AI thought knee X-rays show if you drink beer — they don’t

But one study scientific report It highlights a hidden challenge in using artificial intelligence in medical imaging research – a phenomenon that is highly accurate but can produce misleading results, known as “shortcut learning.”

Researchers analyzed more than 25,000 knee X-rays from the National Institutes of Health-funded Osteoarthritis Project and found that artificial intelligence models can “predict” irrelevant and implausible characteristics, such as whether patients avoid eating fried beans. Mud or beer. Although these predictions have no medical basis, these models achieve surprising accuracy by exploiting subtle and unexpected patterns in the data.

“While artificial intelligence has the potential to transform medicine,” said study senior author Dr. Peter Schilling, an orthopedic surgeon at Dartmouth Health’s Dartmouth Hitchcock Medical Center and assistant professor of orthopedics at Dartmouth’s Geisel School of Medicine. images, but we must remain cautious.

“These models can see patterns that humans can’t, but not all patterns they identify are meaningful or reliable,” Schilling said. “Recognizing these risks is critical to preventing misleading conclusions and ensuring scientific integrity.”

The researchers looked at how artificial intelligence algorithms often rely on confounding variables, such as differences in X-ray equipment or clinical site markers, to make predictions rather than medically meaningful features. Attempts to eliminate these biases have had only modest success—AI models simply “learn” other hidden patterns in the data.

“This goes beyond bias in race or gender cues,” said study co-author Brandon Hill, a machine learning scientist at Dartmouth Hitchcock. “We found that algorithms can even learn to predict Years of taking X-rays. This is harmful – when you prevent it from learning one of the elements, it learns another element that it previously ignored. This danger can have some real consequences.

The findings underscore the need for rigorous evaluation standards for AI-based medical research. Overreliance on standard algorithms without deeper review may lead to erroneous clinical insights and treatment pathways.

“When it comes to using models to discover new patterns in medicine, the burden of proof increases significantly,” Hill said. “Part of the problem lies in our own biases. We can easily fall into the trap of assuming that models ‘see’ the same way we do. Ultimately, that’s not the case.”

“Artificial intelligence is almost like dealing with alien intelligence,” Hill continued. “You want to say that this model is ‘cheating,’ but that personifies the technology. It learns a way to solve the task given to it, but not necessarily how a human would solve it. It doesn’t have the way we typically understand it. Logic or reason about it.

Schilling, Hill and study co-author Frances Koback, a third-year medical student at Dartmouth’s Geisel School, conducted the study in collaboration with the Veterans Affairs Medical Center in White River Junction, Vermont.