Imagine being told you have a condition that quadruples your risk of developing colorectal cancer. That's the stark reality for individuals living with ulcerative colitis (UC), a chronic inflammatory bowel disease. But here's the catch: not all UC patients with precancerous lesions will actually develop cancer. This leaves both patients and doctors in a difficult position, unsure whether to opt for aggressive monitoring or preventative surgery. And this is the part most people miss: the uncertainty can lead to unnecessary procedures or, worse, delayed treatment.
A groundbreaking study from the University of California San Diego is changing the game. Researchers have developed an artificial intelligence (AI) system that, when combined with biostatistical models, can predict with remarkable accuracy which UC patients with low-grade dysplasia (LGD) are most likely to progress to cancer. Published in Clinical Gastroenterology and Hepatology, this research offers a beacon of hope for more informed, personalized care.
The team harnessed the power of AI to analyze the medical records of 55,000 patients from the U.S. Department of Veterans Affairs (VA) healthcare system—the largest dataset of its kind in the U.S. The AI workflow scoured colonoscopy and pathology reports, identifying UC-LGD patients and assessing their individual cancer risk.
But here's where it gets controversial: The AI didn't just rely on traditional risk factors; it extracted critical insights directly from narrative clinical notes. "Large language models accurately pinpointed colitis-associated colorectal cancer risk factors—like lesion size, multiplicity, and colon inflammation—from the doctors' own written observations," explained Kit Curtius, PhD, assistant professor of medicine at UC San Diego School of Medicine. This raises questions about the future role of AI in interpreting complex medical data—could machines one day rival human expertise?
The AI system achieved impressive results:
- It categorized patients into five risk groups based on dysplasia size, lesion removal completeness, number of affected sites, and inflammation severity.
- Its predictions matched real-world patient outcomes with high accuracy over more than a decade.
- Nearly half of patients were correctly classified into the lowest-risk group, with a 99% chance of avoiding cancer within two years.
"Many UC patients have small, low-risk lesions, and until now, it's been challenging to provide clear guidance," said Curtius, also a research health scientist at VA San Diego Healthcare System. "This tool could allow us to extend surveillance intervals, reducing the burden on low-risk patients."
However, the AI also uncovered a surprising finding: patients with unresectable visible lesions—those deemed too risky for complete surgical removal—face a significantly higher cancer risk than many clinicians assume. This challenges conventional wisdom and may prompt a reevaluation of current treatment protocols.
The implications for patient care are profound. The AI models seamlessly integrate into clinical workflows, providing automated, precise risk assessments. This could revolutionize decision-making, from scheduling colonoscopies to planning surgeries, while easing the workload on healthcare teams.
"Risk counseling today is often subjective, with limited data to support recommendations," Curtius noted. "This AI pipeline can analyze clinical notes and provide a concrete risk score, transforming how we communicate with patients."
The technology could also improve follow-up care by identifying patients at risk of missing critical appointments, a major factor in colorectal cancer progression. But this raises another controversial question: as AI takes on more responsibilities, how do we ensure it doesn't replace the human touch in medicine?
Looking ahead, the researchers plan to validate the AI tool in diverse patient populations beyond the VA system and incorporate emerging risk factors, including genetic data. "Genomics plays a huge role in cancer progression, and integrating this information will further refine our predictions," Curtius added.
What do you think? Is AI the future of cancer risk prediction, or does it risk dehumanizing patient care? Share your thoughts in the comments below.
Study co-authors include Brian Johnson, Hyrum Eddington (UC San Diego), Samir Gupta, Shailja C. Shah (UC San Diego & VA San Diego Healthcare System), and Misha Kabir (University College London Hospitals NHS Trust). The research was funded by the U.S. Department of Veterans Affairs and the National Institutes of Health.
