A team of Norwegian researchers, through international collaboration, found that the use of artificial intelligence (AI) during colonoscopy increased the detection of polyps by about 8%, but the risk of cancer was minimally affected.

Colorectal cancer is one of the most common forms of cancer in the western world. In Norway, approximately 5,000 people are affected by colorectal cancer each year. Norway has recently initiated a national screening program for colorectal cancer.

Currently, colorectal cancer screening in Norway is performed using a test that detects blood in stool (FIT screening). If a certain amount of blood is detected in the stool, the patient is referred for a colonoscopy. However, the Norwegian Directorate of Health has clarified that the stool test will be replaced with colonoscopy for all 55-year-olds in Norway within the next five years.

During colorectal cancer screening with colonoscopy, AI is often used as an assistant to identify polyps during the examination.

The clinical benefits of AI need to be documented like other tools for diagnosis and treatment. This was the reason for developing international guidelines for the use of AI in colonoscopy examinations, through BMJ Rapid Recommendations, led by the Norwegian foundation MAGIC.

Such tools also represent the most explored area for AI in clinical medicine. A recent systematic review conducted as part of the guideline work summarized 44 randomized studies with over 30,000 patients. AI detects more polyps but has not assessed long-term effects such as cancer occurrence and mortality.

A new large microsimulation study, led by the University of Oslo, has analyzed the effect of AI on the occurrence of cancer and mortality, as well as drawbacks in the form of more follow-up colonoscopies when using AI in colorectal cancer screening. The researchers have reviewed data from about 15 studies in addition to the systematic overview.

The findings were urgently published this week in BMJ Medicine, as a basis for the guidelines BMJ Rapid Recommendations.

“The project shows that while artificial intelligence can reduce cancer risk by improving the detection of polyps, it must be weighed against the drawbacks of increased follow-up burden. The risk of cancer, which is the most important for patients, is minimally affected,” says physician and Ph.D. candidate Natalie Halvorsen from the Clinical Effectiveness Research Group at UiO.

Three different guidelines based on the Norwegian researchers’ work

“The guideline was urgently published on March 27 in BMJ due to its significant impact on the use of AI in colorectal cancer screening,” explains Halvorsen.

The recommendation in advises against using AI as a routine for general colonoscopy screening of adults. However, it is noted that there is uncertainty in the recommendation and that health economic assessments should be included in the next dynamic update, which may change the recommendations.

In addition to the recommendation in BMJ Rapid Recommendations, the American Gastroenterology Association (AGA) and the European Society of Gastrointestinal Endoscopy (ESGE) published their own recommendations on AI this week, based on the work led by the Norwegian researchers. They draw different conclusions than BMJ.

AGA states that no recommendation could be made (gastrojournal.org) either for or against using AI-assisted colonoscopy. ESGE recommends (thieme-connect.de) use based on the assumption that a majority of patients undergoing colonoscopy screening would prefer assistance.

“This shows the complexity of the recommendations and how the different recommendation criteria are weighed differently, such as drawbacks of overdiagnosis and increased burden on the health care system,” says Halvorsen.

“The guidelines show that we have good methods to evaluate AI tools and that we can collaborate globally on knowledge and recommendations for practice. So far, there are few examples of AI tools benefiting patients, but the development of knowledge is so rapid that we must dynamically update recommendations. We achieved this during COVID-19 and should also do so for AI,” emphasizes Professor Per Olav Vandvik, founder of MAGIC and leader of BMJ Rapid Recommendations at the University of Oslo.

Burden on patients and health care systems

The new guideline and large study published in BMJ Medicine are part of the EU-funded Opera-project (uio.no). Opera, led by UiO Professor Yuichi Mori, aims to investigate whether the use of AI can help reduce cases and mortality of colorectal cancer.

“Our findings show that artificial intelligence detects more polyps but has only a very modest effect on cancer risk, which must be weighed against increased burden on the health care systems. The introduction of AI in health care should be limited to tools that have shown clinical benefits” says Halvorsen.

“It is important to remember that AI tools can have negative effects on individual patients and on the health care system. This means that both benefits and drawbacks should be carefully assessed before introducing AI into health care,” says Mori.

Key findings include:

• Reduction in cancer risk: Using AI in colonoscopy had little or no clinically relevant effect on the risk or mortality of colorectal cancer after screening: AI reduced cancer risk over 10 years from 0.82% without AI to 0.71% with AI. For colonoscopy after FIT screening, the risk fell from 5.82% to 5.77%.
• Increased follow-up: The use of AI resulted in a 20% increase in follow-up colonoscopies for patients over 10 years due to increased detection of polyps. This leads to more examinations for individual patients and more burden on health care systems.

The project shows that although AI can reduce cancer risk by improving the detection of polyps, it must be weighed against the drawbacks of increased follow-up burden. The risk of cancer, which is most important for patients, is minimally affected.