William Karnes, MD, is director of the high-risk program and colonoscopy quality at the UCI Health H.H. Chao Comprehensive Digestive Disease Center in Orange, Calif., and chief medical officer of Docbot, a technology that uses artificial intelligence to detect abnormalities from colonoscopy capsule video.
Here, Dr. Karnes shares his thoughts with Becker's ASC Review on the future of AI in the gastroenterology specialty, and how the technology could help patients and physicians.
Question: Can you tell me a little more about the Docbot technology and how you got involved?
Dr. William Karnes: The story goes back to 2012 when I came to UCI and Dr. Chan brought me on to wipe out colon cancer in Orange County. It was a three-pronged approach but one of the most important ones. And we already had really strong evidence even back then that how well we found polyps was critical to reducing the risk of colorectal cancer.
We knew that abnormal detection rates among different colonoscopists varied widely. And yet the prevalence, the presumed prevalence of abnormals in the screening population is about 50 percent or more. So everybody's abnormal detection rate should be 50 percent or more, but it ranges anywhere from as little as 7 percent to as high as 54 percent. So there's this huge spread among colonoscopists and their ability to find polyps.
We know that for every 1 percent improvement we can make in our adenoma detection rate or our ADR, there's a 3 percent to 6 percent reduction in the risk of getting colon cancer. So that was my primary goal, was to set up a prospectively collected data set of all the quality of our colonoscopists and to challenge them to get better.
I created this browser-based system where the nurses, God bless their souls, put in all the data for us live during colonoscopies and that included every nitty-gritty detail you can imagine from the specific indication for the exam, what prep they got, what their Boston Bowel Prep Scores are. Every single polyp was recorded, images recorded. And then later, a week later, each polyp was linked to a specific pathology. So over the course of several years, we had well over 10,000 colonoscopies in this database with 20,000 or 30,000 polyp images each linked to their own pathology.
Four years ago, I met up with Andrew Ninh for the first time. Andrew was 20 at the time and already had a company. He had already started up a registry for allergy in San Francisco at Stanford. He was brilliant in artificial intelligence and natural language processing. So he and I were looking at this database and thinking it was such a gold mine. Everybody should have this. We should create an application so that everybody in the world can be collecting this data. But it's still such a pain because the nurses have to put it in. And then a light bulb went off for both of us that we can use artificial intelligence to substitute for what the nurses are doing.
Q: How are you applying this technology to your work or research?
WK: There's two major aspects to this. Some of our AIs are going to require FDA for polyp detection, for example, optical pathology. But for report writing it doesn't require FDA because all that's doing is assisting and creating a report. It doesn't affect decision making, any clinical decision making. So that part of it we can just get out there. We're in discussions with FDA now and designing clinical studies, small place center studies to validate our polyp detector and our optical path. I expect that within a few months, maybe by early, mid next year, we'll have FDA clearance and get this out there.
Q: What do you believe the impact on patients will be?
WK: Well, the hope is, No. 1, more polyps will be found and they'll be fewer interval colon cancers.No. 2, at the end of the procedure, the physician is spending less time documenting and has more opportunities while the room's getting turned around to actually talk to the patient. No. 3, they're going to be able to talk to the patient with more information. So it could basically eliminate all this need for follow-up by 90 percent or more. And it provides the patient with all the information they need without that anxiety. So I think those are the main issues that a patient would see as a potential benefit.
Q: How do you hope Docbot will affect physician workflow?
WK: One of my dreams for this is automated recording of adenoma detection rate. It's a quality measure that's reportable to CMS, and it's tied to our Medicare reimbursement. If we're really good with our ADRs, we may get a bonus. If we're under a threshold, we may get penalized. It's a pain to record your ADR. You have to basically hire back office staff with Excel spreadsheets, going through all the path after the fact and keeping track of what were the specific indications. And then after you have all that data, you got to record it through a registry that costs money to get it to CMS. If AI can accurately determine your ADR right then and there and just auto record each procedure, whether an abnormal was found or not, that becomes a nonissue.
Q: What AI technologies are you the most excited about?
WK: First of all, we're just touching the surface on what we can do with colonoscopy. We're building algorithms now that have memory. So they can actually remember a specific polyp and give it a name so that you can start counting polyps. We're educating an algorithm now to size polyps. So automated polyp size is going to be a future reality. Eventually, I suppose, they'll be self-driving scopes. Kind of scary with all of this because we don't want to lose our jobs.
A huge target for AI is capsule endoscopy. That's a big, big, big pain point for gastroenterologists. These are 12-hour videos that have to be reviewed. The literature would suggest that even expert capsule readers miss 20 percent of significant lesions. An AI that could just in minutes find all of the abnormal frames, your time to find the abnormality would be reduced dramatically. Your miss rate could be reduced dramatically. And the time to read could be reduced dramatically. So we're developing AIs for that as well.