Scientists Taught AI to Predict Heart Failure Five Years Away

The first sign something is wrong with your heart often isn't a symptom. It's a hospital admission.

Heart failure builds quietly with the heart straining, the tissue around it shifting in ways that don't show up until it's already too late to do much about it. By the time most people find out, the window for early intervention has closed. Even though, there was a scan somewhere had the information all along.

Oxford researchers say they've found a way to change that. Using a scan millions of people are already getting.

The Coronary CT angiography is standard cardiac care. Hospitals run it daily, read it, file it away.

What the Oxford team found is that those images contain more than anyone's been extracting. The fat around the heart, tissue most analysis ignores, changes in subtle ways when the heart is under stress. Too subtle for the human eye on a routine read. Not too subtle for software trained to look.

They tested the AI on 72,751 adults across nine NHS hospitals. Results published this week in the Journal of the American College of Cardiology. Patients at the top of the risk score were 20 times more likely to develop heart failure than those at the bottom. One in four of the highest-risk group developed it within five years. The model's accuracy held after controlling for everything doctors already track — age, existing coronary disease, cardiovascular risk factors. A lot of medical AI collapses at that step. This didn't.

What it doesn't do

Worth being clear here, because if you or someone you know has had a cardiac scan, the natural instinct is to read this as more than it is. This doesn't predict heart failure for the general population. It flags elevated risk among people already in the cardiac pipeline — already getting this scan.

Also read: Apple Smart Glasses Could Be a Late Rival to Meta Ray-Ban Glasses

Regulatory approval is being sought. It's not in clinics yet. The study also doesn't address performance across different ethnicities and demographics outside the NHS dataset, something that needs answering before this reaches widespread use.

It still matters

The scan already exists. The image is already stored. No new machines, no extra appointments, nothing the patient has to do differently. If this clears regulatory review and performs the way it did in the study, it just slots in.

Heart failure affects 7.7 million Americans in over a year and is frequently developing silently for years before it forces someone into hospital. Most of those people had cardiac scans somewhere along the way. Some of those scans may have had the answer sitting in them the whole time — in the fat around the heart, in the patterns nobody was trained to read.

That's what makes this worth watching. Not that AI beat the doctors. That the scan finally starts telling the whole story.