Penn Medicine's Percival AI reads CT scans for diseases the scan wasn't ordered for

Penn Medicine on Friday, Oct. 16, 2026, described Percival, a CT vision-language model that flags chronic disease traces and future risk beyond the reason a scan was ordered, based on a new study in npj Digital Medicine.

Hospitals already shoot millions of CTs for one question. Percival's bet is that the same scan is a full-body rumor mill — AFib, kidney disease, pancreatic cancer hints — if a model is allowed to look past the order ticket. Open weights plus a BioBank-sized training set make this more than a press-room demo. The catch is the same as every opportunistic-screening pitch: preclinical today, clinic maybe next year, and every false alarm has a patient attached.

Penn Medicine News, dated Oct. 16, 2026, described a new AI model called Percival that analyzes CT scans for health signals beyond the original clinical indication. A CT scan is a 3D X-ray. The study is published in npj Digital Medicine. Those lines are Penn Medicine's.

The name Percival references a Round Table knight and the late radiologist Eugene Percival Pendergrass. The model was built with Penn Medicine teams using Penn Medicine BioBank data: information from more than 350,000 people across the University of Pennsylvania Health System over about two decades. Those lines are Penn Medicine's.

Training scale, as Penn cites it: roughly 400,000 CT scans from approximately 50,000 patients. The model's abilities are still being tested preclinically, meaning outside regular clinic use. Those figures are Penn Medicine's.

Senior author Walter Witschey, PhD, associate professor of Radiology and associate vice chair of Research IT, said the model learns broadly from imaging and connects scans, lab values, physiology, and clinical outcomes for more accurate assessments, or even earlier alarm bells. Those lines are his, as Penn prints them.

Lead author Cameron Beeche, a PhD candidate in Bioengineering at the University of Pennsylvania, said larger, more diverse training can improve performance and give a stronger foundation for adapting a model across hospitals and patient populations. Those lines are his, as Penn prints them.

Penn says Percival uncovered traces of atrial fibrillation, an irregular heartbeat, and chronic kidney disease. Compared with a similar model called TotalSegmentator, Penn says Percival flagged pancreatic cancer, cirrhosis, and bacterial pneumonia more reliably. Those comparisons are Penn Medicine's. The newsroom page does not print exact sensitivity or AUC scores for those flags.

Co-author Charles Kahn, MD, a professor of Radiology, said Percival can help identify unrecognized disease, predict future progression, and anticipate major cardiovascular and liver complications. Those lines are his, as Penn prints them.

Developers hope to test Percival early next year, with an eye toward clinic use. Related Penn tools named on the same page: AInSights, and a cardiac MRI diagnostic model. Percival's code and training-model specifications are described as open source. Funding acknowledged: NIH Office of the Director award OT2OD038048 and NIBIB award R21EB036734. Those lines are Penn Medicine's. The release does not say the Food and Drug Administration has cleared Percival, and it does not say the model is in clinic today.

The picture is Figure 1 from the Percival paper: the Penn Medicine BioBank CT cohort pipeline and a CT-report contrastive vision-language model diagram. A vision-language model learns from pictures and text together. Contrastive, here, means the training pulls matching scan-and-report pairs together in a shared space. The figure is from the open-access npj Digital Medicine article. The frame does not print a calendar date.

In plain terms, Penn Medicine said on Friday that Percival, trained on about 400,000 CTs from about 50,000 patients, can flag disease traces a scan was not ordered to find — including AFib, kidney disease, and, versus TotalSegmentator, pancreatic cancer, cirrhosis, and bacterial pneumonia. The code is open source. The work is still preclinical. The team hopes to test toward clinic use early next year.