New AI Test Exposes Gaps in Reading 24-Hour Heart Monitors

Your heart doesn't take breaks, so doctors sometimes have to listen to it for an entire day. That's what a Holter monitor does: a small device worn for 24 hours, recording every heartbeat while you go about your life. The problem? That's a lot of data to review. Researchers have been trying to teach AI to read these records, but until now, there was no good way to test if the AI was actually learning.

That's where this new study comes in. A team built Holtercare-23K, a huge dataset with 22,980 question-answer pairs pulled from 788 real Holter recordings. The questions test whether AI can do three things: find the exact moment a heartbeat goes wrong, say what the problem is, and summarize the overall pattern. Then they used this dataset to create Holtercare-Bench, a standardized test that anyone can use to grade AI models.

The results are sobering. Even the most advanced AI models made lots of mistakes when faced with these long, continuous heart recordings. Finding the precise time of an abnormality was especially hard — like trying to locate one misplayed note in a three-hour concert. But here's the hopeful part: when the researchers gave the AI extra training on heart data, its scores jumped dramatically. The models weren't broken; they just needed the right practice.

Why should you care? Half the people who have heart rhythm problems only show symptoms once in a while. An AI that can reliably scan a full day of heart data could help doctors spot dangerous patterns earlier and save more lives. The benchmark won't end up in your hospital tomorrow, but it sets a clear scoreboard for progress — so we'll know exactly when these systems are good enough to trust.