We are in the homestretch here on our last session, and, you know, truly this last session, I mean, the moderators, John Vargo and Klaus Mergener, I mean, it's no mistake we've put them together. I call this the presidential session, and the reason being that John Vargo, who's just our immediate past president of the ASGE, I think AI and endoscopy and GI sort of was the seeds were laid when, you know, John, during his presidency, and he started moving things along, and they've really matured under Klaus Mergener, who's the current president of the ASGE under his term, so it's my absolute pleasure to invite both John Vargo and Klaus Mergener to finish us off with the last session, which is integrating AI and GI and challenges now and for the future, so welcome, John and Klaus. Thank you, Prateek, and it's very much an honor to be here on this phenomenal program. We'll get right into this. You know, we've been blessed so far in the program just to see where AI has started its seeds and where it's gone in terms of diversity and, you know, future applications, and obviously one of the most important pieces of this is how do we get this through the regulatory system in order to have standardization and those types of things, so it's really going to be a great session today. Our first speaker is from the medical offices of gastroenterology at the Food and Drug Administration in Silver String, Maryland, and Mark is going to be taking on a topic that has really been on all of our minds, namely the streamlining of regulatory approvals. Mark, good afternoon and welcome. Good morning. My name is Mark Antonino, and I am the team lead for the gastroenterology and endoscopy devices team at the United States Food and Drug Administration. My presentation will address streamlining FDA regulatory approvals for computer-aided endoscopic artificial intelligence devices. Specifically, I will discuss how FDA regulates medical devices, considerations for designing prospective clinical studies to assess the safety and effectiveness of computer-aided endoscopic artificial intelligence devices, and how companies can interact with FDA before submitting a marketing application. FDA regulates medical devices based on risk, that is, the risk to the patient and the user, and classifies medical devices into one of three classifications. Class 1 includes devices with the lowest risk, and class 3 includes those with the greatest risk. Most class 1 devices are exempt from premarket notification. Most class 2 devices require premarket notification, also known as a 510K. Through the premarket notification process, it is demonstrated that the device to be marketed is safe and effective, that is, substantially equivalent to a legally marketed device called a predicate device. Most class 3 devices require premarket approval, or PMA for short. Currently, there are no cleared or approved computer-aided endoscopic artificial intelligence devices for polyp detection in the United States. Based on current FDA understanding of potentially low to moderate risks associated with polyp detection, computer-aided artificial intelligence devices for polyp detection in the United States may be eligible for a de novo classification. The de novo pathway allows for evaluating the safety and effectiveness for which there is no legally marketed predicate device. If a de novo is granted, the new device type is eligible to serve as a predicate for new medical devices. De novos may require clinical data or supporting literature to support a benefit-risk determination. The determination includes an assessment of the probable benefits, extent of the probable risk or harms, and additional factors such as uncertainty, patient-centric assessments, and patient-reported outcomes. Please refer to the guidance document stated on this slide for additional information and available on the FDA.gov website for download. Endoscopic computer-aided polyp detection devices may highlight a polyp by presenting a green bounding box around the polyp to the user, as illustrated on this slide. For the purposes of this talk, such a proposed device will be used to detect polyps and only draw the user's attention to the polyp. It would be the user's discretion to remove or not remove the polyp. The polyp is not characterized by the device. Device manufacturers should determine how their device will be used. FDA will consider the intended use, indications for use, labeling claims, device design, and how the device will be incorporated into clinical practice to understand the risk of the device. Devices that change clinical decision-making may warrant submission of a PMA Class III marketing submission. For example, is the device only detecting polyps? Is it only drawing the user's attention to an area of interest? Or is the device characterizing polyps? If the device is characterizing polyps, is the device stating that the polyp is an adenoma? Is the device only evaluating diminutive polyps, 5 millimeters or less? Again, for the purposes of this talk, I will focus on an intended use of computer-aided polyp detection that draws a user's attention to a polyp and leaves the decision to remove the polyp to the user. To support that the benefits of computer-aided AI devices used for the detection of colon polyps outweigh the risks, FDA recommends leveraging data from prospective clinical studies that utilize a randomized two-arm design comparing AI and non-AI. Such a design may employ co-primary study endpoints of adenomas per colonoscopy and adenomas per endoscopy, also referred to as positive predictive value. An alternative approach would be to utilize a tandem study design to account for the adenoma miss rate. FDA considers the adenoma miss rate to be a valuable metric as it is associated with a decrease in the occurrence of interval cancer and cancer deaths. A tandem study design would require two colonoscopies per patient. It would provide the adenoma per colonoscopy and the adenoma per extraction values on the initial examination and the adenoma miss rate on the second examination. An additional benefit of such a tandem design would be to control for the differences across study arms as the patient acts as their own control. In regard to selecting a patient population, the population should be limited to screening populations and surveillance populations greater than or equal to three years and symptomatic colonoscopy patients and those with screenings less than three years have substantial difference in polyp adenoma rates and heterogeneity compared to screening populations. Leveraging outside the U.S. data is possible but should be justified. To prevent bias, FDA recommends that clinicians in the study represent the skill of an average community gastroenterologist with an adenoma detection rate ranging from 25 to 40 percent, have conducted a minimum of a thousand colonoscopies, and perform an approximately equal number of AI and non-AI procedures. In regard to study endpoints, as stated earlier, FDA recommends studies include two coprimary endpoints of adenomas per colonoscopy and adenomas per endoscopy, also referred to as positive predictive value. Regarding secondary endpoints, FDA recommends considering adenoma miss rate, which would require a TAM study explained before, and adenoma detection rate. Companies should consider additional endpoints such as serrated lesions per colonoscopy and subgroup analysis should be performed for each secondary endpoint. Regarding a statistical analysis plan, studies should demonstrate that computer-aided AI devices improve detection of adenomas per colonoscopy. Statistical superiority should account for such an improvement. A non-inferiority margin for adenomas per endoscopy will allow for determination that the device does not result in unnecessary biopsy when compared to unaided endoscopy. Statistical superiority for adenomas per colonoscopy and the non-inferiority margin of adenomas per endoscopy should be predefined and sample size should be based on both primary endpoints to prevent bias. Standalone testing is an important element of algorithm assessment. Standalone testing should be included so that the algorithm is benchmarked and algorithms, future algorithms, can be compared. The labeling should clearly outline both the standalone and clinical performance. Non-clinical testing should be independent of training or validation data. It should include a full colonoscopy exam to allow for estimation of false positive rates and be inclusive of the entire endoscopy system range that the device is proposed to be used with. Lastly, results should be reported on a per lesion level as well as a frame level. FDA has a program called the Q-Submission Program that allows for companies to send in submissions called pre-submissions. Pre-submissions provide the opportunity for a submitter to obtain FDA feedback prior to an intended pre-market submission such as a de novo request. Please refer to the FDA guidance document referenced on this slide. It is available on the FDA.gov website for download and detailed information. I note that pre-submissions are encouraged and guide product development and or submission preparation. For example, a company may request FDA feedback regarding the proposed clinical trial design or bench testing methodology for their proposed device. The company provides relevant background information such as the device description, proposed indications for use, the draft clinical protocol, and specific questions. An example of a question posed to FDA could be, are the primary and secondary endpoints adequate to support the proposed indications for use? FDA provides written feedback to the company's questions. The company may elect to have a teleconference or meeting to clarify FDA feedback. FDA's responses and written feedback is usually provided within 70 days or five days prior to a scheduled meeting. In summary, based on current FDA understanding of the risk associated with polyp detection, computer-aided artificial intelligence devices for polyp detection in the United States can be eligible for a de novo designation. The de novo pathway allows for classification of novel medical devices for which there is no legally marketed predicate device. To support that the benefits of AI devices used for the detection of colon polyps outweighs the risk, FDA recommends leveraging data from prospective clinical studies. Importantly, the study design should incorporate the co-primary endpoints of adenomas per colonoscopy and adenomas per endoscopy. FDA also recommends non-clinical standalone bench testing be conducted. Lastly, please consider utilizing the pre-submission program to obtain FDA feedback before submitting a marketing application. This concludes my talk. Thank you.

AI integration in GI practices

Gastrointestinal practices and AI

Regulatory approval process for AI devices

Medical device classification and risk

FDA's pre-submission program