who are decompensating, so I
• 16:17don't know that that's generalizable.
• 16:19But in this population, cognitive
• 16:21errors are common.
• 16:23So
• 16:24in my field,
• 16:25which I guess is medical
• 16:26education, reasoning has usually been
• 16:27within the med ed field,
• 16:29but we've looked historically at
• 16:31ways that we can make
• 16:32people better.
• 16:33Three big categories. So first,
• 16:36everyone here who is my
• 16:38age or less,
• 16:40I'm forty,
• 16:41has probably
• 16:42received some education about cognitive
• 16:44biases. Like, actually, raise your
• 16:46hand if you've ever been
• 16:46taught about cognitive biases.
• 16:49Yeah. I would expect to
• 16:50see half or more. So
• 16:52cognitive biases, of course, this
• 16:53is like your anchoring bias.
• 16:54This is recency bias.
• 16:56This has been a major
• 16:57trend in medical education over
• 16:58the last fifteen, twenty years.
• 17:00The problem is that when
• 17:01we study these things experimentally,
• 17:03teaching people about cognitive biases
• 17:05does not make them more
• 17:06less likely to have cognitive
• 17:08biases. I can teach you
• 17:09all about anchoring bias. You
• 17:11are still going to be
• 17:12anchor. You you are still
• 17:13going to anchor. You may
• 17:14have the metacognition that you're
• 17:15doing that, but you can't
• 17:17stop. And why? Like, sometimes
• 17:19I'm pretty amazed. Like, look
• 17:20where we are. Look at
• 17:21this technology. I am like
• 17:23a relatively hairless ape that
• 17:26evolved over fifty million years
• 17:27ago on the plains of
• 17:28Africa. Our brains are not
• 17:29evolved to work in this
• 17:31highly technical world. So it
• 17:33is no not surprising that
• 17:34there are cognitive biases. These
• 17:35are the shortcuts that we
• 17:37evolved with.
• 17:38It it's just part of
• 17:39being human, so that doesn't
• 17:41work. Number two is education
• 17:43about debiasing strategies. This is
• 17:45very popular and kind of
• 17:47optimistic because it works. So
• 17:48if you think about deliberative
• 17:50practice, the literature on deliberative
• 17:51practice,
• 17:52you know, this would be
• 17:53the Malcolm Gladwell ten thousand
• 17:55hours,
• 17:57even though that has been
• 17:59somewhat debunked.
• 18:00But, like, those psychological studies,
• 18:03if you were to spend
• 18:04five hours every single week
• 18:06going over
• 18:07every single pathologic diagnosis you
• 18:09made, reflecting on what you
• 18:10could have done better, seeing
• 18:11how the patients did, you
• 18:13will get better at the
• 18:14diagnostic process. This has been
• 18:16studied in internists.
• 18:18It works.
• 18:19No one has five hours.
• 18:21And if you were to
• 18:21spend five hours doing that,
• 18:23your bosses would get mad
• 18:24at you for not closing
• 18:25out enough cases or for
• 18:26me, like, my my length
• 18:27of stay would go up.
• 18:28Like, there's our system does
• 18:30not allow for these debiasing
• 18:32strategies as effective as they
• 18:33are. I mean, to this
• 18:34day, I keep a follow-up
• 18:35list. I have ten patients
• 18:36on it because that's all
• 18:37I can put because I
• 18:38don't have time. And whenever
• 18:39I put a new patient
• 18:40on, I have to take
• 18:40an old patient off. I
• 18:41spend maybe thirty minutes a
• 18:43week. I don't even know
• 18:44if it makes me that
• 18:44much better, but at least
• 18:46I try. And even then,
• 18:47it's really hard to do
• 18:49that. So that brings us
• 18:50to number three,
• 18:52which, doctor Homer mentioned, AI.
• 18:54We've been talking about AI
• 18:55for a long time.
• 18:57So artificial intelligence has been
• 18:59the third historical and actually
• 19:01the oldest way to make
• 19:02humans better. So this is
• 19:04where,
• 19:06doctor Homer was talking about
• 19:07nineteen seventy nine. I assume
• 19:09they were talking about internist
• 19:10one probably,
• 19:11but we've been talking about
• 19:12this for over a century.
• 19:14So this is the oldest
• 19:15quote I've ever found about
• 19:17artificial intelligence in medicine from
• 19:18Bernard Shaw. So talking about
• 19:20reflecting on how, like, no
• 19:22one does math anymore in
• 19:23in bit, like, old fashioned
• 19:24spreadsheets, paper where we had
• 19:25to do all this math.
• 19:26By the early twentieth century,
• 19:27there were counting machines that
• 19:28did all that. And he
• 19:29says, in the clinics and
• 19:30hospitals of the near future,
• 19:31when they quite reasonably expect
• 19:33that the doctors will delegate
• 19:34all the preliminary work of
• 19:35diagnosis to machine operators.
• 19:38And then he says, the
• 19:38observation of the symptoms is
• 19:40extremely fallible, depending not only
• 19:42on the personal condition of
• 19:43the doctor who has possibly
• 19:44been dragged to the case
• 19:45by his night bill after
• 19:46an exhausting day. I love
• 19:47that that also hasn't changed
• 19:49in over a century. But
• 19:50upon the replies of the
• 19:51patient to questions which are
• 19:52not always properly understood and
• 19:54for lack of the necessary
• 19:55verbal skill could not be
• 19:56properly answered if they were.
• 19:58From such sources of error,
• 19:59machinery is free. So when
• 20:01I talk about artificial intelligence,
• 20:02we are not talking about
• 20:03anything new. These ideas have
• 20:05been around for a long
• 20:06time. And depending on how
• 20:08you define a diagnostic AI,
• 20:11the first artificial intelligence in
• 20:12medicine was made about six
• 20:14months after the first electronic
• 20:15computer was made. So we've
• 20:17been working on this for
• 20:18a long time.
• 20:19Historically,
• 20:20AI clinical decision support
• 20:22has had really big impacts,
• 20:24though, in limited domains. One
• 20:26of my favorite examples is
• 20:27AAP help. This is, sometimes
• 20:29called the Leeds abdominal pain
• 20:30score is for patients presenting
• 20:31with an acute abdomen.
• 20:33A multicenter RCT showed a
• 20:35mortality
• 20:36decrease in acute abdomens of
• 20:38twenty two percent.
• 20:39That is huge. It's a
• 20:41multicenter RCT. It was so
• 20:43important that the US Navy
• 20:45considered it an essential part
• 20:46of our nuclear shield because
• 20:47they put it on all
• 20:48the, the submarines with ICBMs.
• 20:50Because if a semen had
• 20:50an acute abdomen, you needed
• 20:52to know whether to take
• 20:52that submarine and attack that
• 20:54semen, in which case part
• 20:55of your nuclear nuclear umbrella
• 20:56went away. And then, of
• 20:57course, there's, Internus One. I
• 20:59don't know if anyone here
• 21:00knew Jack Myers, but Internus
• 21:01One is one of the
• 21:02coolest AI systems of all
• 21:03time. It was modeled on
• 21:04the brain of Jack Myers,
• 21:06sometimes called so his people
• 21:07who knew him, Black Jack.
• 21:09He is the reason we
• 21:10no longer have oral boards
• 21:11in medicine. He's an eidetic
• 21:12memory chair of medicine at
• 21:13Pitt. But an AI system
• 21:15that was based on his
• 21:15brain at the year nineteen
• 21:16eighty two could solve the
• 21:18New England Journal of Medicine
• 21:19clinical pathological conference better than
• 21:21any human could. So not
• 21:23new technologies.
• 21:24Now
• 21:25one of the sad things
• 21:26is that the technology really
• 21:27stagnated. It worked really well.
• 21:30A lot of these computational
• 21:31strategies worked really well in
• 21:32narrow
• 21:33categories,
• 21:34but from roughly nineteen ninety
• 21:36to the year twenty twelve,
• 21:38there wasn't really any change
• 21:40in the performance. There's actually
• 21:41a review from twenty twelve
• 21:42that shows that differential generators
• 21:44in twenty twelve worked exactly
• 21:45as well as those that
• 21:46had come twenty years before.
• 21:47So real stagnation of the
• 21:49computational techniques.
• 21:50Now
• 21:52language models are the most
• 21:54exciting technology in the clinical
• 21:56reasoning space to come out
• 21:57in
• 21:59forty years, basically since the
• 22:00early nineteen eighties. I'm I'm
• 22:02going to briefly explain
• 22:04what they are, how they
• 22:05work, and why they might
• 22:06be exciting before actually demonstrating,
• 22:09like, how it works on
• 22:10a real case. So a
• 22:12language model, if, anybody, I
• 22:14should say, has everybody, who
• 22:15here has not used a
• 22:16language model before?
• 22:19Oh. A couple people. So
• 22:20most everyone, I imagine, has
• 22:21at least put something into
• 22:23ChatGPT before?
• 22:24Yeah. I I think they're
• 22:26pretty ubiquitous. I I believe
• 22:27ChatGPT was the most downloaded
• 22:29app in history. So, presumably,
• 22:30most people have. So how
• 22:32how does a language model
• 22:33work? So it is a
• 22:35it's a type of neural
• 22:35network. It's actually a transformer.
• 22:37But, effectively,
• 22:38it's autopredict on steroids. So
• 22:41let's say I'm sure everyone
• 22:43has done this. You go
• 22:43to Google. You type in
• 22:44pathologists are, and it predicts
• 22:46a bunch of different words.
• 22:47If I type in internal
• 22:48medicine physicians are, it'll probably
• 22:49say, like, nerdy or awkward
• 22:51or maybe it'll say smart,
• 22:53but it has a lot
• 22:53of predictions on what the
• 22:54next word is based on
• 22:56the tens of thousands of
• 22:57searches that have come before.
• 22:58A language model takes that
• 22:59fundamental technology
• 23:01and puts it on steroids.
• 23:03So it takes
• 23:04we don't not all of
• 23:05human text, but virtually all
• 23:07of human text. We don't
• 23:08know what's in the training
• 23:08corpus of the latest foundation
• 23:10models, but we do know
• 23:11for GPT three class models,
• 23:12which is all of the
• 23:13Internet,
• 23:14a lot of pirated books,
• 23:16a lot of human textual
• 23:17material. We also know from
• 23:18news reports that these companies
• 23:19are doing things like scraping
• 23:21YouTube videos, scraping podcasts. They're
• 23:23hungry for textual data. So
• 23:26this is the script, by
• 23:26the way. Huge nerd here,
• 23:27internal medicine physician. Shocker. This
• 23:29is a script for Star
• 23:30Trek two, the wrath of
• 23:31Khan, which is because I'm
• 23:32a Trekkie, but it's also
• 23:33because GPT two had a
• 23:34very famous experiment where you
• 23:36trained it only on the
• 23:37scripts from Star Trek because,
• 23:38again, the people who built
• 23:39these models were nerds, and
• 23:41someone actually built a dataset
• 23:42only on scripts of Star
• 23:43Trek, which I love.
• 23:44So
• 23:45it breaks every
• 23:47single,
• 23:49word, every single piece of
• 23:50text into units called tokens.
• 23:52And a token is a
• 23:53piece of a word. It
• 23:55is a basic semantic unit
• 23:57of a language model.
• 23:58And just like in the
• 23:59example where you type in
• 24:00pathologist r and it predicts
• 24:02the next word, it predicts
• 24:03the next token based on
• 24:04this huge training set of
• 24:06information.
• 24:07What makes it a large
• 24:08language model is that it's
• 24:10based on a technology called
• 24:11a transformer, which doesn't just
• 24:13predict the next token,
• 24:14but it it predicts the
• 24:15next token in a vector
• 24:16string of tokens, which means,
• 24:18semantically, it's predicting the next
• 24:20word in the context of
• 24:21a sentence, in the context
• 24:22of a paragraph, in the
• 24:23context of an entire book.
• 24:25And because these are so
• 24:26large and computationally intensive, it's
• 24:29doing these these calculations. You
• 24:30know, you go to Google,
• 24:31you type it once, that's
• 24:32one. It's doing that four
• 24:33hundred billion times per parameter.
• 24:36So massive amounts of calculations
• 24:38to figure out what the
• 24:40next token in a string
• 24:41should be.
• 24:42And now
• 24:44I, as somebody well, I'm
• 24:45losing my voice. I guess
• 24:46I'm not surprised. I've been
• 24:47talking nonstop for the last
• 24:48twenty four hours. As someone
• 24:49who studies human reasoning,
• 24:51humans are both, like, wonderful
• 24:52creatures, and we're
• 24:54we're terrible sometimes. Right? So
• 24:57we wrote the declaration of
• 24:58the rights of man. We
• 24:59wrote the, the Torah. We
• 25:00wrote the Bhagavad Gita. We
• 25:02also wrote Mein Kampf and
• 25:03the website four Chan, and
• 25:04all of those are encoded
• 25:06in language models. So the
• 25:08final step that makes these
• 25:10models there is a pretraining
• 25:11step that I won't go,
• 25:11but the final step that
• 25:12makes them so eerily human
• 25:14is
• 25:14reinforcement learning through human feedback.
• 25:16Sometimes it's called fine tuning.
• 25:18There's many ways to fine
• 25:19tune, though. But the idea
• 25:20here is that a human
• 25:21being actually sits down and
• 25:22talks a language model and
• 25:24says, oh, it's a Skinner
• 25:25box. Right? You did good
• 25:26language model, you get a
• 25:27cookie, or you did bad,
• 25:28you get an electric shock.
• 25:29It's a one or a
• 25:30zero, but same idea.
• 25:31And then over time, through
• 25:33all these training cycles, it
• 25:34makes them remarkably human. So
• 25:36a fun fact about the
• 25:37word delve, everybody knows that
• 25:38ChatJPT loves to say, let's
• 25:39delve into that. Right? And
• 25:41there was a great study
• 25:41that was just published, well,
• 25:43about a year ago that
• 25:44showed that in the scientific
• 25:45literature, the word delve was
• 25:46almost never used, and now
• 25:47it's incredibly used because everyone's
• 25:48using ChatJPT to help write
• 25:50their papers. Well, why is
• 25:51delve in there? It turns
• 25:53out that OpenAI
• 25:54used contract workers in Nigeria
• 25:56and Kenya to do most
• 25:57of their RLHF.
• 25:58And DELV sounds weird in
• 25:59American English, but in Kenyan
• 26:01and Nigerian English, DELV is
• 26:03commonly used. It's not doesn't
• 26:04sound weird to them. So
• 26:06now DELV is a big
• 26:07part of American scientific literature
• 26:10because of the RLHF of
• 26:12a language model of contract
• 26:13workers in Kenya and Nigeria.
• 26:15So just one of those
• 26:16fun things. And, of course,
• 26:18what you get is something
• 26:19that is remarkably human. You
• 26:20know, I tell Chat GPT
• 26:22I'm doing community theater. I
• 26:23wanna do the wrath of
• 26:24Khan. And it says, farewell,
• 26:25noble admiral. Hold not your
• 26:26breath. The enterprise cannot save.
• 26:28She's marked for death. From
• 26:29the sky, soon shall be
• 26:30torn asunder, a fiery end,
• 26:31a final echoing thunder. And
• 26:33Shakespeare. Right? So he still
• 26:34goes, Khan.
• 26:36So I did just do
• 26:37this so I could do
• 26:37this meme. You're welcome. But
• 26:39to also point out that
• 26:41the technology that is underneath
• 26:43this, a transformer,
• 26:44can be used to do
• 26:46output of any human creative
• 26:47output. This is just DALL
• 26:49E three. It's captain Kirk,
• 26:50at the Globe Theater. DALL
• 26:51E three is what's called
• 26:52a diffusion model. There are
• 26:54now video models. If anyone
• 26:55has seen Sora, that's the
• 26:56OpenAI video model. There's this
• 26:58crazy thing. Like, there are
• 26:59video games that are diffusion
• 27:00video games, meaning that you,
• 27:02like, walk through a video
• 27:03game, and it renders
• 27:04every single scene from an
• 27:06AI, like, hallucination. And then,
• 27:08of course, most concerningly, you
• 27:09can clone people's voices really
• 27:11effectively now, about ninety seconds
• 27:13of audio, and you can
• 27:14sound like anybody. And as
• 27:15someone who has, like, thirty
• 27:17hours of podcasting out, someone
• 27:18could easily clone me. And
• 27:19so if I call you
• 27:20up, don't trust that it's
• 27:21me, especially if I'm trying
• 27:22to get your Social Security
• 27:23number.
• 27:24Okay. Why does this matter
• 27:25for diagnosis?
• 27:27Because that's cool. Right? I,
• 27:29I've told a couple people
• 27:30this because I was working
• 27:31I I got into all
• 27:32of this research because I
• 27:33was working on my second
• 27:34book, which was about clinical
• 27:35reasoning. And I actually used
• 27:36I talked to a bunch
• 27:37of data scientists before. I
• 27:38had used GPT three before,
• 27:41like, CATGPT was released. And
• 27:43I thought it was stupid,
• 27:44and it wasn't gonna accomplish
• 27:45anything, and I was not
• 27:46impressed at all. So
• 27:47with that caveat, maybe you
• 27:49shouldn't listen to anything that
• 27:50I say. But, like, why
• 27:51do language models appear to
• 27:53work in diagnosis? So this
• 27:54is a real case of
• 27:55mine. This is a patient
• 27:56in whom I made a
• 27:57diagnostic error.
• 27:58The patient died.
• 28:00And what I will say
• 28:01is that the patient was
• 28:03like, I am friends with
• 28:04his wife. I have permission
• 28:05from his family. He was
• 28:06also a huge trekkie. So
• 28:07this patient had night sweats,
• 28:10monocytosis,
• 28:11a daily fever, ground glass
• 28:12opacities on the X-ray. He
• 28:13had been treated for bladder
• 28:14cancer, like, four or five
• 28:15years before, BCG.
• 28:18He did not improve with
• 28:19antibiotics, which is when I
• 28:20met him in the hospital.
• 28:21He still had the fevers
• 28:22despite high dose antibiotics.
• 28:23Liver enzymes were crazy. And
• 28:24then he had spinal hardware
• 28:26in his back from previous
• 28:27surgeries. And,
• 28:29I know, about two weeks
• 28:30before JPT four was released,
• 28:32I got back from the
• 28:33state lab what he had
• 28:34actually died from. So this
• 28:35poor man, we thought he
• 28:35had culture negative endocarditis.
• 28:37He died from a freak
• 28:38case of m bovis bacteremia,
• 28:40presumably reactivated from his BC
• 28:42BCG treatment. Quite rare,
• 28:45but misdiagnosis.
• 28:46It wasn't just me. Like,
• 28:47I was the the hospitalist
• 28:48with the residents. Like, there
• 28:49were a lot of doctors
• 28:50involved. But, you know, one
• 28:51of the first things that
• 28:51I did when GPT four
• 28:53came around is I I
• 28:54asked for a second opinion
• 28:56based on my problem representation
• 28:57at the time. And what
• 28:59happens is that it says,
• 29:01you know, the the number
• 29:02one diagnosis was what the
• 29:03patient died from, and the
• 29:04number two diagnosis was what
• 29:06I was wrong about.
• 29:08So, you know,
• 29:09this is me in twenty
• 29:11twelve, and my first thought
• 29:12is what if I had
• 29:13had this technology six months
• 29:14ago?
• 29:15Would this have changed anything?
• 29:17And we actually don't know
• 29:18the answer because second opinions,
• 29:20we haven't there's not great
• 29:21studies of second opinions, at
• 29:23least in internal medicine. We
• 29:24do know
• 29:25that there are discrepancies that
• 29:27being quite large, these are
• 29:28pathological diagnoses, so these are
• 29:30actually pathology studies.
• 29:32There's only one good perspective
• 29:33study on this from the
• 29:34Netherlands, which did find, like,
• 29:37frequent switching of diagnoses
• 29:39and improved symptoms. The patients
• 29:41actually did better when they
• 29:42got a second opinion.
• 29:43But, you know, the data
• 29:45is pretty early. So how
• 29:47do language models presumably do
• 29:49a good job at diagnosis?
• 29:51Well,
• 29:52our we've done some cool,
• 29:53like, ablation studies. Like, I
• 29:55my job is really cool.
• 29:56I basically gotta do psychology
• 29:57on both humans and machines,
• 29:58but we can do ablation
• 29:59studies where we knock out
• 30:00parts of the language model
• 30:01to figure out what's going
• 30:02on. And what it appears
• 30:04to happen is is that
• 30:05the reason that language models
• 30:06can make diagnoses is because
• 30:07of their similarity
• 30:09to how human brains make
• 30:10diagnoses. Right? So if you
• 30:11think about token prediction, you
• 30:13think about the log probs
• 30:14of different words,
• 30:16that's what a script is.
• 30:17So LLMs are basically
• 30:20system one on steroids, and
• 30:22they encode far more knowledge
• 30:24than we do.
• 30:25They are not perfect, but
• 30:27it does appear that this
• 30:28is what gives them their
• 30:29remarkable abilities in diagnosis.
• 30:31So with that, wait, who
• 30:33is who is my volunteer
• 30:34internist for this?
• 30:36Yeah.
• 30:38That's it. You work on
• 30:39Yelp?
• 30:40So I am going to
• 30:42show an example of
• 30:44how this technology works.
• 30:46Oh, jeez.
• 30:47So no. No. Five years.
• 30:48It's I'm sorry. No. This
• 30:51is, so I'll just explain
• 30:53what we're doing. So right
• 30:54now, if you go to
• 30:54the BI and you get
• 30:56admitted to the emergency room
• 30:57and you meet certain inclusion
• 30:58or exclusion criteria, you will
• 30:59be pulled into our data
• 31:00pipeline
• 31:01to study the effect of
• 31:04different,
• 31:06second opinions at different what
• 31:07we call diagnostic touch points.
• 31:09Because one of the ideas
• 31:10is when you evaluate these
• 31:11things, they,
• 31:13it depends on the information
• 31:14density. Right? Real diagnosis, you
• 31:16don't get a case vignette.
• 31:17You're often operating in
• 31:19poor information settings. So what
• 31:21we are doing here is
• 31:22I am going to, walk
• 31:24you through this is all
• 31:25real material. I've stripped it
• 31:26of PHI. This is a
• 31:27real patient.
• 31:28Just so you know, it's
• 31:29not a I intentionally just
• 31:31chose a random patient. This
• 31:32is not a zebra. Okay.
• 31:34This is not like a
• 31:34CPC. So what I want
• 31:36you to do is this
• 31:37is the information that was
• 31:39available to the port emergency
• 31:40room resident, and I want
• 31:41you to tell me what
• 31:42your thinking is, what you
• 31:44would wanna do, and then
• 31:45I'm going to ask the
• 31:46model, and you can tell
• 31:47me how that changes your
• 31:48thinking.
• 31:49Okay. So I I'll read
• 31:50it out loud if you
• 31:51want even though I'm losing
• 31:52my voice. So this is
• 31:53a a young woman who
• 31:54walks into our emergency room.
• 31:56This is the ED triage
• 31:57note. So this is someone
• 31:58this history is taken not
• 31:59in the emergency room, but
• 32:00in the waiting room, and
• 32:01a nurse has taken this.
• 32:03So,
• 32:04chief complaint, they have to
• 32:05put in a ICD code.
• 32:07So this is chest pain,
• 32:08tachycardia.
• 32:09The triage history, patient reports
• 32:10a new PE diagnosis and
• 32:11left lower extremity DBT at
• 32:13an outside hospital five days
• 32:14ago. Put an Eliquis ten
• 32:15BID since four days ago.
• 32:17Patient now arrives here with
• 32:18worse than chest pain, cough,
• 32:19and tachycardia.
• 32:20The nurse appropriately put this
• 32:22to the highest severity level,
• 32:24one, which means that the
• 32:25doc she goes back immediately
• 32:26and the doctor sees her.
• 32:27And vitals, fever, one zero
• 32:29one. Heart rate, one forty.
• 32:30Reperatory rate, twenty six. Blood
• 32:32pressure and o two sats
• 32:33are fine.
• 32:35K.
• 32:38So you want me to
• 32:39Yeah. Just say what you
• 32:39would what you would do.
• 32:42Okay. So,
• 32:43let's get a chest x-ray.
• 32:46Let's get an eye
• 32:49Let's get a chest x-ray.
• 32:51Let's
• 32:51get,
• 32:56I'm worried she is a
• 32:58I'm worried she still has
• 32:59a PE. Yeah. I'm worried
• 33:00she has either,
• 33:02I'm worried she has a
• 33:03new infection now for some
• 33:04reason, and I'm worried she
• 33:05has, like, a,
• 33:07like, a dissection possibly. Correct.
• 33:10So so the top your
• 33:11top worries are new PE
• 33:13or worse than PE infection,
• 33:15something like an aortic dissection.
• 33:20Me and Max don't get
• 33:21along.
• 33:23Okay.
• 33:25So one of the questions
• 33:26that I always get when
• 33:27people come to my lab
• 33:28is they're like, can I
• 33:28see the AI? Because they
• 33:29think there's something really cool
• 33:30when you see an AI.
• 33:31The reality is that it's
• 33:32a it's an Excel spreadsheet.
• 33:34It's a JSON database. So
• 33:36it's not exciting. So what
• 33:37I'm doing here this is
• 33:38the model I showed you
• 33:38earlier. Is this is not
• 33:40the model we're using. In
• 33:41reality, I'm using a llama
• 33:42three model. This is o
• 33:43one, which is the latest
• 33:45and greatest model from OpenAI
• 33:47that really freaks me out.
• 33:48But, we'll see what it's
• 33:50going to show. And to
• 33:51be clear, I haven't run
• 33:52this three zero one yet,
• 33:53so I don't know what
• 33:54it's going to show.
• 33:55Okay.
• 33:56So
• 33:58wow. It's going fast.
• 33:59It agrees with you that
• 34:01number one on this differential
• 34:02is a worsening pulmonary recurrent
• 34:03or worsening pulmonary embolism. Infection
• 34:05is the number two. It's
• 34:06picking up that fever. Could
• 34:07this be a pneumonia? Number
• 34:09three, it's now considering also
• 34:10pericarditis.
• 34:11Could pericarditis be going on?
• 34:13Fair enough.
• 34:15ACS, it's considering.
• 34:17Again, infection,
• 34:19and then subtherapeutic
• 34:20anticoagulation,
• 34:22pneumothorax.
• 34:22These are all very unlikely
• 34:24or COVID nineteen infection. And
• 34:25it also wants very similar
• 34:27things to you. It wants
• 34:27an EKG, a chest X-ray.
• 34:28It wants a repeat CTA.
• 34:30This is this is its
• 34:31management plan. What what do
• 34:32you does any of this
• 34:33change your thinking?
• 34:35No.
• 34:37Is it helpful?
• 34:39It's It's okay to say
• 34:41no. I mean, it just
• 34:42kind of confirms
• 34:43what I was
• 34:45thinking anyway, I guess. Yeah.
• 34:46So it's it's confirmatory. It
• 34:47makes you more confident in
• 34:48what you were thinking.
• 34:51Okay. On to the next
• 34:52aliquot. So this is very
• 34:54unique to our workflow. So
• 34:55what happens next is the
• 34:57poor ED resident who presumably
• 34:58has like twenty other patients
• 35:00has to come immediately to
• 35:01see this patient because it's
• 35:02an ESI of one.
• 35:03The ED resident, she writes,
• 35:05patient is a young female
• 35:06presented with the ED for
• 35:07chest pain and worsening shortness
• 35:09of breath. Patient notes that
• 35:10she has a history of
• 35:11lupus, was diagnosed with a
• 35:12PE five days ago after
• 35:13having a CT for shortness
• 35:14of breath and chest pain,
• 35:15started on Eliquis. I think
• 35:17all of this is the
• 35:17same. Today, she had worsening,
• 35:19shortness of breath, chest pain,
• 35:20worsening palpitations, which prompted her
• 35:22to present to the ED.
• 35:23And then she was triggered
• 35:24at all of this, we
• 35:25already know. So this is
• 35:26all other stuff. She's a
• 35:26history of lupus. She's a
• 35:28history of lupus. So that's
• 35:29and that this just started
• 35:30today are the additional things
• 35:32that the ED resident picked
• 35:33Does any of that change
• 35:34your thinking?
• 35:35I'm worried
• 35:36now more that she's got,
• 35:38like, a cons like, a
• 35:38constrictive pericarditis
• 35:41or
• 35:42but the the PE is
• 35:43still number one. Yeah.
• 35:45So you're considering other things,
• 35:46but it hasn't really changed
• 35:48yet.
• 35:50Okay. Let's see. Oops.
• 35:56You would think I know
• 35:57how to use a Mac.
• 35:58It's embarrassing.
• 36:05So what I love yeah.
• 36:07Turn this off. What I
• 36:08love about this is you
• 36:09can actually see what the
• 36:10model itself is thinking. So
• 36:11the model, again, similar to
• 36:12me, the fact that it
• 36:13started on just one day
• 36:14means that something like, that's
• 36:16acute.
• 36:17So it's changing its thinking
• 36:19also. Let's see. Did it
• 36:20pick up on the lupus
• 36:21here? Yes.
• 36:24Ah, now that's this is
• 36:25why I like seeing its
• 36:26thinking because you can see
• 36:27it's like, okay. A clot
• 36:29in lupus, could it be
• 36:30antiphospholipid antibody syndrome? So still,
• 36:33like you, it thinks that
• 36:34a recurrent or worsening PE
• 36:35is still the number one
• 36:36thing on the differential. It's
• 36:37now considering antiphospholipid
• 36:39antibody syndrome on the differential,
• 36:41which I think is reasonable
• 36:42given that additional history of
• 36:44lupus.
• 36:45Pericardial fusion. So what you
• 36:47were mentioning, could this be
• 36:48pericardial fusion? It's actually worried
• 36:50about tamponade. I don't know
• 36:51why because the blood pressure
• 36:52is normal, but it mentions
• 36:53that. Blood pressure is normal.
• 36:55I so no hypotension.
• 36:57Arrhythmia, pericarditis,
• 36:59still considering ACS and pneumothorax.
• 37:02It does put panic attack.
• 37:03And then what it wants,
• 37:04I don't think it's changing
• 37:05what it wants. It wants
• 37:06an EKG, a CTPA, a
• 37:07TTE, chest x-ray, all the
• 37:09standard things.
• 37:10Does this change like, did
• 37:12this
• 37:13second opinion change your thinking
• 37:14at all?
• 37:16Really? Does it make you
• 37:17feel more confident or not?
• 37:20Yes.
• 37:22Yeah. It makes me feel
• 37:23yeah. I guess it makes
• 37:24me feel more confident.
• 37:26It's okay to say nothing.
• 37:28Priest health care is last
• 37:29week.
• 37:31Well, that's I mean, the
• 37:32we get we'll get into
• 37:33this. Right? So if you
• 37:34were to give AI generated
• 37:36second opinions even if very
• 37:37effective, it might just lead
• 37:38to overtreatment of everything.
• 37:40So a lot of what?
• 37:41We'll talk we'll talk about
• 37:42that after. Because this is
• 37:43a this is a big
• 37:44concern about when and in
• 37:46what situation you should do
• 37:47this. Okay. Exam isn't gonna
• 37:49help you any. I'll put
• 37:50it in the system, but
• 37:51this is the ED documented
• 37:52exam. K. They document it
• 37:54as completely normal. Okay. I
• 37:56this is not one of
• 37:57my like, literally, this is
• 37:58a random patient that
• 38:01I picked out. So I
• 38:02have no idea if it
• 38:03was actually normal. I'm just
• 38:04gonna put that in there
• 38:05so it knows. And then
• 38:05we'll move on to the
• 38:06next piece of information, which
• 38:07is the imaging. So
• 38:09the resident
• 38:10orders actually pretty much everything
• 38:12that you ask for. They
• 38:14also order an EKG. It
• 38:15is a problem in my
• 38:16data pipeline that I'm not
• 38:17able to pull in EKGs
• 38:18yet, so there is no
• 38:18EKG here. But, X-ray shows
• 38:20bibasilar opacity is compatible with
• 38:22small bilateral pleural effusions.
• 38:26CTA,
• 38:27right,
• 38:28lower,
• 38:29right lower lobar and segmental
• 38:31PE without right heart strain
• 38:32and a small pericardial effusion,
• 38:34and then this bilateral
• 38:36axillary lymph nodes,
• 38:38as well as this hypodensity
• 38:40in the liver. TTE is
• 38:41performed. Big picture, there's no
• 38:43strain seen on the TTE.
• 38:45No tamponade.
• 38:46But so, yeah, those images,
• 38:47does that change your thinking
• 38:48at all or pretty much
• 38:49where you are?
• 38:51It it I mean, it
• 38:53knocks down pneumonia
• 38:54a little bit. It knocks
• 38:56down, like
• 38:57she's sounds like her cardiopulmonary
• 38:59silhouette is normal, so, like,
• 39:00I'm not worried. Like, she's
• 39:01got a big
• 39:02pleural effusion or anything like
• 39:04that.
• 39:07Yeah. I think
• 39:08I'm still
• 39:09on the on the PE
• 39:11train. You're still on the
• 39:12PE train. Okay.
• 39:14Let me
• 39:16try to not freak out
• 39:17my AI model too much.
• 39:19Oh, why am I scrolling
• 39:21the wrong direction? Embarrassing Adam.
• 39:23But I want an EKG.
• 39:26I don't have it.
• 39:28This is epic. This is
• 39:29a snowflake problem.
• 39:31The EKGs are stored in
• 39:33another database, so they're actually
• 39:34not very easy to pull
• 39:35in until this is the
• 39:36problem. Until the cardiologist confirms
• 39:38the read, and then you
• 39:39can extract it. So because
• 39:40my data pipeline is running
• 39:42live, none of our patients
• 39:43have EKGs.
• 39:46So this is what we
• 39:47were talking about. So much
• 39:48so much of this comes
• 39:48to, like, understanding where the
• 39:49data comes from and the
• 39:50limitations.
• 39:51Okay. So let's see
• 39:54how the AI model has
• 39:55changed its thinking. So like
• 39:57Hume, it still thinks recurrent
• 39:59or worsening PE is the
• 40:00number one diagnosis.
• 40:01It still is worried about
• 40:04anaphosolipid antibody syndrome. But sorry.
• 40:07Can I just say Yeah?
• 40:07Please. Anaphosolipid syndrome is not
• 40:09causing her acute, like
• 40:12I know. So, like, it
• 40:13doesn't like, that's
• 40:15it can say that all
• 40:16at once, and it might
• 40:17be APS, but, like, she's
• 40:18still if she's the APS
• 40:20is causing something. Right. Right.
• 40:22Right. I it doesn't That
• 40:24doesn't help
• 40:25you.
• 40:26Well, it helps you down
• 40:27the line, but not It
• 40:28doesn't help you in the
• 40:29acute setting. Exactly.
• 40:31Pericardial effusion, possibly lupus related
• 40:33infection,
• 40:34pericarditis,
• 40:35arrhythmia.
• 40:36Yeah. None of these these
• 40:36are all pretty much things
• 40:37that were on your differential.
• 40:38Right?
• 40:40None of this changes your
• 40:41thinking at all. Doesn't. Except
• 40:42except to get annoyed because
• 40:44you're like
• 40:45even if it's anti antiphospholipid
• 40:47antibody syndrome, it's still a
• 40:48pee.
• 40:49Okay.
• 40:51Labs,
• 40:52I don't know that this
• 40:53is gonna help you much,
• 40:54but she admit our lactate
• 40:55cutoff is one point six,
• 40:56so this is a slightly
• 40:57elevated lactate. Sodium is one
• 40:59thirty one.
• 41:00These other labs are all
• 41:02relatively normal. A troponin was
• 41:04negative. That's a a negative
• 41:05proBNP.
• 41:06She is anemic. Hemoglobin is
• 41:08nine one.
• 41:09Her INR is elevated one
• 41:11point seven,
• 41:12with a PTT of thirty
• 41:13five.
• 41:14That diff is normal, and
• 41:16then when they repeated the
• 41:17lactic acid after fluids, it
• 41:18was one five, which is
• 41:20one below the cutoff. So
• 41:21it's normal in our system,
• 41:22and the repeat retardant was
• 41:23negative. I'm going to well,
• 41:25did the labs change anything
• 41:26for you? No. Yeah. I
• 41:27I wouldn't think they would.
• 41:29And let's see if they
• 41:30change anything via AI model.
• 41:42Come on. Show me what
• 41:43you're doing.
• 41:47You're running a different chat
• 41:49GPT.
• 41:50This is very slow compared
• 41:51to my experience.
• 41:53This is, a new model
• 41:54called o one. You can
• 41:55see what it's doing. It
• 41:57is using an internalized chain
• 41:58of thought process
• 41:59that, well, that's what it's
• 42:01doing. It's thinking it's think
• 42:02thinking through different steps. So
• 42:04that's why it's going so
• 42:05slow. In reality, we're running
• 42:06this on a core with
• 42:07a llama model that's like
• 42:08it's like that. It's very
• 42:09fast.
• 42:11Okay.
• 42:12So I'm gonna guess it's
• 42:13gonna say the same things,
• 42:14but let's check it out.
• 42:16Recurrent PE,
• 42:18it's it's not changing. Right?
• 42:19So it's basically saying the
• 42:20same thing.
• 42:23I don't think any of
• 42:23these are different.
• 42:25Okay.
• 42:28So I'll just go over
• 42:29the last bit because our
• 42:31final touch point is when
• 42:31medicine sees the patient, so
• 42:33when medicine actually admits the
• 42:34patient. So the, I bolded
• 42:36the relevant thing. So, the
• 42:38ED no. Sorry. The medicine
• 42:40intern sees the patient still
• 42:41in the ED, and the
• 42:42patient says,
• 42:44Oh, this happened to me
• 42:45before six years ago. I
• 42:47had a lupus flare with
• 42:48very similar symptoms,
• 42:49and, the medicine resident finds
• 42:51out that the patient
• 42:53outpatient rheumatologist
• 42:54for the last couple weeks
• 42:55has felt that she's having
• 42:56a lupus flare and has
• 42:57been modifying her medications
• 42:59and that these are the
• 43:00current medications. So methotrexate,
• 43:02twenty q seven days, which
• 43:03has just been increased, plaquenil,
• 43:05folic acid, and then apixaban,
• 43:07which is new.
• 43:08Does any of that
• 43:11change
• 43:13your
• 43:16thinking?
• 43:18It's okay to say no.
• 43:20No. It it it doesn't.
• 43:21It's just you know, I
• 43:22think there's
• 43:24two things going
• 43:27on.
• 43:30Do you want me to
• 43:30tell you what the final
• 43:32Like, I guess my question
• 43:33is, is it pleuritic?
• 43:34Is it pleuritic? Yeah. You
• 43:36wanna know what the final
• 43:36diagnosis was from the the
• 43:37medicine team? Is it is
• 43:39it a lupus flare? It's
• 43:40lupus flare. Yeah. So the
• 43:41the final diagnosis is actually
• 43:43that this the poor woman
• 43:44has pericarditis and pleuritis from
• 43:46a lupus flare and had
• 43:47a PE secondary to that.
• 43:49So it is two different
• 43:50things going on. And, you
• 43:52know, the ED team did
• 43:54everything completely appropriate. Right? When
• 43:56you have a patient like
• 43:56this come in, obviously, you
• 43:57wanna make sure they're not
• 43:58having something devastating. So it's
• 43:59not even that this patient
• 44:00was mismanaged
• 44:02in any way.
• 44:03It's just not what the
• 44:05initial diagnosis was. So I
• 44:06was more curious.
• 44:07Come on.
• 44:09It's very slow. It's being
• 44:10very finicky. So I'm curious
• 44:12what the AI model is
• 44:13going to end up saying.
• 44:15You can see how slow
• 44:16it is. This is all
• 44:17the different steps.
• 44:20Well
• 44:21Friction
• 44:23I did not see this
• 44:25patient. This is all mediated
• 44:27through the physical through the
• 44:28chart, and I will tell
• 44:29you that in our data
• 44:30pipeline, which pulls in only
• 44:31two different diagnostic
• 44:33physical exams, no one documented
• 44:35it, that doesn't mean she
• 44:36didn't have it.
• 44:40I highly doubt that the
• 44:41EP resident did the exam
• 44:42later. Okay. So let's see
• 44:43what the model says.
• 44:45Okay. So this is endo
• 44:47it it this is the
• 44:48right final diagnosis, which had
• 44:49lupus flare with psoriositis.
• 44:51And then the number two,
• 44:52the patient has, it's a
• 44:54pulmonary embolism. And, of course,
• 44:55what we found out or
• 44:57I'm not this patient's doctor,
• 44:58but what ended up happening
• 44:59is they get the out
• 45:00the outside CTA, and it
• 45:01shows that, in fact, the
• 45:03PE is no larger. It's
• 45:04even a little bit smaller.
• 45:05So this is not a
• 45:05recurrent PE. The symptoms were
• 45:07likely driven from pericarditis and
• 45:09pleuritis, so lupus serositis.
• 45:12The patient was worked up
• 45:13for antiphospholipid antibodies from all
• 45:14the tests were negative. So
• 45:16that's the that's the final
• 45:17diagnosis. So reflecting back, like,
• 45:19would this have been helpful
• 45:20if you were getting this
• 45:21and would have driven you
• 45:22in the wrong direction?
• 45:25I don't know if it
• 45:26would have driven me in
• 45:27the wrong direction, but it's
• 45:30it was confirmatory.
• 45:31But I mean,
• 45:33I think
• 45:35well, two things. One, we're
• 45:36in an ed setting. So,
• 45:38you know, I think the
• 45:39most important thing is ruling
• 45:41out the things that are
• 45:41gonna kill her in the
• 45:42next hour
• 45:44so, you know
• 45:46It might be a lupus,
• 45:47but you don't want to
• 45:48miss right you don't want
• 45:48to miss a pe or
• 45:49a dissection
• 45:51I guess I don't know
• 45:51if that's anchoring, but, like,
• 45:53that is my top things
• 45:54are I wanna make sure
• 45:55she's not tamponading. I wanna
• 45:57make sure she's not having
• 45:58a massive PE. I wanna
• 45:59make sure she's not having,
• 46:00like, a dissection
• 46:02or she's having, like, a
• 46:03big heart attack. So, like,
• 46:05other than that,
• 46:06it didn't really change anything.
• 46:08Because you would have done
• 46:09everything exactly the same. And
• 46:10you were considering those cannot
• 46:12miss diagnoses from the very
• 46:13beginning, obviously.
• 46:14Yeah. Yeah. I don't think
• 46:15it would have changed much.
• 46:17Would it have driven you
• 46:18in the wrong direction? Right?
• 46:18Would getting a second opinion
• 46:20like this have made you
• 46:21second guess yourself or I
• 46:23think it would have made
• 46:24me order more tests.
• 46:26And in in particular, like,
• 46:28in the ED, you would
• 46:28have ordered a bunch of
• 46:29those tests?
• 46:31Maybe not in the ED,
• 46:32but, like, it was I
• 46:33saw it. It was, like,
• 46:34get a cardiac MRI.
• 46:36I hope no one well,
• 46:37it wouldn't matter. Cardiology wouldn't
• 46:38do the cardiac MRI, but,
• 46:39yes, that is not an
• 46:41appropriate test for this workout.
• 46:42Yeah. But,
• 46:43yeah, maybe order more I
• 46:45I might have ordered more
• 46:46labs if I'm being honest.
• 46:47Yeah.
• 46:48Alright. Well, that's thank you
• 46:50very much. I'll give you
• 46:51a a hand. I was
• 46:52very
• 46:53sorry that I made you
• 46:54breakfast in general medicine again.
• 46:56You thought you escaped.
• 46:59No. That that's oops. Well,
• 47:01that's cool that you have
• 47:02a, one of these things
• 47:03here. So that's, I mean,
• 47:04this is an example
• 47:06of what it looks like
• 47:08in practice in a randomly
• 47:09selected case. And you can
• 47:10start to already see when
• 47:11you go through it some
• 47:12of the challenges of implementing
• 47:14a system like this. I'm
• 47:15gonna go over some of
• 47:16the data, including some of
• 47:17the new data before, seeing
• 47:18if anybody has any questions
• 47:19and before I lose my
• 47:20voice.
• 47:22So LLMs encode lots of
• 47:23knowledge.
• 47:24I'm sure that everyone saw
• 47:25that, you know, it can
• 47:26pass the USMLE.
• 47:27I don't care about this.
• 47:29You guys should not care
• 47:30about this either. It turns
• 47:31out, this is actually from
• 47:32some of our interesting ablation
• 47:33studies, that LLMs'
• 47:35performance on exams
• 47:37has less to do
• 47:38what they're doing is that
• 47:40they have learned the semantic
• 47:41structure of multiple choice questions,
• 47:43meaning that they are effectively
• 47:45good test takers.
• 47:46Some of my colleagues did
• 47:47a really cool experiment where
• 47:48they made up two new
• 47:49organ systems, and then they
• 47:50had test writers write up
• 47:51multiple choice questions with those
• 47:52fake organ systems. And the
• 47:54LLMs still did really well
• 47:55on it because they learned
• 47:56to understand what a question
• 47:58looks like and guess the
• 47:59right answer from that. And
• 48:00I think everyone here knows
• 48:01that if you're being honest
• 48:02with yourself about what multiple
• 48:03choice like, you start by
• 48:05excluding a couple things. We
• 48:06all know how that works.
• 48:07So none of that matters.
• 48:09This empathy thing, I think,
• 48:10is overplayed also. You should
• 48:12know that
• 48:13this is the justification for
• 48:15having LLMs write portal messages,
• 48:16that patients find their communication
• 48:18more empathetic.
• 48:19The standard, of course, is
• 48:21compared to a very overstretched
• 48:23PCP who's just trying to
• 48:24commute
• 48:25communicate your CBC results. So
• 48:26these are not actually empathy
• 48:28in person
• 48:29communications, but at least in
• 48:30written communications, people do find
• 48:32the LLM to be more
• 48:33empathetic.
• 48:34For what I care about,
• 48:36LLMs are able to make
• 48:37diagnoses
• 48:38on a lot of the
• 48:40benchmarks
• 48:41that, like, the field has
• 48:42accepted. LLMs have long since
• 48:44surpassed humans,
• 48:46but a lot of these
• 48:47are relatively artificial because they're
• 48:48very information dense settings, very
• 48:50complicated diagnoses,
• 48:52and a lot of when
• 48:53you look at the diagnostic
• 48:54errors are not coming from
• 48:56lupus nephritis. It's coming from
• 48:57people misdiagnosing
• 48:59common things.
• 49:01They have this is fascinating
• 49:02they have an emergent probabilistic
• 49:04reasoning,
• 49:05so there's no reason that
• 49:07semantic, like language should give
• 49:09you a probabilistic understanding of
• 49:11disease states, But in fact,
• 49:12this is studied with Dan
• 49:13Morgan. When you compare it
• 49:14to large groups of humans,
• 49:16they have a better sense
• 49:17of understanding the pretest probability
• 49:19of disease and how that
• 49:20changes with subsequent tests. That
• 49:22holds up
• 49:23pretty well. The post test
• 49:25probability of disease, it's not
• 49:27really any better than humans,
• 49:28but for after a positive
• 49:30test, but after a negative
• 49:30test, it's a lot better
• 49:31than us.
• 49:33They can forecast similarly well.
• 49:34So if you ask it,
• 49:36what do you think the
• 49:36percentage chance of the final
• 49:38diagnosis is? This this was
• 49:39done in neurologists, ID doctors,
• 49:41and pediatricians.
• 49:42It outperforms every single individual
• 49:44human in every single group
• 49:46of humans,
• 49:47only being beaten when you
• 49:48take the best groups and
• 49:49put them together.
• 49:52They are able to display
• 49:54reasoning.
• 49:55So when it comes to,
• 49:56like, how will you communicate
• 49:58with a human,
• 49:59there's this whole question of
• 50:00human computer interaction. What would
• 50:02an AI
• 50:06LLMs actual cases, as new
• 50:09information comes in and asks
• 50:10them to update their thinking
• 50:11and you compare that to
• 50:12humans, it outperforms humans consistently.
• 50:16It outperforms
• 50:17attendings who outperform residents, and
• 50:18there's no difference in efficiency,
• 50:20accuracy, quality, or cannot misdiagnoses.
• 50:22It does hallucinate more. So
• 50:24this is actually pretty high
• 50:25hallucination. Right? Right? It makes
• 50:26up stuff twelve percent of
• 50:27the times compared to only
• 50:28three percent of the times
• 50:29with humans. Now the hallucinations
• 50:31are relatively minor. Some of
• 50:32them are kind of funny.
• 50:33One of them was a
• 50:34patient who had diverticulitis,
• 50:36and the L. M. Wanted
• 50:37the human to keep gastroenteritis
• 50:39in mind because the patient
• 50:40had recently traveled to, Texas,
• 50:42and going to Texas was
• 50:43a risk factor for enterotoxigenic
• 50:45E. Coli, which I'm pretty
• 50:47sure is not true. So
• 50:48that is a hallucination,
• 50:49but probably not one that
• 50:51would harm the patient.
• 50:52This study was done by
• 50:53my colleagues at Google,
• 50:55very controversial when it came
• 50:56out because what they did
• 50:58is it's actually not a
• 50:59very high performing model. It's
• 51:00a palm two model, but
• 51:01the model itself could solve
• 51:02CPCs. You can see humans
• 51:03are on the bottom. It
• 51:04could solve clinical pathological conferences,
• 51:06and they randomized humans to
• 51:09either solve conferences themselves
• 51:11using Google search, using the
• 51:12AI model, or the AI
• 51:13model itself. And this was
• 51:15controversial when it came out
• 51:16because when you gave humans
• 51:17the AI model, it actually
• 51:18made the model not perform
• 51:19as well. So adding humans
• 51:21into the mix lowered performance.
• 51:23This is against
• 51:24the kind of standard precepts
• 51:26of the informatics field, so
• 51:27quite controversial when it came
• 51:28out. Unfortunately,
• 51:30I I we my group
• 51:31ran a large randomized control
• 51:33trial looking at very nuanced
• 51:34measures of reasoning in real
• 51:35cases, so not CPCs,
• 51:38and we found the same
• 51:38thing. The human performance is
• 51:41on the right by itself
• 51:42in blue.
• 51:43Humans using the AI are
• 51:44in green, and the AI
• 51:45model by itself is in
• 51:46red. So we found the
• 51:47same thing, and because we
• 51:49did a very nuanced measure,
• 51:50I can tell you why.
• 51:51And it's humans, when the
• 51:53AI model tells them that
• 51:54they're wrong, disregard those pieces
• 51:55of information. In particular,
• 51:57humans don't like
• 52:00the humans don't like an
• 52:01AI model critiquing
• 52:02or disconfirming the things that
• 52:04they think.
• 52:06Another randomized controlled trial that
• 52:08was just accepted, this is
• 52:09the one I just just
• 52:09heard from nature.
• 52:10So this is in management
• 52:12decisions. Management decisions are notoriously
• 52:14tricky to measure.
• 52:15In this one, LLMs
• 52:17did improve people's ability when
• 52:18they used it to make
• 52:19management decisions, when you randomized.
• 52:21But when we look at
• 52:21the subgroup, it's not what
• 52:23you would think. Like, people
• 52:24aren't using the LLM to
• 52:25say, what is the right
• 52:26dose of apixaban, or even
• 52:27should I give apixaban? What
• 52:28the LLM did was cue
• 52:30them to, for example, apologize
• 52:32after making a medical error
• 52:33or communicate better with other
• 52:35providers or take patient factors
• 52:37into account when following a
• 52:38likely cancerous nodule. So it
• 52:40actually improved performance
• 52:41not in, like, what we
• 52:42think of as the standard
• 52:43management
• 52:45domains, but in things that
• 52:46we think of humans are
• 52:47good at.
• 52:49A lot of the work
• 52:49that I'm doing with Google
• 52:50is on building models that
• 52:51can collect data. This is
• 52:53from the Omni system.
• 52:55This is standardized patients, not
• 52:57real patients, but a true
• 52:58Turing test where a
• 53:00standardized patients talk to a
• 53:01a terminal, and they don't
• 53:02know whether it's a human
• 53:03or an AI on the
• 53:04other side. And,
• 53:06on twenty six of twenty
• 53:07six patient domains, patients before
• 53:09the AI, and in twenty
• 53:11eight of thirty two axes
• 53:12from the physician graders,
• 53:14AI was preferred,
• 53:15and this held up in
• 53:16every single diagnostic category. So
• 53:19we're running this in clinical
• 53:20trials in human like in
• 53:21actual patients now, It's still
• 53:22performing quite well, but they're
• 53:24increasingly able to collect data.
• 53:27Now
• 53:28the, the the unpublished data
• 53:30that I'm about to show
• 53:31you is from my grad
• 53:31student that I put in
• 53:32this presentation this morning because
• 53:34the models continue to improve.
• 53:36And if you would ask
• 53:36me six months ago, I
• 53:37would say we're seeing convergence
• 53:39of model performance. I don't
• 53:40think there's going to be
• 53:41like, there's still gonna be
• 53:43incremental improvements,
• 53:44but this is for, this
• 53:46is for
• 53:47solving CPC. So this is
• 53:48one of these benchmarks that
• 53:49goes back almost sixty years.
• 53:51And the new models have
• 53:52surpassed everything that came before,
• 53:54and you can see humans
• 53:55are in brown in the
• 53:55bottom. This is the one
• 53:56that freaks me out because
• 53:58this is not an HCI
• 53:59study, right? This is just
• 54:00looking at the human baseline,
• 54:01but I showed you these
• 54:02are real cases for the
• 54:03diagnostic and management decisions.
• 54:05The colors are different, but
• 54:06these are the old graphs,
• 54:07and this is the new
• 54:08model on the left. And
• 54:09you can see that the
• 54:10new models are performing
• 54:13in
• 54:13a far better than any
• 54:15other system, not only much
• 54:16better than the humans, but
• 54:17better than the previous AI
• 54:18systems. So we're continuing to
• 54:20see performance gains.
• 54:22Eric Horvitz, the Microsoft group
• 54:24published their, med prompt follow-up
• 54:26on o one today, and
• 54:27they actually they came to
• 54:28the same conclusion as my
• 54:29paper, which is, like, these
• 54:30things have gotten so good
• 54:31that we need new benchmarks.
• 54:34We or clinical trials because
• 54:35they are outperforming everything that
• 54:37we throw at them.
• 54:39I can go over these
• 54:40quickly. In reality, so a
• 54:42lot of tools are now
• 54:43being used in clinical practice.
• 54:45They're actually kind of underperforming
• 54:46from what we were sold.
• 54:47So if you look at
• 54:48some of the early performance
• 54:49of AI,
• 54:51scribes, which I know you
• 54:52guys are using here at
• 54:53Yale and some of the
• 54:54clinics, some of the early
• 54:55studies actually suggest there is
• 54:56no efficiency gain because they
• 54:58hallucinate, and the doctors have
• 54:59to go back and check
• 55:01the models.
• 55:02And, yeah, people like it,
• 55:03but it's not really saving
• 55:05anybody time and, you know,
• 55:07people always care about money.
• 55:08It's not saving anybody any
• 55:09money either.
• 55:10The same thing is happening
• 55:12with the
• 55:13patient portal messaging. One of
• 55:15the very depressing things from
• 55:16the JAMA study on this
• 55:17is that it actually took
• 55:18more time, seven percent more
• 55:20time, physician time, when the
• 55:22AI wrote the initial drafts
• 55:23because it hallucinates or says
• 55:25something harmful, and the doctor
• 55:26has to go back and
• 55:27edit it. Again, the patients
• 55:28liked the responses more, but
• 55:29it took the doctors more
• 55:30time.
• 55:31And then what everybody should
• 55:33know, I don't think I
• 55:34need to say this, but
• 55:34LLMs are racist and sexist.
• 55:36They actually encode
• 55:38because they are trained on
• 55:39our language and then fine
• 55:40tuned by humans. They encode
• 55:41all of the biases that
• 55:43humans have. Now they do
• 55:45appear I I just published
• 55:46a study in, JAMA. They
• 55:47appear to be less racist
• 55:49and sexist than us, but
• 55:50they are still racist and
• 55:52sexist.
• 55:52So in a world where
• 55:54we're trying to get past,
• 55:55like, race based medicine,
• 55:57especially as LLMs get more
• 55:58and more powerful, we should
• 55:59know that they are showing
• 56:01human, not only cognitive biases,
• 56:03but racial and gender biases,
• 56:04which is concerning.
• 56:06And then we talked a
• 56:08little bit, but the you
• 56:08know, HCI is actually quite
• 56:10challenging because
• 56:12if used inappropriately, this technology
• 56:14probably will drive overtreatment.
• 56:16It
• 56:17different people need different opinions
• 56:19at different times. Like, a
• 56:20second opinion is not universally
• 56:21helpful.
• 56:22Also, HCI is unpredictable. There's
• 56:24a great study from some
• 56:24of my colleagues at MIT
• 56:26that showed that the best
• 56:27radiologists
• 56:28actually have their performance lowered
• 56:30by a high performing AI
• 56:31because they second guess themselves.
• 56:33So just because an AI
• 56:34model works well, even if
• 56:36it consistently works well, like,
• 56:38in silica, that in silico,
• 56:40doesn't mean that it's actually
• 56:41going to improve human performance
• 56:42because, you know, again, we
• 56:44are hairless apes that evolved
• 56:45to, like, live on it'd
• 56:47be hunter gatherers, and now
• 56:48we're trying to do complex
• 56:49medicine in the twenty first
• 56:50century. So,
• 56:52whew, I'm gonna lose my
• 56:53voice. That is it for
• 56:54this presentation. So if anybody
• 56:56has any questions or wants
• 56:57to talk about pathology, I
• 56:58am happy to, entertain them.
• 57:05And thank you very much.
• 57:08Are the new models based
• 57:10on the performance of the
• 57:11previous models formed from Zetta
• 57:13testing? So the new model,
• 57:15Owen, this is really interesting,
• 57:16has no new data in
• 57:18it. There is it is
• 57:19the same
• 57:20data as Ford Turbo. So
• 57:21the cutoff is like last
• 57:23year. So what in is
• 57:24improving its performance has nothing
• 57:25to do with the training
• 57:26data, but what they're doing
• 57:27is chain of thought. So
• 57:29if you get a model
• 57:29to speak out at SOTA,
• 57:30it does better. And what
• 57:31they've done is reinforcement learning
• 57:33for the chain of thought.
• 57:34So they're teaching it how
• 57:35to think out loud and
• 57:36then reinforcing that over time.
• 57:38So these models, these are
• 57:39all computational techniques. It has
• 57:40nothing to do with the
• 57:41underlying data, and there's no
• 57:43more scale. The parameters of
• 57:44the model are exactly the
• 57:45same, which is one of
• 57:46the reasons I'm so freaked
• 57:47out because I didn't think
• 57:47we could get such impressive
• 57:49performance gains without increasing the
• 57:51number of parameters.
• 57:54Yes.
• 58:12Yeah. Yeah. So this is
• 58:13a great question. Like,
• 58:15what would it look like
• 58:16in practice?
• 58:17So
• 58:18I'm assuming
• 58:19we're talking Epic here. Right?
• 58:21So the the reality of
• 58:23the situation is that Epic
• 58:24is working on clinical decision
• 58:26support software.
• 58:28This is not a huge
• 58:29priority. If you look at
• 58:30what Epic is working on,
• 58:31they're mostly efficiency. They're working
• 58:33on like tech summarization.
• 58:34However,
• 58:35Epic does make it fairly
• 58:37easy to have a data
• 58:38pipeline to put information in.
• 58:40So even at my own
• 58:41institution, we have a pipeline
• 58:41through Amazon Web Services that
• 58:41I can push a second
• 58:42opinion in through the chart,
• 58:42trivially
• 58:43second opinion in through the
• 58:44chart, trivially easy. Like, the
• 58:45any any health system could
• 58:46do this. Any third party,
• 58:47there are vendors right now
• 58:48who want to sell you
• 58:49this technology. No one should
• 58:50buy it, by the way,
• 58:52because this is not tested,
• 58:56and I'm pretty certain that
• 58:57it will lead to worse
• 58:58care if used, like, routinely
• 59:01on every single patient. So
• 59:02from a technological standpoint, you'd
• 59:04need, like,
• 59:06fifteen hours of a programmer's
• 59:07time to build a pipeline
• 59:08to do this. The question
• 59:10becomes, like, what are the
• 59:11other strategies that you're going
• 59:12to do to make sure
• 59:13that you're giving a second
• 59:14opinion to the right person
• 59:15at the right time? At
• 59:16the BI and what we're
• 59:17doing through the, Home Run
• 59:19Network is we're looking at
• 59:20serving second opinions at clinical
• 59:22decompensation. So at the moment
• 59:23that a patient's about to
• 59:24go to the ICU, based
• 59:26on this logic that the
• 59:27patient is already really sick,
• 59:30we do diagnostic timeouts anyway.
• 59:32So this is just another
• 59:33part of the diagnostic time
• 59:34out. But, so a lot
• 59:35of our work is like
• 59:36looking at audit logs, trying
• 59:37to get a sense on
• 59:38which patients or which providers
• 59:39need second opinions, and that's
• 59:40much more computationally intense.
• 59:43My guess is, like, in
• 59:44five years, Epic will just
• 59:45build this into Epic.
• 59:48Yes.
• 59:50So I mean, the question
• 59:51is, how do I state
• 59:53of,
• 59:54let's say, current out as
• 59:55well?
• 59:56Is it capable of creative
• 59:58information? And if not,
• 01:00:00is it accurate to study?
• 01:00:03You are asking the right
• 01:00:04questions. So okay. What I'm
• 01:00:05gonna say is controversial.
• 01:00:08LLMs,
• 01:00:10large language models, codify human
• 01:00:11knowledge.
• 01:00:12They are and there's actually
• 01:00:14there are computational tests to
• 01:00:15test their ability to be
• 01:00:16creative for things outside of
• 01:00:17their training set. I do
• 01:00:19not think there is any
• 01:00:20reason to think that any
• 01:00:21large language model will ever
• 01:00:23be able to be creative
• 01:00:25in that outside of its
• 01:00:26training set. They are effectively
• 01:00:29locking in human knowledge. Now
• 01:00:30they can be updated, and
• 01:00:31they can read things and
• 01:00:33integrate that new knowledge, but
• 01:00:34they're still fundamentally limited by
• 01:00:36what's in their training set.
• 01:00:37And that gets to, like,
• 01:00:38well, what are the impacts
• 01:00:39for medicine? The fact of
• 01:00:40the matter is when it
• 01:00:40comes to diagnosis,
• 01:00:42ninety nine, ninety eight times,
• 01:00:44we're not being creative, but
• 01:00:45sometimes that's necessary. And what
• 01:00:47does this do for human
• 01:00:48creativity? I mean, everyone's seen
• 01:00:49this. When you work with
• 01:00:50an LLM, you have it
• 01:00:51write something. It's very average.
• 01:00:53It's very milquetoast.
• 01:00:54It literally is picking out
• 01:00:56the average of its training
• 01:00:57set. That's actually one of
• 01:00:58the reasons it works well
• 01:00:59in diagnosis, but there's gonna
• 01:01:00be downstream effects and the
• 01:01:02lack of creativity is one
• 01:01:03of them.
• 01:01:07Well, you mean in science
• 01:01:08or in These are advanced.
• 01:01:10Any that is that we
• 01:01:11touch. I mean, this is
• 01:01:12this is a very real
• 01:01:13concern.
• 01:01:15Yeah. You're not wrong.
• 01:01:17I
• 01:01:18I think this is is
• 01:01:20that is that depressing? I'm
• 01:01:21sorry. You're maybe looking for
• 01:01:22a more optimistic answer there.
• 01:01:24LLMs
• 01:01:25will, I don't think, ever
• 01:01:26be capable of creativity in
• 01:01:28the way that a human
• 01:01:29is.
• 01:01:32Oh, I have all the
• 01:01:32time in the world.
• 01:01:34That's not true. But
• 01:01:36What's sometimes specifically the physical
• 01:01:39exam? You know, how can
• 01:01:40we be doing a diagnostic
• 01:01:42differential
• 01:01:43based on the exam?
• 01:01:45And and this is a
• 01:01:46bigger question. I got really
• 01:01:47the way,
• 01:01:49the best side of,
• 01:01:51schedules that we need,
• 01:01:53or
• 01:01:55errors. And there's also implications
• 01:01:57for what we should be
• 01:01:57teaching our students.
• 01:01:59A physical exam, in terms
• 01:02:01of collecting data is not
• 01:02:02something that an LLM can
• 01:02:03do now.
• 01:02:05Multimodal models this is why
• 01:02:06I'm telling you pathology everyone
• 01:02:08who confidently predicts that pathology
• 01:02:09is going to be computerized,
• 01:02:11that technology is way ways
• 01:02:12away. Multimodal models do not
• 01:02:14perform very well.
• 01:02:16And
• 01:02:17maybe five to ten years
• 01:02:18from multimodal models being able
• 01:02:20to perform at a human
• 01:02:20level. But so in the
• 01:02:22interim, an accurate physical exam
• 01:02:24becomes incredibly important. And you
• 01:02:26saw that exam that the
• 01:02:27student that the resident put
• 01:02:28in there. That's a templated
• 01:02:29exam.
• 01:02:31I mean, probably the resident
• 01:02:32did an appropriate exam for
• 01:02:34someone who they thought might
• 01:02:35have a dissection or PE,
• 01:02:36but she documented just the
• 01:02:38templated exam and that can
• 01:02:40throw off a language model.
• 01:02:41So when it gets to,
• 01:02:42like, what are the things
• 01:02:43that are uniquely good at
• 01:02:44humans in a world where
• 01:02:45we're working more and more
• 01:02:46with AI models,
• 01:02:47good observational skills and learning
• 01:02:50how to do those skills
• 01:02:51and then
• 01:02:52accurately
• 01:02:53represent them is really important.
• 01:02:55I think we'll call it
• 01:02:56a day. We'll get Adam
• 01:02:58a chance to do a
• 01:02:58Phew. Yeah.