At St Mark's Frontiers 2026 in London, in the session "Future visions of colorectal care", Prof David Tate (UZ Gent) asked what a good endoscopy unit looks like when every procedure becomes data. The talk follows one patient through the four parts of a unit described by ESGE and UEG: before the scope, in the room, learning, and after. At each stop it places one digital tool, always as an assist to the physician. The central argument is that detection is not the same as recording. CADe raises adenoma detection, but it saves nothing, benchmarks nothing and tells no one how they are performing. In the talk's analogy, it is a car with headlights but no dashboard. Recording comes first, and the rest of the talk shows what becomes possible once it does: bleeding that can be counted, competence that can be assessed from video, ergonomics read from the room camera, and feedback that reaches each endoscopist. The talk closes on an open question for the field: if every procedure were recorded, who should see the numbers?
The headlights analogy in one self-contained passage: CADe lets you see the road but gives no dashboard, followed by the ACCEPT finding that ADR fell once the AI was switched off. It states the talk's central argument without needing prior context.
At St Mark's Frontiers 2026 in London, in the session "Future visions of colorectal care", Prof David Tate (UZ Gent) asked what a good endoscopy unit looks like when every procedure becomes data. The talk follows one patient through the four parts of a unit described by ESGE and UEG: before the scope, in the room, learning, and after. At each stop it places one digital tool, always as an assist to the physician.
The central argument is that detection is not the same as recording. CADe raises adenoma detection, but it saves nothing, benchmarks nothing and tells no one how they are performing. In the talk's analogy, it is a car with headlights but no dashboard. Recording comes first, and the rest of the talk shows what becomes possible once it does: bleeding that can be counted, competence that can be assessed from video, ergonomics read from the room camera, and feedback that reaches each endoscopist.
The talk closes on an open question for the field: if every procedure were recorded, who should see the numbers?
"I liken polyp detection to kind of driving with headlights on but no dashboard, so that you can see where you're going, but you're not really knowing how much fuel you have in the tank."
The frame comes from the ESGE and UEG performance measures for endoscopy services. Leadership is the roof. Measurement is the floor. In between are the rooms a patient walks through. Each room gets one digital assist, and none of them makes the decision.
"We missed 84% of the clinically significant bleedings in those videos with manual review."
Scheduling by complexity. Lists built on fixed slots do not match the work. In a randomised trial of 1,027 patients, a 9-minute withdrawal gave an ADR of 36.6% against 27.1% for 6 minutes (Zhao, 2022). Later cases on a list also do worse. Procedure data already tell us how long each type of case takes, so a tool can propose the list and the clinician signs it off.
Talking to the patient beforehand. Across 12 randomised trials of smartphone education, adequate bowel preparation rose from 76.7% to 86.6%, and ADR followed (Li, 2022). A next step in development: the patient photographs their final output, and a model flags a likely failed preparation before they travel (Ramprasad, 2025).
Consent. The tool can explain, answer the common questions and check understanding. The individual conversation about a complex intervention stays with the endoscopist.
"In my hospital, we are so bad at obtaining data that we don't even know the KPIs of the endoscopists overall."
Headlights, but no dashboard. COLO-DETECT showed an ADR gain of 8.1 percentage points with CADe in NHS practice (56.5% vs 48.4%; Seager, 2024). But CADe records nothing. In the observational ACCEPT study, ADR in non-AI colonoscopies fell from 28.4% to 22.4% after endoscopists had been exposed to AI (Budzyń, 2025). Whether that is deskilling or reversion to the mean, no one would know without the data.
Bleeding we never wrote down. In 20 recorded ESD procedures, general-purpose AI models screened the video and three blinded experts confirmed every clinically significant bleed. The manual record had captured 28 of 180. If bleeding cannot be counted, ESD cannot be benchmarked. (Unpublished data, in preparation.)
Driver assist. The next step is an assistant that does not take the wheel. At each decision point in a polypectomy (technique, injection, snare placement, defect inspection, closure) it alerts, suggests, and records the alert and what the endoscopist did with it.
"I think we as physicians need to give industry pointers as to what they should be developing for artificial intelligence, because it's definitely not just polyp detection."
AI as a second assessor. GPAT, a polypectomy assessment tool developed through ESGE, is rarely used because nobody has time to watch the videos. When experts scored polypectomy videos and then saw the AI's score and reasons, one in five scores changed, and most of those moved toward the AI. Agreement between experts improved. The AI does not give the score. It makes assessors more consistent and points the trainer to the clips that matter. (Unpublished data, in preparation.)
Ergonomics from the room camera. A generative model rated endoscopist posture from a fixed room camera with high self-agreement. Injury from endoscopy accumulates slowly and silently, and a camera that is already in the room could warn before it happens. (Unpublished data, in preparation.)
The patient. Results explained in plain language on the day, an automated check-in at one week with a nurse calling every flag, histology explained when it arrives, and a surveillance reminder when it is actually due. One unit already sends an SMS at one week and phones those who report a problem (PACTS, 2022).
The endoscopist. In the Dutch FIT screening programme (239,217 colonoscopies), a patient with no high-risk polyps but a low-ADR endoscopist had a higher risk of post-colonoscopy colorectal cancer than a patient with high-risk polyps and a high-ADR endoscopist (HR 2.11; van Roermund, 2025). In the UK national evaluation, 63.9% of endoscopists met all minimum colonoscopy standards (Beaton, 2025). Feedback works where it matters most, in low performers (rate ratio 1.62; Bishay, 2020), and the Cochrane review of audit and feedback across medicine agrees (292 trials, OR 1.47; Ivers, 2025).
Put together, recording gives three layers: the procedure in real time, the endoscopist after every list, and the department every quarter. Each depends on the layer below having recorded something. The GIEQs Box is one attempt to fill all three from the room: capture, de-identify on site, recognise the moment, and feed back. It is in development and for research use only.
The question left with the audience: if every procedure were recorded tomorrow, who should see the numbers? The endoscopist, the unit, or the patient?
Access the full course with expert demonstrations, case studies, and clinical pearls.