A new study from the University of California San Diego shows that AI models can analyze clinical notes to predict which ulcerative colitis patients with low-grade dysplasia are most likely to develop colorectal cancer. Using data from over 55,000 patients in the U.S. Veterans Affairs system, the automated workflow extracted key risk factors—such as lesion size, number of dysplastic sites, resection completeness, and inflammation severity—from narrative colonoscopy and pathology reports.
The model accurately stratified patients into five long-term cancer risk categories, correctly identifying nearly half as low risk, with ~99% avoiding cancer within two years. By turning unstructured clinical documentation into actionable risk scores, the approach could help guide surveillance intervals, surgical decisions, and follow-up timing—potentially reducing delays in colonoscopy and improving personalized care in UC-associated colorectal cancer.