Penn Medicine’s AI finds disease signals beyond a CT scan’s original purpose in study

A CT scan ordered for one medical problem may contain clues to others. Penn Medicine researchers say their AI model, Percival, identified disease signals beyond the reason scans were taken. The findings, published in npj Digital Medicine, point toward broader use of existing images, but the model remains in preclinical testing—not routine care.

Looking beyond the original question

In the study, Penn Medicine reported that Percival identified signs of atrial fibrillation, a form of irregular heartbeat, and chronic kidney disease. It also flagged pancreatic cancer, cirrhosis and bacterial pneumonia more reliably than another AI model, TotalSegmentator. The announcement gives no numerical comparison to show the size of that advantage.

The intended use is “opportunistic screening”: looking for health information beyond a scan’s original purpose. Senior author Walter Witschey, a Penn Medicine associate professor of radiology, offered the example of finding early traces of cancer in an image taken to assess pneumonia.

Co-author Charles Kahn, a Penn radiology professor, framed the goal as earlier recognition. Many patients already have substantial heart or metabolic disease when they reach specialty care, he said, making routine imaging a potential opportunity for earlier intervention.

Learning from scans already collected

The team trained Percival using Penn Medicine BioBank data, drawing on roughly 400,000 CT scans from approximately 50,000 patients. The study is titled “Generalizable CT vision-language modeling for population health and disease risk.” The broader BioBank contains information from more than 350,000 people, collected across the University of Pennsylvania Health System over two decades.

Lead author Cameron Beeche, a Penn bioengineering PhD candidate, said larger and more diverse datasets can improve performance. He described that breadth as a stronger starting point for adapting a model to other hospitals or patient populations.

Toward a clinical test

The developers hope to test Percival in early 2027, with the aim of eventually bringing it into regular clinical use. They are also examining whether larger models and data spanning multiple institutions could make it more robust and predictive.

Percival’s code and training-model specifications are open source. Witschey said the project benefited from researchers sharing code, methods and insights, and described the release as a contribution to that same approach.