New AI Reads Your Whole Heart, Not Just One Test

Doctors rarely diagnose a heart problem from one test. They piece together a rhythm tracing, a heart ultrasound, a chest X-ray, bloodwork, and what the patient tells them. AI usually doesn't work that way. Most medical AI models specialize in one type of test and only combine information at the very end, if at all. A team of researchers decided to fix that by building a system that learns from all of it at the same time.

The trick is a technique called a masked autoencoder, which you can think of as a fill-in-the-blank exercise. The AI is shown parts of a patient's record with chunks hidden, then asked to guess what's missing. To do that well, it has to build an internal picture of how heart rhythms, ultrasound images, and lab values relate to one another. Trained on more than 1.2 million hospital stays from a large U.S. hospital database, it learned those connections on its own, without anyone labeling the data by hand.

The payoff showed up in practical tasks. Compared with AI trained on single test types, this model was better at predicting in-hospital death, estimating how long a patient would stay, and assigning the correct ICD-10 and DRG codes — the codes hospitals use to bill insurers. Coding errors cost hospitals and patients real money, so that last one is more important than it sounds. Notably, the model stayed accurate even when only one test was available, which matters because real patients rarely have every scan on file.

The honest caveat: this is a preprint, meaning it has not yet been reviewed and confirmed by other scientists. It was trained on one hospital dataset, so it may not travel well to other hospitals with different equipment and patient populations. There is also the privacy question that hangs over any AI learning from medical records. Most importantly, nothing here is approved to diagnose anyone. It is a promising step toward AI that thinks more like a clinician, not a replacement for one.