The use of artificial intelligence (AI) in medical care is a salient example of the technology’s value. Yet AI’s improvements may be marginal and may also introduce downsides, such as higher costs and redundant or unnecessary care.
The recently released official guidelines on the use of AI in colonoscopy procedures by the American Gastroenterological Association are an example of these tradeoffs. A clinical panel, comprised of 10 experts in the field, came down with a “neutral” recommendation on using AI for the procedure.
“The panel reached the conclusion that no recommendation could be made for or against the use of CADe-assisted colonoscopy in light of very low certainty of evidence for the critical outcomes, desirable and undesirable (11 fewer colorectal cancers per 10,000 individuals and two fewer colorectal cancer deaths per 10,000 individuals), increased burden of more intensive surveillance colonoscopies (635 more per 10,000 individuals), and cost and resource implications,” the panel concluded in the guidelines, published in March 2025.
In fact, the neutral guidance suggests that, in medical care at least, the popular imagination’s perception of AI as some sort of magical breakthrough is often far from the truth.
“This guideline highlights the close tradeoff between desirable and undesirable effects and the limitations in the current evidence to support a recommendation,” the authors concluded. “The panel acknowledged the potential for CADe (computer-aided detection systems) to continually improve as an iterative artificial intelligence application. Ongoing publications providing evidence for critical outcomes will help inform a future recommendation.”
“It wasn’t unanimous by any means,” said guideline co-first author Dr. Dennis Shung, M.D., director of digital health in digestive diseases and of applied artificial intelligence at the Yale Center for Healthcare Simulation at Yale University. “Some people still felt very strongly that we needed to push this for patients, because if you are depriving them of technology that might identify that polyp that could become a cancer, why would you not?”
However, Shung said, the balance of all factors led most the panel to issue a neutral vote.
“From a patient standpoint it’s almost impossible to figure out ‘Do I need this or not?’ and that’s part of the reason why we came down with the recommendation we had,” Shung said. “For patients, the marginal benefit is unclear. We may find a few more adenomas (non-cancerous tumors). Do those really need to be removed at that point? Maybe, maybe not.
“But what’s guaranteed is that that patient wants to come back earlier for another colonoscopy. And that is expensive for the patient, for the system, and also again, what is that actual benefit to the patient other than having to do another procedure at a shorter interval?”
Factors Beyond Image Analysis
Official guidelines aside, how quickly gastroenterologists will adopt AI for colonoscopy depends on factors beyond whether or not the technology helps them recognize small anomalies in the colon wall faster than they would with the naked eye. Factors including interface design, billing, and patient preference also come into play.
Medtronic, a medical technology company in the gastroenterological field, is running consumer-oriented advertising that touts the AI features of its product, GI Genius. Medtronic did not respond to requests for comment.
As to the cost-effectiveness of such “ask your doctor” marketing, the costs of direct-to-consumer advertising for pharmaceuticals in the U.S. were examined in a 2023 paper by researchers affiliated with Harvard, Yale, and Dartmouth. They found that just 27% to 30% of 81 highly advertised drugs were rated as having high therapeutic value. They said $15.9 billion was spent on television advertising of drugs with low therapeutic value from 2015 to 2021.
Shung said the gastroenterological clinical community is wary of adopting technology that will lead to higher costs without commensurate returns. “It doesn’t make sense economically from their standpoint,” he said. “Reimbursements for colonoscopies are decreasing every year in real terms anyway; they’re not adjusted for inflation. So, to add an additional expense and make that the cap of what we can make, that’s why nobody wants to open up the billing codes.”
Shung also mentioned overly-busy user interface design that leads to frustration akin to the “alarm fatigue” nurses often report in hospital settings, as machines monitoring patients’ vital signs often activate for trivial developments.
“Some people don’t want to use it,” Shung said. “It’s distracting. Some people who use these systems say it distracts them from doing a good quality exam. It keeps flashing whenever it sees a polyp or a lesion; even if you’ve seen it, it will still flash. So, you can imagine if you’re doing a procedure and you keep getting distracted by this flashing light, you are not actually looking at the rest of the colon. Some providers hate it so much they turn it off.”
Anthony Borrelli, executive director of endoscopy product and marketing for Fujifilm Healthcare Americas, said the company works closely with the clinical community to address such concerns. For example, he said, Fujifilm’s Cad Eye endoscopy AI platform, which received U.S. Food and Drug Administration approval in 2024, can be turned on and off as the user prefers; in addition, Borrelli said continual refinement of its algorithm can also help alleviate false positives that often trigger unnecessary notifications.
“If that AI is continually going off, then you either mentally turn it off in your mind or physically turn it off,” he said. “We have gotten very positive feedback on not having a significant number of false positives.”
Borrelli said in today’s medical economic climate, return on investment has to be considered up front.
“Years ago, you could develop a technology and throw it out there, and you could get some folks to adopt it, regardless of the price point. Those days are gone. So, we truly cannot afford to bring technologies to market that don’t answer a clinical question, but also don’t answer the economic question as well.”
Guiding the Scope
In a blog post highlighting the Cad Eye technology, Borrelli also posited features of endoscopic AI that go beyond image recognition, such as data that could help a clinician guide the scope through a colon’s acute turns in the body.
Guideline co-author Shung said that feature could be a great boon to practitioners. Currently, he said, scopes lack the number and types of sensors that would allow real-time modifications to guiding the scope through the colon.
“That would be fantastic,” he said. “The difficulty there is you have access to data in images only, and to then identify depth perception with that is very difficult. You need LiDAR. At the end of the scope, you need additional sensors to help you map what is happening.
“That’s not happening now. People aren’t adding more sensors, though that makes the most sense in my mind, because without better sensors, all you have is light and a series of two-dimensional images. What you need are additional inputs to be able to do things you want to do.”
Other Specialties and AI
The use of AI in medical care is extends beyond gastroenterology. AI has made quick inroads in specialties including neurology (notably stroke care) and dermatology. Yet firm guidelines for clinical use of AI are still either in their infancy or missing entirely.
For instance, the American Heart Association, in a February 2024 scientific statement in its flagship journal Circulation, called for concerted effort in developing AI and cardiac care: “Artificial intelligence may transform cardiovascular medicine,” the AHA said in announcing the statement. “For now, though, many challenges remain, and few AI tools have been proven to improve care.”
In a recent issue of Practical Dermatology, Dr. Daniel Schlessinger, M.D., also updated readers on guideline efforts by the American Academy of Dermatology (AAD).
“The AAD’s Augmented Intelligence committee, of which I am a member, is tasked with creating dermatology-specific AI guidelines and standards,” he wrote. “Thanks in part to its work, there are now standards for how AI tools are reported, how dermatology AI algorithms are created, and how industry should engage with the AAD regarding AI.”
Schlessinger concluded by mentioning the first principle of care, one that is applicable to any specialty, and one that should be kept in mind with any new practice or technology: “More work is still needed to ensure these tools are effective, fair, and true to the same rule that most of us, donning our white coats, recited to begin the practice of medicine: ‘First, do no harm.’”
Gregory Goth is an Oakville, CT-based writer who specializes in science and technology.
