@oziadias Good explainer for how this discovery was made @NatureNV https://www.nature.com/articles/d41586-026-01806-z
Congrats @oziadias and colleagues for this extraordinary work!
The model flags a slurred downstroke after the R-peak.
@oziadias Good explainer for how this discovery was made @NatureNV https://www.nature.com/articles/d41586-026-01806-z
Congrats @oziadias and colleagues for this extraordinary work!
Many users praise the AI model for discovering a new ECG biomarker for sudden cardiac death risk because it surfaces useful signals humans missed and could save lives with no new tests needed.
No Digg Deeper questions have been answered for this story yet.
This is very impressive!
A model was trained to predict probability of death from ECG, which can be used to flag patients with high risk of sudden cardiac death.
it was combined with a generative model for explainability, leading to the discovery of a new biomarker that cardiologists were not aware of before!
AI can help us *do* useful things in medicine – great!
Below - one such thing, from our @nature paper: - AI can flag people at high risk of dropping dead - and help us decide who gets implanted defibrillators
But what can we *learn* from AI?
My friend the brilliant @kevin_volpp dropped dead
He survived (read abt it: https://tinyurl.com/ypcpdzpz)
But every year 300-400,000 in the US don't
This week: our @nature article (+A.Schubert J.Ross @m_sendhil M.Lingman) applies AI to this medical mystery
https://www.nature.com/articles/s41586-026-10674-6

If you made it this far: Separate thread on “AI for science” part of the paper - coming up next
If you're interested in your own ECG prediction... we're working on doing that in a safe and responsible way
Leave us your email here https://eccentric.health/ecg-interest/

Sudden death is a great problem for AI
There's a direct line from predictions to saving lives.
If you knew you were going to drop dead from an arrhythmia, you'd want an implanted defibrillator
These devices are a miracle. They detect and terminate arrhythmias before they kill.

But… defibrillators have existed since 1980 – why are people still dying?
Because we’re bad at predicting who needs one.
Exhibit A: 3-400,000 preventable deaths per year
Exhibit B: 2/3 defibrillators never fire – because we implant them in people who turn out to be low risk.

What we do:
Pull all 440,000 ECGs in a Swedish region
Why ECGs? - They’re cheap and easy to do - They contain a lot of signal about your heart rhythm
Why Sweden? - Because we can link ECGs to high-quality death certificates (Swedes are *very* good at death certificates)

Are high-risk deaths preventable?
Lots of caveats (it's observational data; we’re working on a RCT!)
But interesting fact:
There’s one group the model does *terribly* in: those with defibrillators.
They die *way* less than they “should”
Suggests defibrillators could work

As some have observed, 100–7=93% - true!
But 7% is *annual*.
And 93%^5 (for example) = 70%
Which means a 5-year, potentially preventable mortality rate of 30%.
And 86% of high-risk people are unsuspected based on current state of the art (LVEF)

Important check: can we predict for non-Swedish people? Yes.
In the US: - high-risk patients: *25% annual rate* ventricular arrhythmias – proximal cause of sudden death – in EHRs.
In Taiwan: model distinguishes - future adjudicated arrhythmic arrests - vs. controls

Then we build a deep-learning model that looks at the ECG waveforms
It learns the difference between ECGs - from people who had sudden cardiac death - vs. people who didn’t.
It flags a high risk group (2% of our sample)
with a *7% risk* of dropping dead in the next year

@oziadias @kevin_volpp @Nature @m_sendhil The model doesn’t ask for new data or new tests. It extracts better signal from the exact ECGs cardiologists already order and read every day. That’s the difference between interesting research and something that can actually move into guidelines.

@oziadias @kevin_volpp @Nature @m_sendhil How much of this could have been analyzed prior to generative AI and ChatGPT? Seems like it could have been, using the data analysis tools available then?

@iScienceLuvr I'm scared of looking for this feature on my old ECG

@oziadias @kevin_volpp @Nature @m_sendhil thank you for this!!! countless lives will hopefully be saved

@oziadias @kevin_volpp @Nature @m_sendhil Congrats Ziad - super interesting!

@oziadias Dr. Volpp's story gives me goosebumps every time I read it, so many important lessons to be learned.

@oziadias @pmarca @kevin_volpp @Nature @m_sendhil Seems less a triumph of "deep learning" and more a signal of how cardiology ignores meaningful signals because it already knows everything there is to know.

@calimagna Great idea and good guess, it’s annoying to get the data

@oziadias @kevin_volpp @Nature @m_sendhil Without better discrimination between arrhythmic and non-arrhythmic deaths, difficult to make the jump from this model —> ICD. Presumably a lot of PEA deaths included.
Would love to see how model performs in patients w/ primary prevention ICD and subsequent therapies

@oziadias @kevin_volpp @Nature @m_sendhil wilddd that this works