AI Misreads X-Rays, Thinks They Reveal Beer Drinking

A recent study by Dartmouth Health researchers has uncovered a significant challenge in applying artificial intelligence (AI) to medical imaging: AI models can produce highly accurate yet misleading results by exploiting unintended patterns in data—a phenomenon known as “shortcut learning.”

The researchers analyzed over 25,000 knee X-rays from the National Institutes of Health-funded Osteoarthritis Initiative. They found that AI models could “predict” unrelated and implausible traits, such as whether patients abstained from consuming beer or refried beans. These predictions, while lacking medical basis, achieved surprising levels of accuracy by leveraging subtle, unintended data patterns.

Dr. Peter L. Schilling, an orthopedic surgeon at Dartmouth Health’s Dartmouth Hitchcock Medical Center and senior author of the study, emphasized the need for caution: “While AI has the potential to transform medical imaging, we must be cautious. These models can see patterns humans cannot, but not all patterns they identify are meaningful or reliable.”

The study also revealed that AI algorithms often rely on confounding variables—such as differences in X-ray equipment or clinical site markers—to make predictions, rather than focusing on medically relevant features. Attempts to eliminate these biases were only marginally successful, as the AI models would identify other hidden data patterns.

Brandon G. Hill, a machine learning scientist at Dartmouth Hitchcock and co-author of the study, noted, “This goes beyond bias from clues of race or gender. We found the algorithm could even learn to predict the year an X-ray was taken.”

The findings underscore the necessity for rigorous evaluation standards in AI-based medical research to prevent misleading conclusions and ensure scientific integrity.

In a related development, researchers at the University of Jyväskylä and the Central Finland Health Care District have developed an AI-based neural network capable of detecting early knee osteoarthritis from X-ray images. The AI matched doctors’ diagnoses in 87% of cases, highlighting its potential to support early diagnosis in primary health care.

These studies collectively highlight both the promise and challenges of integrating AI into medical diagnostics, emphasizing the importance of careful implementation and thorough validation to ensure reliable and meaningful outcomes.

Original Story: AI thought X-rays of your knees show if you drink beer—they don’t. | News | Dartmouth Health