Heart · Diagnosis and imaging
An AI that reads a routine heart tracing to spot hidden heart disease held up in a new hospital
Researchers tested a free, public AI that reads an ECG (the quick sticky-pad heart tracing) to flag structural heart disease. On 45,878 patients from a separate hospital database, all already sent for a heart ultrasound, it scored 0.790 (where 1.0 is perfect), close to its 0.820 at home. It's a study of past records, not tested in everyday screening. Rating: Early research.
How far away is this?
A peer-reviewed study (European Heart Journal: Digital Health, September 2026) using past hospital records. Nobody's care was changed by the AI, so it's Early research.
Get the weekly How Far
One email every Friday morning (Sydney time) on which medical AI actually reached patients, and what it means for you. Free, in plain English.
Where is it approved?
| Region | Regulator | Status | Notes |
|---|---|---|---|
| United States | FDA | Not approved | Research model; the study describes no FDA clearance. |
| European Union | CE mark / EMA | Unknown | Register not searched yet. |
| United Kingdom | MHRA | Unknown | Register not searched yet. |
| Australia | TGA | Unknown | Register not searched yet. |
A research model, not an approved product, as far as the study says. "Register not searched yet" means we haven't searched that regulator's register; it doesn't mean the product is or isn't available. Approval in one country isn't approval in another.
What is it?
Structural heart disease means problems with the heart's valves, walls or pumping. It's normally found with an ultrasound scan (echo), which is pricier and harder to get than an ECG. EchoNext-Mini is a released AI that looks at an ECG, plus age and sex, and estimates whether an echo would show disease. This team, at Thomas Jefferson University, tried it on a separate set of hospital records (the MIMIC-IV database) without retraining it.
What's the evidence?
The team used 45,878 patients who had an ECG and then an echo within a year. 47.2% of them, nearly half, had structural heart disease on the echo.[1]
Evidence stage: Early research · #c1
The AI's score for any structural heart disease was 0.790 (AUROC, where 0.5 is a coin toss and 1.0 is perfect), against 0.820 on its own test data. For single conditions it ranged from 0.679 (thick heart walls) to 0.837 (a weak right side of the heart).[1]
Evidence stage: Early research · #c2
The overall yes/no score was close to well calibrated, but the AI's percentages for single conditions overestimated risk in both hospitals. A simple fix based on the released training data mostly corrected this, and the authors say those numbers need adjusting before being read as real risks.[1]
Evidence stage: Early research · #c3
These were patients already sent for an echo, so it doesn't show how the AI would do screening everyday people. The authors declare no conflicts of interest.[1]
Evidence stage: Early research · #c4
What does it mean for me?
One day a normal ECG at your GP might come with an AI flag saying you should get an echo. It isn't there yet. If you've got symptoms like breathlessness, swollen ankles or a heart murmur, ask your doctor about an echo rather than waiting for an AI.
What could I ask my doctor?
- Would an echo be useful for me, given my symptoms?
These are general starting points, not a recommendation to choose any particular test or treatment.
Short videos, under a minute
We make short, plain-English videos about AI in medicine, including what it can't do yet. Follow How Far for the next one.
Change history
- : First published.
General information only, not medical advice. Talk to your doctor about your situation. AI-assisted source check. Not clinically reviewed.
References
- European Heart Journal: Digital Health (2026): Transportability of an AI-ECG model for structural heart disease (EchoNext-Mini in MIMIC-IV). doi.org (2026). Peer-reviewed study. https://doi.org/10.1093/ehjdh/ztag147
- Full text on PubMed Central (PMC13617667). pmc.ncbi.nlm.nih.gov. Peer-reviewed study. https://pmc.ncbi.nlm.nih.gov/articles/PMC13617667/
Publisher shown as the website hosting each source. Dates are given where the source label includes them.