AI workflow predicts colorectal cancer risk in ulcerative colitis with 99% accuracy
A study published on February 17, 2026 validated an AI workflow that predicted colorectal cancer risk in ulcerative colitis with 99% accuracy using over 55,000 patient records. The model identified low-risk and higher-risk patients from clinical notes, colonoscopy reports, and pathology findings.
A new AI workflow utilizing large language models can predict with 99% accuracy which ulcerative colitis patients will remain free of colorectal cancer for at least two years. A study published on February 17, 2026, validates a new AI workflow capable of predicting colorectal cancer risk with 99% accuracy in patients with ulcerative colitis.
Patients with this inflammatory bowel disease are up to four times more likely to develop colorectal cancer than the general population, making precise risk stratification for low-grade dysplasia a critical clinical necessity. This primary study utilized large language models to analyze over 55,000 patient records from the Department of Veterans Affairs health system, the largest dataset of its kind.
The automated workflow effectively sifts through narrative clinical notes, colonoscopy reports, and pathology findings to identify subtle risk factors like lesion size and inflammation severity. The system identifies high-risk features from narrative clinical notes that are often overlooked, potentially allowing for more personalized and less invasive surveillance schedules.
By accurately grouping patients by risk, the AI identified approximately half of the cohort as low-risk, suggesting they could potentially safely extend their surveillance intervals. Conversely, the model flagged visible lesions as higher risk than clinicians typically estimate, advocating for more timely follow-up or preventative surgery in those cases.
The integration of this technology into gastroenterology practices could standardize the management of dysplasia, which is currently subject to high inter-observer variability among pathologists. The automated nature of the tool also saves significant physician time by providing a “one-click” risk score within the electronic health record.
However, the reliance on high-quality digital clinical notes may limit the tool’s effectiveness in health systems with less structured records or disparate reporting styles. Further prospective validation in non-veteran populations and community settings is still required to confirm these high accuracy rates across broader demographics. Currently, the research team is looking to adapt the model for other inflammatory conditions to improve long-term monitoring.