Patients diagnosed with ulcerative colitis (UC), a chronic inflammatory bowel disease, are up to four times more likely to develop colorectal cancer than the general population.
Low‑grade dysplasia (LGD), abnormal or precancerous lesions, can be an early warning sign, but only a fraction of UC‑LGD cases progress to cancer. This makes it challenging for clinicians and patients to make informed care decisions, ranging from continued surveillance to preventative surgery. Because of the uncertainty in existing prognostic data, current guidelines recommend incorporating expert opinion into management decisions. However, these decisions are generally imprecise.
Now, a new study led by researchers at the University of California, San Diego, has found that large language models (LLMs), also referred to as artificial intelligence (AI), when combined with biostatistical risk models, can accurately predict which UC‑LGD patients are most likely to develop cancer. The findings have the potential to significantly improve patient counseling, decision‑making, and timely follow‑up care. The study, funded in part by the U.S. Department of Veterans Affairs Biomedical Laboratory Research and Development Service and the National Institutes of Health, was published on February 17, 2026, in Clinical Gastroenterology and Hepatology. [1]
Automated workflow
The researchers created a fully automated AI workflow to sift through the past medical records — including colonoscopy and pathology reports — of 55,000 patients in the U.S. Department of Veterans Affairs (VA) health care system to identify UC-LGD patients and assess their individual cancer risk. The dataset is the largest of its kind in the U.S.
“Large language models accurately derived colitis-associated colorectal cancer risk factors — such as how big the low-grade dysplasia lesion is, whether there are multiple lesions and if the colon is extremely inflamed — from the narrative clinical notes themselves,” noted Kit Curtius, PhD, assistant professor of medicine in the Division of Biomedical Informatics at UC San Diego School of Medicine and a member of Moores Cancer Center.
The AI workflow and statistical risk model predictions:
- Correctly grouped the patients into five risk categories based on four established factors: dysplasia size, lesion resection completeness and visibility, number of dysplastic sites, and severity of inflammation
- Matched real‑world patient outcomes with high accuracy for more than a decade after diagnosis
- Classified nearly half of the patients into the lowest-risk group, correctly determining that almost 99% will avoid a cancer diagnosis within two years
“A lot of people are low risk. They have small dysplastic lesions, and it’s been hard to know what to confidently tell these people until now,” explained Curtius, who is also a research health scientist at VA San Diego Healthcare System.
“Normally, it is recommended that patients with small lesions return for cancer surveillance in two years. “With this tool, there may be a potential to increase the surveillance interval so patients who are at this low risk don’t have to come back so often,” Curtius added.
The AI model also revealed that patients with unresectable visible lesions, lesions that cannot be safely and completely removed through surgery due to size, location or extent of spreading, are at significantly higher risk than many clinicians typically estimate.
A boon to patient care
The study indicates that AI models can integrate naturally into clinical workflows, offering precise, automated risk assessments to guide clinician and patient decision-making—from timing their next colonoscopy to when to consider surgery, while reducing the burden on care teams.
“Currently, the process of advising people about levels of risk is a somewhat subjective thing, and doctors don’t have enough data to back up what they feel,” said Curtius.
“This AI pipeline could read the clinical notes and tell you your risk score, rather than just having a list of risk factors and no real way to turn that into a number during a patient visit,” he added.
The technology may also help flag patients who need to return to the clinic, helping prevent delays in follow-up colonoscopies, a major contributor to colorectal cancers.
The next steps include validating the AI tool in patient populations outside of the VA system and incorporating emerging risk factors and patient genetic information.
We know that genomics plays a big part in driving cancer progression,” Curtius concluded.
Reference
[1] Johnson B, Eddington H, Kabir M, Gupta S, Shah SC, Curtius K. Automated phenotyping with AI predicts future advanced neoplasia risk in colitis-associated low-grade dysplasia DOI 10.1016/j.cgh.2026.01.037.
Featured image: Large Learning Model, Artificial Intelligence. Photo courtesy: © 2024 – 2026 Philip Oroni. Licensed under Unsplash+ License.
DOI
