LLMs approach expert-level clinical knowledge and reasoning in ophthalmology

https://www.ft.com/content/5b7a76be-467c-4074-8fd0-3e297bcd91d7

by thm • 2 years ago

78 102 2 years ago

ft.com

dedosk• 2 years ago

behind paywall

mtmail• 2 years ago

But please don't post complaints about paywalls. https://news.ycombinator.com/newsfaq.html

danesparza• 2 years ago

How do I read the article?

mewpmewp23• 2 years ago

So it doesn't match?

lionkor• 2 years ago

If smart people made language,yea, youd be right. But most people speaking english dont have the capacity, so no, "all but" means the "completely", for no logical reason.

SiempreViernes2• 2 years ago

No, in this case "all but" means "as good as, or worse, than the bottom 2/5 of doctors tested on a written exam", see table 1.

ethagknight• 2 years ago

I know you are being funny, but I would just say this title is an incorrect use of the idiom. I’m assuming the intent is to have the words “model matches doctors” as clickbait. “All but” idiom means “very nearly”, not completely.

mdrzn• 2 years ago

https://archive.is/w3Sk0

monkeydust• 2 years ago

Link to paper:

https://journals.plos.org/digitalhealth/article?id=10.1371/j...

gwern2• 2 years ago

FRCOphth Part 2 questions were sourced from a textbook for doctors preparing to take the examination [17]. This textbook is not freely available on the internet, making the possibility of its content being included in LLMs’ training datasets unlikely [1].

I can't believe they're serious. They didn't even write any new questions?

hyperorca• 2 years ago

Funny thing is that this textbook can be easily found on LibGen. I do not know a lot about LLM datasets, but they probably include books from these shadow libraries, right ?

dagmx• 2 years ago

I’m sure they’ll be surprised how many books aren’t freely available on the internet that are in common training sets.

It would have taken them a few minutes to learn about the current lawsuits around book piracy.

I agree with you. That they didn’t try asking novel questions or even look into what training data is used makes this paper bunk.

westurner• 2 years ago

"Large language models approach expert-level clinical knowledge and reasoning in ophthalmology: A head-to-head cross-sectional study" (2024) https://journals.plos.org/digitalhealth/article?id=10.1371/j...

[...] We trialled GPT-3.5 and GPT-4 on 347 ophthalmology questions before GPT-3.5, GPT-4, PaLM 2, LLaMA, expert ophthalmologists, and doctors in training were trialled on a mock examination of 87 questions. Performance was analysed with respect to question subject and type (first order recall and higher order reasoning). Masked ophthalmologists graded the accuracy, relevance, and overall preference of GPT-3.5 and GPT-4 responses to the same questions. The performance of GPT-4 (69%) was superior to GPT-3.5 (48%), LLaMA (32%), and PaLM 2 (56%). GPT-4 compared favourably with expert ophthalmologists (median 76%, range 64–90%), ophthalmology trainees (median 59%, range 57–63%), and unspecialised junior doctors (median 43%, range 41–44%). Low agreement between LLMs and doctors reflected idiosyncratic differences in knowledge and reasoning with overall consistency across subjects and types (p>0.05). All ophthalmologists preferred GPT-4 responses over GPT-3.5 and rated the accuracy and relevance of GPT-4 as higher (p<0.05). LLMs are approaching expert-level knowledge and reasoning skills in ophthalmology. In view of the comparable or superior performance to trainee-grade ophthalmologists and unspecialised junior doctors, state-of-the-art LLMs such as GPT-4 may provide useful medical advice and assistance where access to expert ophthalmologists is limited. Clinical benchmarks provide useful assays of LLM capabilities in healthcare before clinical trials can be designed and conducted.

RandomLensman5• 2 years ago

On questions, not actually examining eyes, though.

SiempreViernes2• 2 years ago

It also scored 0% on ethics :D

(Was just one question though, 8/10 humans got that one right)

JaimeThompson• 2 years ago

Maybe it enabled modern MBA mode for that question ;)

bidandanswer2• 2 years ago

On that note, legacy computer vision models have exceeded humans at identifying pathologies from images of retinas, etc., for probably decades now.

greybox• 2 years ago

So. . . It doesn't work

hestefisk• 2 years ago

“What this work shows is that the knowledge and reasoning ability of these large language models in an eye health context is now almost indistinguishable from experts,” said Arun Thirunavukarasu, the lead author of a paper on the findings published in PLOS Digital Health journal.”

FTFA.

boyka• 2 years ago

FTFA?

alex_suzuki• 2 years ago

From The F**ing Article, presumably.

nicklecompte• 2 years ago

What this work actually shows is that a bunch of scientists are ignorant about how LLMs work, but are rushing to publish papers about them anyway. It is ridiculous for Thirunavukarasu to draw this conclusion from GPT-4's performance on a written ophthalmology exam.

From the good folks at AI Snake Oil[1]

Memorization is a spectrum. Even if a language model hasn’t seen an exact problem on a training set, it has inevitably seen examples that are pretty close, simply because of the size of the training corpus. That means it can get away with a much shallower level of reasoning....In some real-world tasks, shallow reasoning may be sufficient, but not always. The world is constantly changing, so if a bot is asked to analyze the legal consequences of a new technology or a new judicial decision, it doesn’t have much to draw upon. In short, as Emily Bender points out, tests designed for humans lack construct validity when applied to bots.
On top of this, professional exams, especially the bar exam, notoriously overemphasize subject-matter knowledge and underemphasize real-world skills, which are far harder to measure in a standardized, computer-administered way. In other words, not only do these exams emphasize the wrong thing, they overemphasize precisely the thing that language models are good at.

Also[2]:

Undoubtedly, AI and LLMs will transform every facet of what we do, from research and writing to graphic design and medical diagnosis. However, its current success in passing standardized test after standardized test is an indictment of what and how we train our doctors, our lawyers, and our students in general. ChatGPT passed an examination that rewards memorizing the components of a system rather than analyzing how it works, how it fails, how it was created, how it is maintained. Its success demonstrates some of the shortcomings in how we train and evaluate medical students. Critical thinking requires appreciation that ground truths in medicine continually shift, and more importantly, an understanding how and why they shift. Perhaps the most important lesson from the success of LLMs in passing examinations such as the USMLE is that now is the time to rethink how we train and evaluate our students.

[1] https://www.aisnakeoil.com/p/gpt-4-and-professional-benchmar...

[2] https://journals.plos.org/digitalhealth/article?id=10.1371/j...

Filligree• 2 years ago

That might be the thing that makes me most optimistic about AI.

Not because they’re super useful. They are, if and only if you use them right, which is a skill few people seem to have.

But because they’re illuminating flaws in how we’re train our students, and act as a forcing function to _make_ the universities and schools fix that. There’s no longer any choice!

thiago_fm• 2 years ago

Lots of people publish papers to get cited and be first in the field, without any rigorous research behind it, to try to net them a good career.

Mostly recently the best to do is to ignore this, and soon enough the things that are true will emerge, without academia.

The media is the worst, that tends to look for whatever research that would be click-baity. It's a real waste of time to follow FT and others.

aantix9• 2 years ago

This doesn’t surprise me.

I find most doctors are terrible at debugging root cause.

They’re simply working through a protocol of most common symptoms/cause and if your ailment doesn’t fit these narrow requirements, you’re screwed. Your 10 minute consult has ended and they’re not thinking about the problem anymore.

mewpmewp23• 2 years ago

This is exactly what a doctor told me during my health check up as well.

He casually hit up a monologue on how dumb doctors are. I didn't really ask, but I wonder if he does this monologue to every patient he has.

It seemed like for years he had developed this theatrical and interactive comedy piece for all the patients he was doing the check up on.

RandomLensman5• 2 years ago

Common things are common, so going through common things first doesn't strike me as wrong. If someone suffers from something really exotic then I'd argue that is what specialists are for.

The 10 minute window is more an external/financial constraint, I'd say.

wredue• 2 years ago

One of the main issues doctors face today is that everyone is really fat, and being fat leads to every diagnosis under the sun.

People don’t like being told that being fat is causing their problem (when it almost definitely is), and this is creating a lot of distrust in doctors.

I mean, you see these examples of it being more than just overweight, but that’s the minority, and everyone being fat unfortunately contributes to increasing the rate of these misdiagnosis.

aantix9• 2 years ago

I saw multiple cardiologists in Lincoln, NE, and San Francisco, CA.

Multiple nephrologists throughout the 25 years.

These are the "specialists" I was referred to from my general physician.

I had to work side-by-side with them to manage my condition.

I had lost significant weight in the past, but it never reduced my blood pressure medication intake.

But there was nothing special about these doctors.

They simply had no answers beyond the typical.

The equivalent for software would be that a bug report comes in and the engineer responds with "have you tried rebooting the servers?" Just superficial, unhelpful recommendations.

magicalhippo• 2 years ago

That's how I approach analyzing a software issue as well. Start with the basics, is the application even running, is it actually reading the file you think it's reading etc.

So often it's not something weird.

resource_waste2• 2 years ago

The most horrifying thing has been being an adult and seeing how terrible medical doctors are at diagnosis (of my kids).

If there are no objective measurements, do your own research and get second opinions. Maybe this works because I come from a science background.

Heck today, I'm getting closer and closer to trusting ChatGPT4 over a doctor. My wife has used it to help diagnosis patients and 2 of them had life changing success. (okay hold up before my keyboard warrior pals think we are using chatgpt4 like a magic 8 ball, closing our eyes and showing the patient the response. No it dosent work like that)

aantix9• 2 years ago

I had blood pressure issues of 25 years.

“Lose more weight”. “Less sodium.” “More medications. “ (four to control my BP at its peak)

Four cardiologists. Three nephrologists. All have WebMD level recommendations at best.

I do my own research. I see hyperaldosteronism as a possibility. Cardiologist didn’t think so. But reluctantly gives an endocrinologist referral.

Yes, I have it. Had my left (overactive) adrenal gland removed because of it.

Now, BP is normalized with significantly less medication.

Your health is in your hands. Do your own research.

No one else will save you, including the doctors.

aantix9• 2 years ago

I wonder how we get better physician ratings?

Right now, all online reviews are superficial at best.

"He listened to my questions."

"She made my family feel at ease."

"The front office followed up for my next appointment."

I don't care how kind or rude the physician is.

I want them to get my diagnosis correct.

I'd like to see data on each physician - each patient's five- and ten-year follow-up data. Reveal to me if the diagnosis was correct.

I understand the messiness of this data gathering. I know some patients may not want to reveal such sensitive data. Doctors would probably game the scoring system. There may be confounding conditions as well.

But this would be my dream data scenario.

emporas• 2 years ago

They are also getting paid by the drugs they prescribe, not by the diseases they cure. Incentives guide actions of many people. Not saying their actions are dictated only by incentives and nothing else, but still, a strong influence there.

svaha1728• 2 years ago

Ask the model to provide an image of its answers written on a chalkboard. Or, ask it to provide you an image of pathological eye conditions and look at the text labels.

sethammons• 2 years ago

My wife had gone through half a dozen doctors or more over a decade, giving them the same description of her issues. They never correctly diagnosed the issue. We typed her descriptions into chatgpt and it suggested multiple sclerosis. It was right. It must be what other folks feel when they talk to tech support, but I've never, not once, been impressed by a doctor and have found them lacking in nearly every interaction.

xanderlewis2• 2 years ago

The best doctors are the ones who listen carefully to the patient before dispensing their expert opinion. Very few do.

Perhaps that’s just because they’re jaded from having to deal with the public all day long…

joelfried2• 2 years ago

I agree with you - listening is very important.

I worked in an ophthalmology practice when I was quite young many years ago. The head of the practice often had over 50 patients a day (one of my responsibilities was pulling the folders for all daily patients from the wall-o'-folders; this was in the age everything was on paper).

50 patients over 8 hours gives you less than 10 minutes per patient. I suspect a nontrivial part of the problem is being rushed.

swader999• 2 years ago

Figure out how they are paid and that will tell you much. In Alberta they get paid by the visit and many are using ten minute appointments as the norm. Often takes six weeks for me to get an appointment with my family doc.

Zenzero4• 2 years ago

HN is going to select for a very small subset of the population, so unfortunately people's experiences have to be interpreted in that light.

A massive chunk of my caseload is the direct result of people not listening to basic instructions. Between people not listening and people who don't understand shockingly basic fundamentals of health, you become jaded.

Add to that the active disincentives to doctors really digging into cases, and we end up where we are. It feels much of my knowledge goes unused and wasted.

As an aside, I have messed around with chat gpt 4 for how well it can mimic what I know. My personal experience is that I can see how it would impress non-medically trained people, but to my eyes it has a long way to go. It struggles to accurately rank differentials. Also once you press deeper into decision making it gets really scary. It was consistently presenting treatment plans that would have killed patients if followed. I would love more than anything to have AI cut down some of our work but I was really disappointed with how it currently performs. Hopefully improvements keep coming.

rjzzleep2• 2 years ago

The way med school works by definition is that it filters for very conformant hard working individuals and then fills them up with the most common symptoms, diseases and treatments and turns them into human dictionaries. There are of course many exceptions to that rule, but if you look at medical textbooks that is the structure most of them follow. A human competing on who's the best dictionary is a terrible way to solve this problem although it might have historically been helpful.

I recently had a retinal detachment, not entirely sure why, no doctor can tell me anything useful besides you must have been stressed, but I got it lasered despite my better judgement, because I knew fixating something in the eye somewhat arbitrarily must cause problems, so it left me with blurriness and aniseikonia(a condition with one eye seeing things smaller).

I don't need an LLM for that, just the right google queries do the trick. This is by far not the only issue I have with doctors, but my experience with doctors has been vastly different between exceptional doctors and people with absolutely no interest into any kind of in depth root cause analysis, but the vast majority of them has been in the latter category.

bidandanswer2• 2 years ago

I don't need an LLM for that, just the right google queries do the trick.

LLMs are clearly superior at presenting the information and have tangible room for improvement, whereas Google has regressed over the last decade and is getting more brittle by the day.

Open technology (LLMs) are going to offer much more robust and reproducible solutions to such questions as "here are my symptoms, what's wrong with me?" than Google.

You can nitpick my statements all you want and I will probably agree with you, but the overarching takeaway should be that LLMs are a much better solution, especially for laypeople who cannot use Google effectively.

staticman22• 2 years ago

In my experience (just having used it the other day to look at a blood test) ChatGPT is quick to tell a healthy person they are sick.

henriquenunez• 2 years ago

I do not trust doctors in general (unless they are scientists); and I also think that clinical knowledge is not that hard in ophthalmology…

troq13• 2 years ago

Well it is not going anywhere near my eyes.

rafaelero5• 2 years ago

I am very bullish for GPT-5. It's going to surpass even the best diagnosticians we have right now. And we will all have access for $20 a month. Crazy and exciting times.

squeegmeister• 2 years ago

There are reasons to believe we’ve reached a plateau with gpt 4. Check out Gary Marcus’s Substack

rafaelero5• 2 years ago

Unless he has access to GPT-5 his opinion is worthless.

daveguy3• 2 years ago

Do you think hallucinations will be solved with GPT-5? If so, that would be an amazing breakthrough. If not, it still won't be suitable for medical advice.

rafaelero5• 2 years ago

It will certainly decrease. Also, there are multiple ways to deal with hallucinations. You can sample GPT-4 not once, but 10, 100, 1000 times. The chances of it hallucinating the same things asymptotically reaches 0. It all depends on how much money you are willing to invest in getting the right opinion, which in the field of medicine, can be quite a lot.

caeril6• 2 years ago

This is a solid point.

When humans have vague or incomplete knowledge on a specific topic, the very LAST thing they do is fill in the gaps with their best guesses. Humans are built with no ego, and when they don't have a high degree of confidence in their answer, they always admit to it, and they NEVER fill in knowledge gaps with low-confidence information. Hallucination is a problem that ONLY exists with these evil machine models, not the flawless Vulcan Logic Devices inside our heads.

What these LLM fetishists don't understand is that this hallucination problem is a Very Big Deal, and we should always rely on the infallible human brain, which has never made assumptions, lied, or, made shit up.

Whenever you're looking at the third leading cause of death in the United States (preventable medical errors), seeing that a quarter million Americans are murdered annually, with a significant proportion of those errors being gross misdiagnosis, it's Super Important to remember that an LLM that hallucinates will NEVER be suitable for medical advice, because the human brain is a Magical Jesus Box that has never made similar mistakes.

Sometimes I like to watch World War 2 documentaries, and I tear up at the realization that nearly half a million American soldiers lost their lives in four years to those evil Japanese and Nazi scum. War is Hell.

But what is DEFINITELY NOT HELL, is exactly the same number of American fatalities at the hands of human doctors with their perfect, infallible, flawless-in-every-way human brains in only half the time, which is what we have now.

Imagine how bad it would be if we allowed these LLMs to hallucinate medical advice! People might die! It would be horrible!

Euphorbium• 2 years ago

Doctors cannot be automated soon enough, and nothing of value will be lost.

VeejayRampay• 2 years ago

doctors are by and large terrible

trained to condescend and ignore / infantilize patients, generally thinking highly of themselves, never providing explanations for anything, thinking they're better than everyone, I can't wait for them to get humbled

LLMs approach expert-level clinical knowledge and reasoning in ophthalmology

Related Stories