- Researchers have developed an artificial intelligence (AI) powered model that predicts colorectal cancer risk in patients with ulcerative colitis and low-grade dysplasia.
- Using data from more than 55,000 individuals, the tool could accurately identify very-low-risk patients, potentially helping to reduce unnecessary surveillance colonoscopies.
- The findings suggest AI could support more personalized surveillance strategies while complementing clinician decision making.
Colorectal cancer describes any cancer affecting the colon and rectum. Also known as bowel cancer, it is the
Risk factors for developing colorectal cancer
People living with IBD, especially if untreated,
Although dysplasia can be an early warning sign, detecting which patients are most likely to progress to cancer is a clinical challenge, which can leave patients and clinicians uncertain about when to increase surveillance or consider preventive surgery.
Now, a new study published in Clinical Gastroenterology and Hepatology, suggests that an AI model can accurately predict those most likely to develop cancer, potentially paving the way for more personalized care.
The research team, led by the University of California, San Diego, developed a fully automated AI pipeline that uses large language models to extract relevant clinical information from electronic health records, including colonoscopy and pathology reports.
These records came from more than 55,000 patients in the U.S. Department of Veterans Affairs (VA) healthcare system.
The AI system identified key predictors of cancer progression. This included lesion size, inflammation severity, and whether lesions could be completely removed. The system then integrated these predictors with traditional risk factors into a comprehensive risk model.
The model successfully categorized patients into 5 distinct risk groups that aligned closely with real-world outcomes over more than a decade of follow-up.
Notably, the tool correctly determined that nearly 99% of patients in the lowest-risk category would not develop colorectal cancer within 2 years.
Kathleen Curtius, PhD, assistant professor of medicine in the Division of Biomedical Informatics at UC San Diego School of Medicine, and study author, spoke to Medical News Today about how this tool could help reduce unnecessary surveillance procedures for low risk individuals:
“Current guidelines suggest patients in this low-risk group should come back for a follow-up colonoscopy in 2 years.”
“The data for this group of U.S. Veterans, however, matched our model’s prediction — these patients are at ~1% risk of high-grade dysplasia or cancer by 2 years, and so the 2-year surveillance interval can likely be safely extended in practice. This would save healthcare costs and decrease worry for these patients,” Curtius said.
It can be challenging for clinicians to estimate the cancer risk for a person living with low-grade dysplasia, which can result in frequent colonoscopies.
Using this AI approach, clinicians may be able to personalize screening intervals more effectively, thereby reserving intensive surveillance for those with the highest predicted risk and minimizing interventions for those at low risk.
“Our study shows that the cancer risk prediction model we developed and tested in U.K. patients with ulcerative colitis and low-grade dysplasia also performs well in U.S. populations,” Curtius told MNT.
“This is a major step toward broader clinical use. The statistical model uses established clinical risk factors, which can be pulled directly from doctors’ notes using large language models, highlighting how easily it could fit into real-world clinical workflows.”
— Kathleen Curtius
Interestingly, the model also flagged patients with unresectable visible lesions. This describes lesions that cannot be safely removed due to size or location. The AI system highlighted that individuals with these lesions are at significantly higher risk than many clinicians typically estimate in routine clinical practice.
“Doctors often underestimate the imminent risk of high-grade dysplasia and/or colorectal cancer developing after a visible low-grade dysplasia lesion cannot be completely resected,” Curtius noted.
“This is important to get right because patients decide on major [preventive] surgery partly based on the cancer risk their doctor tells them. Using our tool will help doctors and patients weigh accurate risk estimates when deciding on treatment options, including partial or full colon removal to prevent likely cancers,” she said.
The technology could also help flag individuals who need to return to the clinic, potentially preventing delays in follow-up colonoscopies.
Importantly, the research team notes that the AI tool is designed to complement clinician judgement. The predictions may offer additional evidence to support shared decision making between patients and their healthcare teams.
Curtius spoke about how this tool could be integrated into real-world clinical workflows:
“This tool is already available as a web tool for doctors to use with their patients, but it can next be readily integrated directly into the electronic Health record system.
AI tool + doctor insight“The risk factor data is already collected routinely by doctors, and then the patient’s electronic health record dashboard could compute and display the future cancer risk over time during clinical decision making.”
— Kathleen Curtius
Although the results are promising, the authors emphasize the need to validate the model in diverse patient populations outside the VA healthcare system.
Curtius notes that this model may help to support shared decision making:
“This approach could help reduce unnecessary surveillance colonoscopies and surgeries by giving doctors and patients confidence when someone’s cancer risk is very low.”
“At the same time, giving doctors and patients clear numbers and a visual tool to convey when cancer risk is very high can make shared decision making easier and help people better understand the risks of a ‘watch-and-wait’ approach,” she said.
The research team also plans to explore integrating emerging genetic risk factors into the algorithm to further enhance its predictive accuracy.
