Percival AI spots hidden disease in CT scans, still preclinical

A CT scan taken to check for pneumonia can carry far more than that answer. Somewhere in its slices, a radiologist's eye may miss the early signs of a heart rhythm problem or a failing kidney, because nobody asked the question. Penn Medicine researchers have built an artificial intelligence model that tries to find exactly those things, and a new study in npj Digital Medicine reports what it can pick up.

The model, called Percival, was trained on roughly 400,000 CT scans from about 50,000 patients, drawn from the Penn Medicine BioBank, which gathers data on more than 350,000 people across the University of Pennsylvania Health System over two decades. Most AI tools have to be pointed at a specific condition. Percival is designed to run in the background and surface patterns nobody thought to look for. The team says it found traces of atrial fibrillation, an irregular heartbeat, and chronic kidney disease in scans. Against another AI model, TotalSegmentator, Percival flagged pancreatic cancer, cirrhosis and bacterial pneumonia much more reliably, according to the study's authors.

The researchers call this approach 'opportunistic screening': finding disease on a scan ordered for something else. Co-author Charles Kahn, a radiology professor, argues that many patients already have substantial heart or metabolic disease by the time they reach specialty care, so an image taken for another reason could bring that forward. The results are still preclinical. They come from the team's own testing, and the model has not yet been used on patients in routine care.

For patients, the practical gain would be earlier recognition, and for hospitals, a way to get more out of imaging they already perform. The developers hope to test Percival early next year, aiming to bring it into the clinic for regular use. Until then, the gains are a promise to be checked: the team is examining how larger datasets and data from several institutions could make the model more robust and predictive. Its code and training specifications are open source, so other groups can test it too.

Penn Medicine is also developing other tools. A separate model for cardiac MRI scans, built to help diagnose new conditions, reached near-perfect accuracy in one test, and the team's earlier AInSights tool is expected to help guide Percival's path to the clinic.