Agreed, and honestly this is one of the easiest and best behaving use case for LLMs. We’re not talking replacing any diagnoses, but freeing the humans to do what they do best, using their intuition.

AIs performed better at replacing human diagnoses long ago. Physicians don't like to use them because, well 8-10 years of specialized study tends to make you feel like you know what you're talking about vs some computer

LLMs aren't intelligent. They won't replace an EMR. They'll hallucinate information, associate the wrong details and just generally be not deterministic in the way you want an EMR to be. But they can be a powerful interface that does reduce documentation and time spent hunting down information in an EMR.

The future of EMRs is more machine learning models of various kinds(not just LLMs) all built into the EMR to turn it into a medical scribe with great memory and a keen eye for what's important.

Nurses and doctors aren't going anywhere until we achieve true artificial intelligence, but until then the goal is to keep their eyes off the computer and on the patient as much as possible.

> Multimodal LLMs will also replace much of medicine and nursing which is good.

You should read up on the history of AI in medicine. 30-40 years ago they had rudimentary (by our current era understanding) systems that basically equalled and beat physicians in diagnosing and prescribing medicine to patients when given a list of symptoms (sources at the end).

None of these ever had any uptake because physicians didn't want to use them, even when shown they performed more accurately.

Ultimately credentialed professionals like medicine and law will never be replaced by LLMs because LLMs cannot be held responsible for the medical regimes they place patients under (or the legal advice they give).

Sources:

https://en.m.wikipedia.org/wiki/DXplain

https://www.dhinsights.org/news/does-ai-matter-if-human-clin...

In most cases the diagnosis is the easy part. The hard part is gathering the signs and symptoms necessary to make the diagnosis. AI isn't particularly helpful for that.

Yep - and the hardest part about getting the input signs and symptoms data is that patients often don’t know the words to describe what they’re feeling.

LLMs are not going to help if the input data is also garbage, let alone hallucinations.

“Are you dizzy?” - sometimes, I’m not sure… etc

Never is a very strong word. In particular considering what changed in just 100 years in medicine.

It is hard to imagine that any kind of textual computer input will persist more than 20 years in the future.

Always bet on text (http://graydon2.dreamwidth.org/193447.html and on HN at https://news.ycombinator.com/item?id=8451271)

But really. Always, always bet on text. Chances are it'll work in 10000 years if we still exist.

> Ultimately credentialed professionals like medicine and law will never be replaced by LLMs because LLMs cannot be held responsible for the medical regimes they place patients under (or the legal advice they give).

It’s not quite the same as what you are saying, but things are changing quickly.

Software is diagnosing pathology in radiology now. Software made up at least 3/4 of the pixels in the MRI scans I acquire. Sometimes it makes up more than that as every second slice is made up too.

Clinic letters are being written by applications that listen to the consult with the patient and write up a summary (for the clinician to authorise).

Why would things be changing quickly now when old-school AI have been out-diagnosing doctors for years?

Again, like I said, doctors are ultimately responsible for the final call made. So no amount of AI will replace any doctors. It may remove some toil, but they cannot be replaced.

Was the previous test Done with massive amounts of structured input which would be incredibly time-consuming for the doctor involved? I think the difference now is that we’re talking About making diagnosis with minimal extra work or time from the clinicians

> Why would things be changing quickly now when old-school AI have been out-diagnosing doctors for years?

The tools actually save time and effort now, and don’t have a steep learning curve.

At least that’s my view from radiology as a tech.

"Multimodal LLMs will also replace much of medicine and nursing which is good"

Mind elaborate how is this done?

For starters, ward nurses spend 25-35% of their time doing documentation. This is doing patient notes, transcribing vitals, filling in forms and checklists, logging in and out of workstations, etc etc.

Much of this could be replaced by audio/video input.

Secondly, if the LLM can identify at risk patients or drug checks, you don’t need a nurse or 2 nurses to dispense medications or identify at risk patients. Multimodal input LLM plus a caring low skilled person can do the necessary things far more often.

These 2 factors alone can reduce nursing workload by half.

As for family doctors, I would personally be very happy to transfer my basic general care to a good LLM right now. Everything I need is straightforward, protocolised and the hassle of making appointments, delays, waiting rooms, form filling etc is more of a barrier to my healthcare than current LLM deficiencies.

Sounds like you don't have a very good understanding of what nurses actually do nor do you appreciate them.

Ok, I’ve spent the last 18 years working in hospitals, and the stats in my post are accepted in nursing research. Very open to hear where I’m wrong.

tech background here. Feel free to drop me a line if you need a teck parter to revamp what you described above

Tangential, will we ever certify current LLMs as "medical grade", for example they reveal suicide steps too easily, pick up Redditor snark as serious content, contatenate parts of random sentences together etc?