Deadly risk hidden in plain sight: Artificial intelligence can pinpoint which ulcerative colitis patients with precancerous lesions are most likely to develop colorectal cancer—and it does so with impressive accuracy. And this is the part most people miss: such AI-driven insight could transform how doctors warn patients, plan follow-up, and intervene early.
Overview
People with ulcerative colitis (UC), a long‑standing inflammatory bowel disease, face a markedly higher risk of colorectal cancer—up to four times greater than the general population. Among UC patients, low‑grade dysplasia (LGD) refers to abnormal, precancerous tissue and can signal trouble ahead. Yet only a subset of UC‑LGD cases progress to cancer, making it tough for clinicians and patients to decide between continued surveillance and preventive surgery.
A new study from researchers at the University of California San Diego shows that combining artificial intelligence (AI) with biostatistical risk models can accurately identify UC‑LGD patients who are most likely to develop cancer. The findings could improve how we counsel patients, tailor decision‑making, and schedule timely follow‑ups. The study was published on February 17 in Clinical Gastroenterology and Hepatology (full text: https://www.cghjournal.org/article/S1542-3565(26)00066-2/fulltext).
What they did
- Built a fully automated AI workflow to comb through decades of medical records from the U.S. Department of Veterans Affairs (VA) health system, including colonoscopy and pathology notes, to locate UC‑LGD cases and estimate individual cancer risk. This dataset is the largest of its kind in the United States.
- Employed large language models to extract meaningful risk factors for colitis‑associated cancer directly from narrative clinical notes, such as lesion size, the presence of multiple dysplastic sites, and the level of bowel inflammation.
- Combined AI outputs with established statistical risk factors to categorize patients into five risk levels based on four core factors: dysplasia size, completeness and visibility of lesion resection, number of dysplastic sites, and inflammation severity.
- Demonstrated that the model’s risk predictions aligned with real-world outcomes for more than ten years after diagnosis.
- Found that roughly half of UC‑LGD patients fall into the lowest‑risk group, with nearly 99% not developing cancer within two years.
Key insights for care
Many patients have small, seemingly low‑risk lesions, and for years clinicians had limited data to confidently reassure them. The AI tool can quantify risk instead of relying solely on subjective judgment, potentially allowing clinicians to extend surveillance intervals for truly low‑risk individuals. In the study, some patients with unresectable visible lesions—lesions too large or in locations that prevent safe, complete removal—were identified as being at substantially higher risk than clinicians typically estimate.
Impact on clinical practice
The AI system is designed to integrate into everyday clinical workflows, delivering precise, automated risk scores to guide decisions about when to colonize next or whether to pursue surgical options—while easing the workload on care teams.
Curtius, a physician‑scientist at UC San Diego and VA San Diego Healthcare System, emphasizes that risk communication often feels subjective. With this AI pipeline, clinicians could quote a concrete risk score rather than listing risk factors without a clear numeric interpretation. This clarity could also help flag patients who need timely follow‑up colonoscopies, reducing delays that can contribute to cancer progression.
Next steps and broader implications
Researchers plan to validate the tool in populations beyond the VA system and to incorporate additional risk factors, including genetic data. Genomics is increasingly recognized as a key driver of cancer progression, and integrating such information could further refine predictions.
Collaborators include Brian Johnson and Hyrum Eddington (UC San Diego), Samir Gupta and Shailja Shah (UC San Diego and VA San Diego Healthcare System), and Misha Kabir (University College London Hospitals NHS Trust).
Funding came in part from the U.S. Department of Veterans Affairs Biomedical Laboratory Research and Development Service and the National Institutes of Health. Disclosure: Kit Curtius reports no competing interests.
Bottom line
- AI‑assisted risk prediction can stratify UC‑LGD patients more accurately than traditional methods.
- This enables personalized surveillance schedules and timely decisions about surgery, potentially reducing cancer risk and easing patient anxiety.
- The technology holds promise for broader use across diverse patient groups, provided further validation and integration with genetic and other biomarkers.
Would you like this rewritten version to lean more toward a clinical audience (doctors and researchers) or a general audience (patients and the public)?
