Artificial intelligence (AI) has emerged as a tool with great promise in gastroenterology. Machine learning is already changing the detection, characterization and classification of colonic lesions. It may soon also aid in the risk stratification, prediction, diagnosis and treatment of various other GI conditions and disorders, such as inflammatory bowel disease, Barrett’s esophagus, gastric cancer and non-alcoholic steatohepatitis. The technology also has applications in practice management that will affect patient care and physician finances, including the automation of tasks during endoscopy, scheduling, reporting, patient chart review and patient communication. Our panel this month discusses how much AI has already started to change the field, and how much it will do so in the future. Thanks for reading, and thanks to all of the contributors for their efforts!
—Colleen Hutchinson
Colleen Hutchinson is a medical communications consultant at CMH Media, based in Philadelphia. She can be reached at colleen@cmhadvisors.com.
 | Professor of medicine at the University of Kansas Medical Center, in Kansas City, Kan. |
 | Professor of medicine at Mayo Clinic, in Jacksonville, Fla. |
 | Associate professor of medicine and director of clinical care and quality in the Division of Gastroenterology at NYU Langone Health, in New York City |
 | Swedish Medical Group in Issaquah, Wash. |
What is the current state of computer vision in GI?
Dr. Sharma: Computer vision (CV) is rapidly evolving in the endoscopy space. Recent studies have assessed the role of CV in lesion detection and characterization in several GI disorders, including Barrett’s esophagus, gastric cancer, colon polyps, video capsule endoscopy, pancreatic lesions and others. We expect a slew of several algorithms to hit the journals very soon. The algorithms are here, but the bigger questions are how do we ensure that they work in the clinical setting and how do we ensure that these algorithms are well validated and are not biased. That’s the next important step. Implementation and integration after appropriate oversight and regulation to clinical practice is key.
Dr. Wallace: CV for endoscopy is moving forward extremely rapidly given the natural fit between endoscopic imaging and AI. The key applications are in lesion detection: colon polyps, small bowel lesions on capsule and Barrett’s esophagus. Lesion classification—for example, adenoma versus hyperplastic—is also moving forward rapidly, although it faces more regulatory hurdles. Several systems have recently received approval in Europe and Japan, and are being used in practice there. At least one system is already in clinical trials under FDA guidance in the United States.
Dr. Parasa: CV is a technique to help computers understand the content of digital images such as photographs and videos. This means CV algorithms can be developed and validated for endoscopic images to aid in detection and characterization of lesions.
Several algorithms have already been developed for the detection and characterization of colon polyps, and a recent randomized controlled trial comparing CV plus an endoscopist versus endoscopist only showed an improvement in the adenoma detection rate (ADR).
The concept is the same for any lesion detection/characterization: Once a model is trained on a diverse image data set and validated in an external data set, that algorithm can be applied to several use cases. Such algorithms exist for several lesions that we diagnose and treat during endoscopy and as gastroenterologists.
How will AI improve GI care today and in the future?
Dr. Parasa: I will start off with a well-known quote: “AI will not replace clinicians, but clinicians who do not use AI might be replaced.” AI will augment a clinician’s ability to better predict, diagnose and prognosticate several GI disorders and cancers. For example, several algorithms exist to help triage patients presenting to the emergency department with GI bleeding, regarding which patients will need ICU level of care versus not. Another impact will be personalized medicine in IBD. The latest technological advances in health care and biology have paved the way for systematic collection of complex data, usually referred to as multiomic data, and these data sets provide information on DNA/RNA variation, protein abundance, gut microbiota, metabolites and gene expression profiles that are then analyzed along with clinical information using complex machine learning algorithms to risk-stratify patients and predict response to a drug or prognosis. Such multiomics data are being used for drug development in IBD, as well.
AI also could affect the business end of gastroenterology and endoscopy. It can help with scheduling, reporting and communication, endoscopy operations, and workflow and billing.
Dr. Gross: AI will have a global impact on gastroenterology, including helping the endoscopist in the procedure room to identify and then interpret mucosal lesions in the luminal GI tract. The endoscopy procedure could be fully automated by creating the procedure report as the endoscopist is doing the procedure with voice command. Big data from electronic health records will be analyzed by AI platforms to allow for precision medicine to more accurately diagnose and risk-stratify our patients. This will allow physicians and health systems to offer high quality at a lower cost as we continue to move toward value-based care models.
Dr. Sharma: Other potential areas include use of machine learning algorithms for risk prediction models for GI disorders, such as screening/surveillance for colon cancer and with GI bleeding. AI also can have an important role in the business of endoscopy and supply chain management. Natural language processing can help extract relevant data from free-form charts to aid in calculation of quality metrics, patient charting and review of clinic visits.
What are the barriers to widespread adoption of these technologies?
Dr. Parasa: Barriers include adjudication and validation of algorithms; regulatory issues; clarity on medicolegal responsibility; data variability and ability to generalize; disease diversity within and between individuals; and workflow integration. AI algorithms must be extensively validated in diverse data sets to prevent bias while clarifying the medicolegal responsibilities and promote adaptive intelligence. Adaptive intelligence is a marriage between intelligent data-driven solutions and human clinical expertise.
Dr. Gross: The biggest challenge will be the cost of implementing these systems into clinical practice. Health systems will likely have an easier time, since they have more capital funding to support this integration. However, the private practice physician’s adoption of AI technology will likely be slower, since the margins are tighter as reimbursement in medicine changes and often declines.
Dr. Sharma: Validation must ensure that biases are addressed. Although there might be a profusion of algorithms, there is lack of sufficient data to train algorithms and test them in a diverse data set. And we can’t forget about the legal issues: Who is responsible if the algorithm fails, the clinician or the software developer?
Dr. Wallace: The main barrier is regulatory approval, although the pathway for at least common applications has been made clearer by the FDA. For example, the FDA has made it clear it wants colonoscopy systems to increase adenoma detection without significantly increasing false-positive rate.
Besides the use of CV, what other AI applications do you foresee in GI?
Dr. Gross: AI integration during a procedure to improve lesion detection is the tip of the iceberg. The role of AI for lesion detection will expand beyond colon polyp detection, but all aid in detection for other conditions: Barrett’s esophagus, gastric cancer and IBD. AI will also help better risk-stratify GI disease—for instance, using data from the electronic health record and analyzing clinical history, laboratory studies and imaging studies to help better determine who should be admitted for a GI bleed and who can safely go home.
Dr. Wallace: Other areas for lesion detection and classification include IBD inflammation and dysplasia; early gastric cancer; early squamous cell neoplasia; and classification of polyps into noninvasive, superficially invasive and deeply invasive. AI also has strong potential to replace mundane tasks of endoscopy such as the grading of bowel prep and the generation of reports, including billing codes. These applications likely will gain rapid acceptance given the obvious gains in efficiency.
Dr. Sharma: I foresee dictation of endoscopy notes being made easier and seamless, as well as voice recognition software to type notes, pulling up follow-up notes, images/radiology and more. AI can be applied to motility tracings, 24-hour pH monitoring, wireless capsule endoscopy, and the spectrum of wireless devices used in and around the GI lab and clinic.
Can AI affect the quality of care and can it be used to improve quality metrics used in GI?
Dr. Wallace: I think a clear answer is yes. AI can replace our most common metrics with existing technology—ADR, withdrawal time, rectal retroflexion, the Boston Bowel Preparation Scale, and novel measures such as speed of withdrawal, extent of washing and inspection of the surface.
Dr. Parasa: AI can be your big brother to monitor the quality of endoscopy both live in the endoscopy suite and to help extract information to aid in data gathering and automated quality metrics reporting. Live quality metrics in endoscopy could be algorithms that might indicate quality of bowel preparation, percentage of the mucosa visualized and data regarding the percentage of blind spots missed during endoscopy. Such algorithms can serve to provide live feedback and hence improve the quality of endoscopy at an individual level.
Dr. Gross: AI will be a key tool for personalized medicine by analyzing a large number of data points to better risk-stratify patients. For instance, a patient is newly diagnosed with IBD. AI platforms could analyze clinical, laboratory and genomic data to guide the clinician on which drug will lead to the greatest chance of response versus the more traditional approach. During colonoscopy, AI could alert the physician to inspect an area of the colon that was not well visualized or highlight an area where a polyp is located.
What collaborative efforts will be required to move AI in GI?
Dr. Parasa: AI applications in GI have focused on narrow tasks and are far different from the reality of a gastroenterologist’s overall job. To be useful, these AI solutions will have to be integrated into more comprehensive solutions. Although we are starting to remove barriers for integrating image analysis and machine learning into GI practice, many hurdles still exist. Overcoming them will require significant effort and collaboration among academia, health systems, insurance, technology experts, industry partners, regulatory bodies and professional societies.
Dr. Sharma: We need a GI AI ecosystem that consists of regulatory bodies, research entities, GI societies, academia, technology industry and venture capitalists. This was a point highlighted in the first global GI-AI meeting held in Washington, D.C., a few months ago [see article, December 2019, page 1].
Dr. Wallace: A critical need is the development of a large, annotated image repository (such as EndoNet for laparoscopy) where multiple platforms can test and improve their systems. The American Society for Gastrointestinal Endoscopy is taking the lead on developing such a system, as well as defining key parameters such as whether to use still versus video images, image transformation versus raw images, and key outcomes (per patient, per lesion, per image). This will require collaboration between national societies, end users, software providers and computer scientists. Such an effort was launched in 2019 at the first Global Summit for Artificial Intelligence in Gastroenterology, with the second annual summit planned for fall 2020.
Do you see any potential hazards of this technology in health care and GI in particular?
Dr. Parasa: AI should be viewed as a tool to augment a physician’s performance. Many AI algorithms are already being advertised as providing diagnosis with boastful comments that they are more robust than physicians.
As all of us know, medicine has several nuances and the current AI algorithms are not sophisticated enough to function at a human level. They are more suited for narrow AI tasks, like detecting polyps. Given the hype of AI in public media, patients and the general public might be misinformed. Clear guidelines and recommendations from physicians and GI societies are key to help mitigate this problem. AI is a great tool, and we should use it carefully and wisely.
Dr. Gross: AI technology should be viewed as a tool to help the physician. Physicians should not solely accept the AI recommendation, but use the information in their clinical decision making. AI is a potential value add, but not a physician substitute.
Dr. Wallace: I would suggest that we use this technology to move medicine back to its core values—the care of the patient—and not allow technology (as was the case with electronic health records) to increase the separation between doctor and patient. If we apply this carefully, we can return physicians to do what humans do best: combine intelligence with empathy. When AI is best applied, it will replace/reduce mundane technical/administrative tasks, not add to them.