Percival AI model detects hidden health issues in CT scans beyond original diagnosis

Penn Medicine researchers have developed Percival, an artificial intelligence model trained on 400,000 CT scans from 50,000 patients, capable of detecting unrequested medical conditions such as pancreatic cancer, atrial fibrillation, and kidney disease. Operating background analyses during routine imaging, Percival outperforms existing tools like TotalSegmentator in preclinical tests. With open-source code now available, developers plan clinical trials in early 2027 to advance proactive healthcare.

How the Percival AI model operates

• ▪Penn Medicine researchers designed the Percival artificial intelligence model to advance opportunistic screening, allowing clinicians to discover early traces of unrequested health conditions like cancer or cardiovascular disease on scans ordered for other primary reasons.
• ▪The Percival artificial intelligence model operates in the background to identify health patterns, clinical outcomes, and connections between CT scans, lab values, and physiology without requiring specific direction to search for particular outcomes.

Preclinical test results and detection capabilities

• ▪In preclinical tests, the Percival artificial intelligence model detected hidden traces of chronic conditions, including atrial fibrillation and chronic kidney disease, directly from CT scans.
• ▪When evaluated against the AI model TotalSegmentator, the Percival model flagged conditions including pancreatic cancer, cirrhosis, and bacterial pneumonia significantly more reliably.
• ▪Penn Medicine researchers reported that routine medical imaging analyzed by the Percival artificial intelligence model offers opportunities for earlier medical intervention, especially for patients entering specialty care with existing metabolic or heart conditions.

Training data used for Percival

• ▪The Percival artificial intelligence model was trained using data from the Penn Medicine BioBank, which contains information from more than 350,000 people collected across the University of Pennsylvania Health System over two decades.
• ▪Researchers used approximately 400,000 CT scans from about 50,000 patients to train and test the Percival artificial intelligence model preclinically.

Future development plans and clinical testing

• ▪Developers of the Percival model are drawing from their experience with another Penn Medicine AI radiological solution called AInSights and are examining multi-institution datasets to improve model robustness.
• ▪Penn Medicine developers plan to begin early clinical testing for the Percival artificial intelligence model in early 2027 with the goal of bringing the model into regular clinical use.

Publication and open-source release

• ▪The code and training model specifications for the Percival artificial intelligence model have been made open source by the Penn Medicine research team.
• ▪The findings and technical specifications regarding the Percival model were published by lead author Cameron A. Beeche and senior author Walter Witschey in the journal npj Digital Medicine.