New AI Predicts Prostate Cancer Return—But a Catch Emerges

After prostate cancer surgery, men anxiously watch their PSA levels. A rising PSA can mean the cancer is coming back. Doctors call this "biochemical recurrence," and predicting it early helps plan treatment. But current predictions rely heavily on human judgment and are not standardized. A new international challenge—called CHIMERA—set out to build a fair comparison of AI tools that could predict this recurrence.

Researchers gathered data from 267 patients across two hospitals, including 801 MRI scans, 942 microscope images of prostate tissue, and 13 clinical variables like age and tumor stage. They then trained different AI models to predict how quickly cancer might return. The best performer was a simple model using only clinical data, achieving a score of 0.74 on the C-index—a standard accuracy measure where 0.5 is a coin flip and 1.0 is perfect.

Here's the catch: when the researchers scrambled or withheld doctor-provided variables, that top model's performance collapsed to near-random (0.50). In comparison, models that also used MRI and tissue images lost almost nothing—dropping by at most 0.04. That means image-based AI is far more robust in the real world, where electronic medical records are incomplete or contain errors.

CHIMERA is the first public, standardized benchmark of its kind for prostate cancer prognosis. It doesn't claim that imaging alone is better, but it shows that relying only on clinical notes is brittle. For patients, this means future AI tools could give more reliable predictions even when doctors' records are patchy—offering a safety net when human data falls short.