New AI audits colonoscopies

A new study of 19,000 colonoscopies shows AI can accurately assess quality metrics from procedure videos.
A team of Northwestern Medicine scientists developed an AI tool that rapidly and successfully measures colonoscopy quality by reviewing procedure footage.
In a study of nearly 19,000 colonoscopies published in The American Journal of Gastroenterology, the Northwestern team demonstrated the tool’s high accuracy by comparing its quality measurements with those of nurses and other clinicians. The authors say the study marks the first time an AI tool is shown to comprehensively measure the quality of thousands of colonoscopies.
Study lead author Dr Rajesh Keswani, associate professor of medicine in the division of gastroenterology and hepatology at Northwestern University Feinberg School of Medicine, said: ‘Our AI software provides a scalable approach to measuring colonoscopy quality and, ultimately, providing feedback to clinicians to improve care quality.’
Previous studies have shown that measuring colonoscopy performance improves overall colonoscopy quality and, ultimately, reduces colorectal cancer mortality. By automating that quality-monitoring process, the new tool could make colonoscopy feedback more feasible across entire hospitals or healthcare systems, said Keswani, who is also a Northwestern Medicine physician.
He said: ‘The goals of screening colonoscopy are to provide safe and effective care. This is not possible unless an institution can measure the quality of colonoscopy.’
The AI also tracked quality indicators that ‘can't be feasibly measured by humans at scale’, according to Keswani. These included the number of polyps removed during a procedure and how often physicians used cold snare polypectomy, a guideline-recommended technique for removing small polyps.
Last year, a Lancet study made headlines after suggesting that colonoscopists who started using an AI helper to detect polyps became less proficient over time. Keswani said he’s not sure yet how to view AI’s impact on colonoscopies and medicine in general. He also noted that his AI tool assesses quality only after colonoscopies are completed.
He added: ‘One possibility is that physicians rely too much on AI, which leads to deskilling. Alternatively, AI can teach us about blind spots and make us better clinicians.’
Keswani added that his research team is currently studying AI's role and impact in teaching colonoscopy to trainees.