Comment by lucfranken | Hacker News Reader

Comment by lucfranken | Hacker News Reader

Parent Full thread

Why wouldn’t you as a doctor by standard run the images through a certified compliant LLM? The actual cost won’t be it and then you can see if you get any new ideas from it. See if it’s just wrong or that it spotted a little detail you missed?

The LLM doesn’t need to be leading or whatever but then you can have a conversation with the patient. If their ChatGPT reports has differences it can be analyzed as well.

It feels like the time constraint of the 15m doctor sessions is the thing. But if prepared immediately after the scan then why not?

There is always time needed to factor in new developments and innovations and that’s fine. Just moving blindly work from human to LLM is wrong. But learning on and testing with all the ai tools incoming constantly won’t be a waste. There will be more and more tools in those processes outside of human judgement, better improve the workflows now to be able to test and plugin new models and systems when they are ready.

> standard run the images through a certified compliant LLM?

Because they don't exist, yet.

In the UK MRIs and other imaging systems need two opinions. there has been a move to allow the first opinion to be ML based.

The _problem_ is that you are basically doing grey smudge analysis, and thats fucking hard.

I've been starting to think of LLM as a great tool for "lead generation," borrowing a term from sales. Most of the things it comes up with don't pan out, but in many cases it's things we wouldn't have thought of, or at least not as quickly. This is especially in the context of web service or SAAS outages.

Because they might bias you. And because you have your own brain, training and experience

That does make sense but the order of it might be: doctor analysis. Then show LLM as double check to doctor. Doctor assesses and may keep en improve the doctors first analysis?

Another problem is that general models' performance just sucks. From an upcoming conf. talk (in pathology) where they ran 2 Medgemma models on 100 slides with known diagnosis:

> Results: Full concordance with the reference diagnosis was 8% (27B) and 5% (1.5 4B; McNemar p=0.68), while partial matches were 29% vs 20% respectively (McNemar p=0.053). When correct diagnoses anywhere in the differential were counted, 51% (27B) vs 30% (1.5 4B), with 27B significantly superior (McNemar χ²=12.1, p=0.0005). Site-level performance varied widely (30–100%). Both models reported HIGH confidence in ~99% of cases irrespective of correctness.

i.e. highly confident, wrong 95% of time. in 49% of cases the real diagnosis wasn't even on models' differential. Doctor can hardly improve using something they can safely assume to be just noise.

https://ecp2026.abstractserver.com/programme/#/scientific/de...

Totally agree on that, if you have to look constantly at something 95% there is no value. The expectation of course is that it will be better. But if not at a certain level useless.

Not sure how that research compares to the claims being made by many that a second opinion via ai in the end led to changes in treatment. Likely people spent quite some time searching and figuring out. That would be a different and n=1 result. Don't have enough knowledge of that research to determine how much result can be gained when the models are managed in a way that produces better results.

And of course how much time/effort/cost that would take. How much is custom and how much is an automated programmable flow.

Reply on news.ycombinator.com