New AI Fixes Wandering Heart Scans, Catching Problems Earlier

Every year, millions of people get an echocardiogram — a heart ultrasound. Doctors use it to measure how well the heart muscle squeezes and relaxes. One key number is 'myocardial strain,' which sounds technical but simply means how much the heart muscle stretches and springs back with each beat. A weak squeeze can be an early sign of heart disease, sometimes showing up before other symptoms.

Recently, hospitals have started using AI to measure that squeeze automatically from the video. The problem: the AI tracks points on the heart wall frame by frame, and over a full heartbeat those points slowly drift off course. It's like following a runner on a track and ending up half a lap behind. By the end of the heartbeat, the software thinks the heart is in a different spot than it really is, which throws off the strain numbers — and in some cases makes the reading meaningless.

The team behind this paper, from universities and hospitals in France and Norway, extended an existing tracking system (called TAS-Net) with what they describe as 'persistent memory tokens.' In plain terms: instead of judging each slice of video on its own, the AI remembers what it saw earlier in the heartbeat and uses that to stay on track. They then fine-tuned it on real patient ultrasounds so the motion makes physical sense — heart tissue that leaves a spot should come back to it.

In tests, the drift dropped, the readings matched clinical reference measurements more closely, and repeating the same scan gave more consistent results. That last point matters most in practice: if two scans of the same patient give different numbers, doctors can't tell whether the heart changed or the software just wobbled. This is research, not yet a shipping product, so don't expect your next echo to use it tomorrow. But it points to better heart monitoring without buying new machines.