A new artificial intelligence model can uncover health problems hiding in CT scans that were ordered for entirely different reasons, according to research from Penn Medicine. The AI model, dubbed Percival, can detect nearly hidden patterns of disease and potential future health risks that might otherwise go unnoticed by human eyes. Unlike other models that must be directed to look for specific conditions, Percival can effectively run in the background, picking up what is unseen and pointing towards a future of more personalised and proactive healthcare.
What Percival Can See in a CT Scan
In tests, the model uncovered traces of several chronic conditions in routine scans, including atrial fibrillation, a form of irregular heartbeat, and chronic kidney disease. When compared with a similar AI tool called TotalSegmentator, Percival flagged conditions such as pancreatic cancer, cirrhosis and bacterial pneumonia far more reliably. Researchers say the tool learns broadly from imaging data, uncovering connections between scans, lab values, physiology and clinical outcomes to give more accurate assessments of health, or even ring alarm bells early before symptoms appear.
The concept is known as opportunistic screening: finding results beyond the original purpose of a scan, such as spotting the early traces of cancer on a scan ordered to assess pneumonia. Co-author Charles Kahn, a professor of Radiology, said the model can help identify unrecognised disease, predict future progression and anticipate major cardiovascular and liver complications. Because many patients already have substantial heart or metabolic disease when they enter specialist care, routine imaging analysed this way could create an opportunity for earlier recognition and intervention.
For radiologists, the appeal is clear. A standard CT scan contains far more information than the human eye can absorb during a busy clinical shift, and incidental findings have always been a blind spot: subtle signs of disease that are visible in retrospect but missed in the moment. A tool that runs quietly in the background, reviewing every scan for a broad range of risks without adding to the radiologist’s workload, could change that equation. It would not replace the doctor’s judgement, but it could act as a tireless second reader, flagging cases that deserve a closer look and helping clinics catch illness earlier.
How Percival Was Trained
Percival was built on an exceptionally large and diverse dataset. The team drew on the Penn Medicine BioBank, which holds information from more than 350,000 people collected across the University of Pennsylvania Health System over two decades. The researchers then used roughly 400,000 CT scans from about 50,000 patients to hone the model’s abilities. Lead author Cameron Beeche, a PhD candidate in Bioengineering, said training on larger, more diverse datasets improves performance and provides a stronger foundation for adapting a model to a hospital or patient population almost anywhere in the world.
Percival vs Other AI Radiology Tools
What sets Percival apart, its developers say, is breadth. Rather than being trained to answer a single narrow question, it learns general patterns from imaging data and links them to outcomes. That breadth is what allows it to run quietly in the background of a radiology department, screening every scan for a wide range of risks without needing to be told what to look for. The team believes experience gained from another Penn Medicine AI radiology solution, AInSights, will help them refine the new model and prepare it for the demands of everyday clinical work.
The Road to the Clinic
For now, Percival remains in preclinical testing, and its abilities are still being evaluated. The developers hope to begin testing it in real clinical settings early next year, with an eye towards bringing the model into regular use. If that effort succeeds, the humble CT scan, already one of medicine’s most common imaging tools, could become far more than a picture of a single suspected condition. It could turn every scan into a window onto a patient’s broader health, catching disease earlier and giving doctors a head start on treatment.