Hacker News Reader

Comment by andy_ppp | Hacker News Reader

EKGs should be extremely easy for AI to identify every disease with a range of probabilities and even some humans can’t identify from EKGs. Do we have the labelled dataset for this?

View on HN

Eric Topol has a cool post on "machine eyes" -- using deep learning to identify subtle signals in ECGs and medical images that would be invisible to the naked eye: https://erictopol.substack.com/p/the-amazing-power-of-machin...

There are still regulatory/deployment/reimbursement barriers to cross, but this is happening.

There is a decent looking dataset available here, let's see my model can generate random numbers from this for each of the diseases: https://www.nature.com/articles/s41597-022-01403-5

You're assuming EKG is definitive.

I had an EKG last week, the analysis comes back "borderline". Running down every abnormality listed (I know my cardiologist has seen it, didn't consider it notable) has a range of possible causes, including the changes that come with good endurance.

"It should be very easy for an AI to look at an x-ray, CT, ultrasound or MRI image and tell what disease a human got, even some that humans don't know of.

Are there any labeled datasets "

-Some Software dude, every month since 1971

No the examples you give are extremely difficult compared to the 2D graphs of an EKG. The Apple Watch is clearly doing some fairly accurate inference with a single lead around arrhythmias, for example. I’m really sorry for being enthusiastic about machine learning, I just finished doing an intensive ML bootcamp and it was fun getting results. I also actually have some heart issues so I’d love to see if I could get a result. Thanks for your constructive comment though, I’ve never seen ones like this here before!

>I just finished doing an intensive ML bootcamp

Then maybe learn a bit more about the domain before posting silly things.

I'm not talking about me successfully building an AI that can do better than humans or identify all the worlds heart diseases I'm more looking to have fun and play around with some data (and yes learn more about the domain by doing!). I was maybe a bit too excited about seeing what would happen with a real dataset like this, but it's a hacker news comment not a PhD defence, no need to be so negative. Thanks.

I wonder if all symptoms would show on an EKG though. Would reduced coronary blood flow alter electrical signals ?

That’s what’s fascinating about AI it might find patterns with enough data that we don’t see.

That’s a stress test.

Attach an EKG, have the patient ride a bike with significant exertion.

If there is a lack of blood flow, it will show up in the EKG as alteration of the electrical signals through the heart.

Well I've been subjected to these tests and I fell between the lines, I was clearly below normal health (had trouble walking) but they said there was no issues. So I wondered if there's not more subtleties. Like lag between effort and signals showing up, or vascular issues like micro clotting impeding flow.

EKG changes are a late sign. You need structural imaging like CAC or CCTA or echo.

Reply on news.ycombinator.com