ChatGPT Has Passed the United States Medical Licensing Examination (USMLE) | Hacker News Reader

ChatGPT Has Passed the United States Medical Licensing Examination (USMLE)

medrxiv.org

45 pointszekrioca30 comments

The hn title is really pushing way beyond what the study can shoulder. The title from the result chapter sums it up better; "ChatGPT yields moderate accuracy approaching passing performance on USMLE"

The study was done on public available questions that have been used in usmle exams. They used two physician reviewers to judge ChatGPTs textual answers. While really interesting no exam was actually passed, and the authors never claim that.

Public questions really means they have likely been in training data together with their answers

From the paper:

Random spot checking was performed to ensure that none of the answers, explanations, or related content were indexed on Google prior to January 1, 2022, representing the last date accessible to the ChatGPT training dataset.

So its maybe not a problem, who knows

The HN title is misleading. The study is titled:

Performance of ChatGPT on USMLE: Potential for AI-Assisted Medical Education Using Large Language Models

The study did not attempt to get ChatGPT to pass any exam. Rather, ChatGPT was question on material in past exams. The authors are enthusiastic about ChatGPT's performance but nowhere do they claim it "passed the US Medical Licensing Exam" as the title claims.

Indeed, the study concludes that ChatGPT's performance is less than enough to pass the exam:

ChatGPT yields moderate accuracy approaching passing performance on USMLE (line 287 of the manuscript).

@zekrioca, please correct the title so that it more closely reflects the results of the linked study.

Sorry you and all about the misunderstanding, I’d gladly fix it but I guess it is late? I couldn’t get the real title to fit within the space limit, and improvised from the abstract.

Hey, no worries, thanks for posting.

Medical student here. I’m impressed by ChatGPT’s ability to come close to passing Step 2/3. I tried plugging in my own question banks that I used for studying and it got many right. Even when it was wrong it was also “almost right” and its answer would not be disparaged on rounds.

However one thing to keep in mind: a bot passing these exams alone will not disrupt medicine. There are plenty of doctors today that could ace these exams. I don’t think that the ability to outsource medical decision making to a bot will change much, especially since 90% of medical decision making is already outsourced to a flowchart or an “expert attending” physician. Much of the value of doctors today comes from procedures, patient interaction, legal liability, and patient trust, which cannot yet be done by a bot.

Just because it can pass the test doesn't mean that its behavior will be consistently correct. Trusting ChatGPT in medical applications is the same as trusting it in programming applications: It will work for simple to intermediate tasks and may even look as though its explaining itself properly, but for specialized tasks, sufficiently complicated ones, and even in simple tasks, it can't be expected to product correct results consistently.

>Just because it can pass the test doesn't mean that its behavior will be consistently correct.

Now do doctors.

The main observation I've had of these AI tools is they don't learn the same way people do. You see the art generation ones which are able to do hyper realistic texturing and shading, but fail to get the right number of fingers on a hand. In the say way ChatGPT is able to answer some extremely tricky questions right while failing trivial ones.

They don’t even learn, they confabulate. “Correct” responses are an accident, and there is no notion of (especially medical) correctness involved.

I already started using it for medical applications. My friend had allergic reaction to Monster Energy drink today, and while I could find a lot of documents in Google, ChatGPT explained what I need to know, what medicine she has to buy (antihistamine, it worked), when is it important to call ambulance, what to take.

I can Google all these things, but all the SEO spam pages that I get just take a lot of time and effort.

Also she doesn't have money for private doctor, and to get to talk to a doctor takes weeks.

She was also at the emergency last time when she had the reaction, and they made her wait 8 hours there :(

ChatGPT just works for medical problems (but I still double check the answers)

Isn't all that seo spam in ChatGPT's training data?

The looming AI-generated SEO nightmare feels like we are on the cusp of a cousin to the machine translation problem.

Many years ago, someone had the fantastic idea to improve translation by comparing websites available in different languages. This shift made machine translation noticeably less embarrassing. It also meant that instead of hiring a translator to translate a site fully, you could cheaply hire a shop that runs machine translation and fixes the worst errors.

The knock-on effect here is that many websites that would constitute input data for future machine translation systems are essentially already machine translated. So we are in a G-in/G-out situation.

(source: a translator friend of mine in the AAA games industry)

If you had to consult the AI for what to do for an allergic reaction, how do you check the answer?

What do you mean?

Just ask what are the reasons that a Monster drink gave an allergic reaction and what are the treatment options.

They are asking how you would validate the medical advice given is appropriate and not dangerous.

You said “ I still double check the answers”. How do you do it?

Ah, sure, I had to do all the Google searches for that anyways (look at the caffeine content, see how it changes histamine metabolism).

Also I asked my friend if she has itching to confirm it's the same thing ChatGPT is talking about (she said yes).

Asked her if she has problem with breathing (if she needs ambulance).

Basically I used Google and ChatGPT together as tools for helping me understand her situation and the seriousness of the situation better.

It was also easier to do directed queries to Google with ChatGPT's answer.

As a comparision Google gets these kind of ugly SEO pages (the word histamine is only once in it, and very hard to find):

https://reizeclub.com/can-energy-drinks-make-you-itchy/

60% is considered passing (and that's the range the paper says it achieves). So yeah, it's not consistently correct.

,, The most recent iteration of the GPT LLM (GPT3) achieved 46% accuracy with zero prompting12, which marginally improved to 50% with further model training. Previous models, merely months prior, performed at 36.7%13. In this present study, ChatGPT performed at >50% accuracy across all examinations, exceeding 60% in most analyses''

I have heared many times that ChatGPT is basically GPT3, but somehow it feels like it gives more correct answers. Now there's data to show it.

I'm not sure if this is more of a pro in the ChatGPT column or an indictment of the USMLE. Although I don't think the USMLE is known as an easy exam, so I guess ChatGPT really is that good now?

Isn't this yet another case of "you have to memorize and regurgitate a ton of information" which people are bad at but something like ChatGPT is exceptionally good at?

I think this is one point that I don't see being made often enough. ChatGPT is exceptional at retrieval of information, but if it starts passing tests like the MCAT and LSAT and so on, I feel this is more that our testing is just crap, rather than ChatGPT being "smarter" than humans.

Being a doctor does require a lot of just plain information retrieval. Human bodies have a kajillion parts interacting in a kajillion-squared ways. And we usually don't get to look directly at most of the parts; we have to infer what's going on from very indirect evidence.

This isn't one of those "memorize nothing you can look up" situations. You just have to have it at your disposal before you can do actual medicine.

There may be a day when a computer can do most of the memorization scutwork and you can get back to the real work, but we're not there yet. It's not the test's fault that it's testing medicine as it's practiced today.

> In 2020, 98% of DO and MD students passed the exam on their first try. But if you don’t, you are given four attempts to pass each different level of the USMLE exam. If you fail, you receive a score report that offers feedback on the areas in which you struggled.

from https://www.ama-assn.org/medical-students/usmle-step-1-2/usm...

It would seem USMLE is not hard to pass.

Does anyone know the methodology? E.g., what were the inputs and context windows given to ChatGPT?

The full text is available and describes how they approached it. I didn't look to see if the actual inputs are in there.

It's available on an "online data system" says the article. But no where do they say what the damn system is.

Heart palpitations

Reply on news.ycombinator.com