‘Keep the guardrails on’: As AI use expands in IBD care, physicians should remain cautious
Click Here to Manage Email Alerts
Artificial intelligence has been used extensively in gastroenterology for polyp detection and is expanding into the inflammatory bowel disease space in an experimental and “research phase,” quickly making its way into clinical use.
“AI encompasses several different technologies that can engage text, images and other data with initial use cases aiming to reproduce expert opinion with the benefits of automation including standardization, reproducibility and speed,” Ryan W. Stidham, MD, MS, AGAF, associate professor of gastroenterology and internal medicine at Michigan Medicine, told Healio Gastroenterology. “As capabilities expand, artificial intelligence applications will move past recapitulating expert judgement and begin enhancing the clinician’s ability to assess patients and predict future outcomes.”
Source: Sravanthi Parasa, MD, FASGE
Stidham said research in AI goes back over 70 years with “fundamental work on cognitive theory and replication of neural networks.” Research exploring applications of AI in the IBD space only surfaced about 15 to 20 years ago.
According to Stidham, AI has specifically been applied in the past 3 to 5 years to analyze IBD clinical imaging such as endoscopy, CT and MRI, as well as histologic imaging. Conventionally, medical image interpretation requires the expertise and judgement of an experienced pathologist, radiologist or IBD specialist to grade severity and characteristics of disease.
“Now AI methods can be trained with expert knowledge for specific image interpretation tasks with remarkable results, particularly in endoscopic scoring and histologic grading,” Stidham said.
Claudia Diaconu, MD, and colleagues recently reported in Diagnostics that AI is emerging as a valuable tool in IBD diagnosis and management with the potential to analyze biological data through large-volume input for machine-learning models, cross-sectional imaging, endoscopic and histologic imaging, inflammation biomarkers, gut microbiota composition and gene expression.
They wrote that potential applications of AI technology in IBD include assessment of disease severity, diagnosis of IBD, treatment response and prognostic assessment, disease extension assessment, early detection of neoplasia and personalized medicine.
“The use of AI in gastroenterology in general, and in IBD management in particular, continues to evolve rapidly at multiple levels of patient care, enabling more than ever the possibility of ‘precision medicine,’” Diaconu and colleagues wrote.
They added: “AI technology in IBD is still in a research phase that can only be experimentally used. However, the AI applications developed so far in the IBD field already have the potential to improve the standards of patient care, starting from diagnosis to long-term therapeutic decisions and neoplasia surveillance.”
‘Transformative’ AI Coming to GI Practices
Sravanthi Parasa, MD, FASGE, a gastroenterologist at Swedish Medical Center in Seattle, told Healio Gastroenterology that one AI technology being used throughout the medical field — and currently being explored in IBD — is computer vision, or the use of image recognition and classification, such as endoscopes. As with any new technology, the domain expert or user needs to understand how it works and the “constraints under which it performs.”
“For computer vision algorithms, we need to know what the model is trained for, what can you differentiate and in what circumstances it does not differentiate,” she said.
Tyler M. Berzin, MD, FASGE, associate professor of medicine at Harvard Medical School and an advanced endoscopist at Beth Israel Deaconess Medical Center, noted that, with computer vision, “AI will augment assessment of mucosal disease activity, both in the clinical setting and for clinical trial purposes.”
Computer vision could also decrease changeability of reporting standardized disease severity scores, such as the Mayo score, he said. This may lead to “novel and more nuanced” approaches to measure and predict disease activity.
In a review published in Gastroenterology, Stidham and colleagues discussed the emergence of machine learning in IBD, which uses a “high volume” of digital medical data as well as computational methods or algorithms for pattern recognition.
They noted that machine learning shows promise for decision-making in IBD through analysis of readily available conventional information sources such as administrative claims data, diagnostic and procedural codes, and laboratory values. Machine learning in IBD is being applied to therapeutic drug monitoring of thiopurines as well as predictions of response to biologic therapies.
Seth Gross, MD, FACG, FASGE, AGAF, professor of medicine and clinical chief in the division of gastroenterology and hepatology at NYU Langone Health, told Healio Gastroenterology that although machine learning and algorithms have considerable potential, AI algorithms must be validated before approval for use in daily clinical care of IBD.
Another AI technology currently being developed in IBD is natural language processing, which Gross said could potentially improve physician efficiency in more mundane tasks, such as creating reports. For example, a physician could use natural language processing to create reports during a procedure rather than having to complete reports afterward, thus improving workflow.
One form of natural learning process being further explored in gastroenterology and the medical field in general are generative large language models. OpenAI’s Chat GPT application provides a chatbot that allows an individual to ask questions and receive replies in regular conversational language, Stidham noted.
“These chatbots are going to be integrated into many aspects of care delivery,” he said. “Whether we like it or not, it is absolutely going to be accessible to everyone, including our patients. GPT is a major step for AI that will be transformative, but also comes with some dangers.”
He explained that while chatbots do not have their own knowledge, they are programed to reproduce conversations based on millions of examples used in training. However, the ongoing chat replies are delivered as though “you are having a conversation with a person,” Stidham said, which can make the statements appear trustworthy and reliable. “Generative chatbots are very convincing.”
One kink that will need to be refined regarding natural language processing is the removal of the background noise from the physician’s staff talking during a procedure, Gross noted.
“The potential positive for [natural language processing] is that it could lead physicians to spend more time with patients and not get caught up in all the tasks that they have to do throughout their day,” he said.
Chat GPT and other natural language processing technologies will be the most “transformative” AI technologies and will be seen in GI practices within the next few years, Stidham said.
Potential to Improve Patient-Caregiver Relationship
AI is also being applied to diagnosis of disease, risk prediction and response to therapy, noted Nayantara Coelho-Prabhu, MBBS, associate professor in the division of gastroenterology and hepatology at the Mayo Clinic. This is being accomplished through models developed using modalities to include large clinical data sets, patient-reported outcomes, feedback from patient wearables and other apps, and image and video data from capsule endoscopy and colonoscopy.
“This multimodal approach has the potential to significantly impact patient care and result in better outcomes for patients,” she said.
Coelho-Prabhu added that current research assessing AI for IBD has focused on using images or video data to identify disease severity and grading as well as predict endoscopic and histologic healing.
“There are large clinical studies now that show that AI algorithms can be used to accurately grade endoscopic disease activity, thus reducing previously demonstrated variability,” she said. “This can be utilized to guide therapy and has been shown to enhance enrollment into clinical trials for patients. It also has the potential to replace central readers for trials, which will result in decreased costs and better standardization.”
“We are seeing the ability to detect polyps and also bleeding, inflammation, mucosal injury and other types of injury,” Stidham said. “We can train AI to detect any lesion that is visible to our eye.”
Coelho-Prabhu added that “AI will ideally allow physicians to focus on patient care.” AI can potentially amalgamate data from various sources, boost physician’s efficiency in documentation reporting, augment physician decision-making on risk prediction and guide therapy using clinical decision support tools.
With regards to cross-sectional imaging studies and endoscopy, “AI can standardize reporting of findings and also be trained to detect abnormalities or patterns that physicians may have missed, by utilizing large volumes of data correlated to patient outcomes,” she said.
However, Coelho-Prabhu warned, “AI cannot be utilized to replace physician interactions with patients and physicians’ abilities to synthesize various clinical and patient perspective data points into a long-term therapy plan. Given the current lack of robust validated real-world algorithms, AI cannot be utilized solely to create treatment recommendations for patients.”
According to Berzin, AI tools will soon be used to predict IBD disease prognosis and treatment response and is “already playing a critical role in drug discovery.”
“The critical strength of AI is that it has the potential to generate complex predictions — for instance, predicting the likelihood of response to various IBD medications much more quickly and more accurately than humans can,” Berzin said. “The key observation here is that all medical decisions involve some combination of prediction and judgement. AI will ‘decouple’ prediction and judgement and power much higher accuracy predictions in medicine.”
At the 2023 Crohn’s and Colitis Congress, Stidham told attendees that while AI can “reasonably” replicate expert interpretation of imaging in endoscopic and histologic scoring, read cross-sectional imaging and capsule endoscopy, and predict outcomes in IBD, clinicians must be able to maintain the ability to intervene and “take control” when AI technology is incorrect.
“Artificial intelligence will be coming to your clinic soon, so you have to get ready,” he said. “In the coming 5 years, we are going to see artificial intelligence methodologies implemented to help us make decisions, not just measure disease.”
But Stidham cautioned: “We must be careful how much we trust AI. At the least we should maintain some suspicion and verify what the AI is doing. If we use AI in the right ways, we will not only preserve but improve the patient-caregiver relationship, because AI is going to give us back a lot of time to do what we love doing.”
He added that although the future of AI in GI is bright — technologies standardizing practice, elevating the level of care for all patients and offloading a sizeable portion of labor from the clinician — implementation of AI will require in-depth attention.
The GI field has seen AI technology go from research to implementation in dozens of centers, “which is great,” Stidham said.
Creating Framework for AI Implementation
With the growing applications of AI in gastroenterology, the ASGE had developed an AI Task Force, of which Berzin, Gross and Parasa are members, to create guidance on successful implementation of the technology, specifically in GI endoscopy.
In 2020, the task force developed a position statement, which was published in Gastrointestinal Endoscopy, that focused on three key areas: priority use cases for development of AI algorithms in GI, for clinical scenarios and to streamline workflows; data science priorities, including development of image libraries and standardization of methods for storing, sharing and annotating endoscopic images and video; and research priorities, which focus on the importance of trials measuring clinically meaningful patient outcomes.
Priority clinical uses for AI algorithm development identified by the task force included colon polyp detection and diagnosis, detection of dysplasia in Barrett’s esophagus, detection of gastric cancer precursor lesions and early gastric cancer, and detection of dysplasia in IBD.
“The ASGE intends to support the successful integration of clinically relevant AI technologies into GI practice by facilitating the development of standardized approaches for endoscopy data sets and image libraries, by providing guidance for regulatory bodies and industry regarding how best to assess and monitor AI interventions in GI endoscopy, and by directly supporting important research in this field,” the task force wrote.
Parasa added: “AI has a tremendous potential to change the way we practice medicine, and it is coming very soon. However, it is important clinicians understand and try to keep the guardrails on, with frameworks for implementation, as well as evaluation of algorithms. Either you do not use the tool at all and do not reap the benefits or you overuse the tool.”
Gross recommends “clinicians should use AI as indicated with the current algorithms and that is only for screening and surveillance colonoscopy.” Currently there are no algorithms for other diseases, as the focus should be using AI in polyp detection.
AI, Physicians Must Work Together
Both Berzin and Parasa noted the increased “hype” of AI technology in and outside of medicine.
“I expect that we will watch AI tools go through every iteration of the classic Gartner hype cycle — from inflated expectations to disillusionment and back, many times over,” Berzin said. “The biggest con in AI adoption in general is that so much baggage comes along with the AI hype. It can be hard to be entirely impartial and objective when discussing and debating these technologies.”
During the Crohn’s and Colitis Congress, Stidham said he is not without reservations about AI, particularly the risks associated with allowing AI technology to go unchecked by a clinician. He noted that as the IBD population continues to grow, both in North America and around the world, gastroenterologists will need the technology to handle time-consuming tasks, including reviewing laboratory results or imaging while interacting with patients.
“As AI becomes available, and it will be pushed into our clinics, we have to keep our hands on the wheel; we cannot lose sight of who is in control,” he said. “Unsupervised AI can make some catastrophic mistakes. Even when you think that you’re paying attention to the AI, if you don’t, if you take your eyes off the road for one second, something bad can happen. We want to make sure that this does not happen for our patients as we integrate these tools.”
AI is still new in medicine and the technology is “quickly” approaching the field, Stidham said, and it is “impressive in terms of its capabilities.”
“The amazing capabilities of AI can result in enthusiasm that sometimes overcomes the actual AI performance,” he said. “Over the next decade, if not more, clinicians are going to have to supervise AI in the same way that they might supervise perhaps a trainee, medical student or a new team member so that we all can begin to understand where it is useful.”
Parasa echoed Stidham’s concerns, likening the technology to an assistant.
“Although I think [AI] is more assistive, driving will still be with the physician,” she said. “I do not think that at least in the near future, AI will be the sole decision maker; we will need well vetted systems.”
With no shortage of clinical prediction tools that describe what physicians already know, “seasoned” clinicians will remain superior to machines in collecting information from a patient, Stidham said, and will need to supervise machines for the foreseeable future.
Physicians cannot rely on AI alone, Gross added, and must continue to practice medicine and care for patients as they normally do.
“It is more of a team approach, meaning the physician still has to be very detail-oriented and focused when they are managing their patients,” Gross said. “AI is still growing and improving, and it is not perfect. It is important that the physician and AI work together.”
- References:
- Berzin TM, et al. Gastrointest Endosc. 2020;doi: 10.1016/j.gie.2020.06.035.
- Diaconu C, et al. Diagnostics (Basel). 2023;doi:10.3390/diagnostics13040735.
- ‘Keep our hands on the wheel’: Caution, supervision urged as AI enters the IBD space. https://www.healio.com/news/gastroenterology/20230125/keep-our-hands-on-the-wheel-caution-supervision-urged-as-ai-enters-the-ibd-space. Published Jan. 27, 2023. Accessed April 20, 2023.
- Stidham RW, et al. Gastroenterology. 2022;doi:10.1053/j.gastro.2021.12.238.
- For more information:
- Tyler M. Berzin, MD, FASGE, is associate professor of medicine at Harvard Medical School and a fellowship trained advanced endoscopist at Beth Israel Deaconess Medical Center in Boston and can be reached at tberzin@bidmc.harvard.edu.
- Nayatara Coelho-Prabhu, MBBS, is associate professor in the division of gastroenterology and hepatology at the Mayo Clinic, and can be reached at coelhoprabhu.nayantara@mayo.edu.
- Seth Gross, MD, FACG, FASGE, AGAF, is professor of medicine and clinical chief in the division of gastroenterology and hepatology at NYU Langone Health and can be reached at seth.gross@nyulangone.org.
- Sravanthi Parasa, MD, FASGE, is a gastroenterologist at Swedish Medical Center in Seattle and can be reached at vaidhya209@gmail.com.
- Ryan W. Stidham, MD, MS, AGAF, is associate professor of gastroenterology and internal medicine at Michigan Medicine and can be reached at ryanstid@med.umich.edu.
Collapse
Read more about
Click Here to Manage Email Alerts