The latest evidence on AI in colorectal cancer detection supports one claim much more strongly than the one often heard in procurement meetings: computer-aided detection during colonoscopy reliably increases adenoma detection, but it has not yet shown, in full peer-reviewed evidence, that it reduces interval colorectal cancer.
That distinction is not semantic. Adenoma detection rate is a clinically meaningful surrogate, and CADe has a better surrogate-endpoint record than many medical-AI categories. But a hospital committee buying a system, training staff, maintaining software, managing alerts, and absorbing downstream pathology volume is entitled to ask what patient-important outcome has actually been demonstrated.
This appraisal treats randomized-trial meta-analyses as the main evidence base, the 2025 AGA living guideline as the interpretive anchor, and large real-world observational data as hypothesis-generating rather than dispositive. Regulatory clearance and clinical effectiveness are kept separate: market authorization can say a device may be used; it does not, by itself, prove fewer cancers. This article is procurement and research support, not clinical guidance.
What CADe Has Actually Proven
The strongest part of the CADe case is adenoma detection. Makar and colleagues’ 2024 systematic review and meta-analysis in Gastrointestinal Endoscopy pooled 28 randomized controlled trials with 23,861 participants and found that CADe increased adenoma detection rate with a relative risk of 1.20, with a 95% confidence interval from 1.14 to 1.27 and moderate heterogeneity.[1] That is not a fragile one-study signal.
The Annals of Internal Medicine meta-analysis by Soleymanjahi, Shung, and colleagues reached the same general conclusion across a larger RCT evidence set: 43 randomized trials and roughly 35,000 patients. Its subgroup analyses found that the ADR benefit persisted across endoscopist experience levels, which matters because one common procurement hope is that CADe may act less like a specialist replacement and more like a second observer with consistent vigilance.[2]
Qutob and colleagues then brought the question into a GRADE frame. Their 2025 review included 38 randomized controlled trials and judged the evidence for ADR improvement as moderate-to-high certainty.[3] That certainty judgment is important because it moves the finding beyond the familiar AI problem of impressive demos and thin validation.
| Evidence layer | What it supports | What it does not settle |
|---|---|---|
| RCT meta-analyses | CADe increases adenoma detection rate and reduces miss rates across pooled trial evidence | Whether those additional findings reduce interval colorectal cancer |
| GRADE synthesis | The ADR benefit is not merely speculative or dependent on one trial | Whether the benefit is large enough, durable enough, and well-targeted enough for routine adoption in every setting |
| AGA living guideline | CADe is promising for polyp detection as a surrogate endpoint | The guideline body did not conclude that CADe prevents colorectal cancer |
| Large real-world observational data | Interval-cancer reduction is a plausible hypothesis worth testing | Causal effect, device effect, and confounding control |
The procurement implication is straightforward: a committee can fairly describe CADe as evidence-supported for improving adenoma detection during colonoscopy. It should not describe the same evidence as proof that the system prevents colorectal cancer.
The Surrogate Endpoint Is Stronger Than the Outcome Claim
Adenoma detection rate deserves respect. It is not a vanity metric, and the clinical logic behind finding and removing more premalignant lesions is real. The problem begins when the slide changes from “higher ADR” to “fewer colorectal cancers” without showing the evidentiary bridge.
The AGA’s March 2025 living guideline is useful precisely because it did not flatten that distinction. The guideline made no recommendation for or against CADe overall and stated that current evidence was not yet certain to show that AI technology in colonoscopy further prevents colorectal cancer beyond the human eye. It did, however, recognize a conditional role for polyp detection as a surrogate endpoint.[4]
That is the right level of tension. CADe can be promising and insufficiently proven for cancer prevention at the same time. A “no recommendation” guideline stance is not a finding of futility; it is a refusal to promote a surrogate endpoint into a patient-outcome conclusion before the evidence is there.
Readers who want the guideline dispute in more detail can use the related ClinicalMind appraisal on AI-assisted colonoscopy and the AGA guideline stalemate. The short version for this article is simpler: the guideline body saw enough evidence to take detection seriously, not enough to certify cancer prevention.
The Sessile Serrated Lesion Gap Matters
The most uncomfortable mechanistic gap is sessile serrated lesion detection. If CADe mostly improves detection of the lesions already easiest to notice or least responsible for interval cancer, the cancer-prevention argument weakens even when overall ADR improves.
The pooled estimate for sessile serrated lesion detection in the research brief is a relative risk of 1.10, with a 95% confidence interval from 0.93 to 1.30, which is not statistically significant. That matters because sessile serrated lesions account for up to 30% of interval colorectal cancers.[4]
Adenoma detection can still be valuable without a clear SSL benefit. But a committee evaluating AI in colorectal cancer detection should not let a global ADR number do all the work. The missing link sits exactly where many interval cancers arise.
The Interval-Cancer Signal Is Provocative, Not Yet Decisive
The newest reason to keep watching the field is the DDW 2026 TriNetX analysis reported by Butt and colleagues. In more than 1.5 million patients, AI-assisted colonoscopy was linked to a 47% relative reduction in interval colorectal cancer.[5]
That result is large enough to make any evidence reviewer lean forward. It is also exactly the kind of result that should not be treated as a procurement verdict before full peer-reviewed publication and causal scrutiny. The study is observational. It can show a temporal association, not causation, and the investigator explicitly cautioned against overinterpreting it that way.[5]
Confounding is not an abstract nuisance here. Sites adopting CADe may differ from non-adopting sites in withdrawal technique, quality programs, bowel preparation standards, pathology follow-up, endoscopist performance monitoring, patient mix, and investment in colorectal screening infrastructure. Those factors live directly in the pathway between colonoscopy and interval cancer.
The DDW finding should raise the priority of interval-cancer studies. It should not erase the current evidentiary hierarchy. Through mid-2026, the peer-reviewed evidence base still supports “more adenomas found” more firmly than “fewer cancers prevented.”
More Detection Also Means More Cleanup
False positives and clinically low-value removals are not merely workflow annoyances. They create pathology specimens, documentation, procedure time, patient anxiety, and small but real procedural risk. The downstream work is done by the endoscopy team and borne by the patient, not by the detection box.
The pooled evidence described in the research brief found a 39% increase in non-neoplastic resection, with a relative risk of 1.39 and a 95% confidence interval from 1.24 to 1.56.[1] That does not negate the ADR benefit, but it changes the numerator in any value calculation. A system that finds more adenomas while also increasing removal of non-neoplastic lesions is not a frictionless quality upgrade.
Withdrawal time appears less concerning operationally. The reported prolongation is about 9 seconds, which is clinically modest, though still methodologically relevant because even small behavior changes during withdrawal can complicate claims about whether the algorithm or the altered procedure produced the detection gain.[1]
Cost-Effectiveness Depends on the Unproven Translation
Cost-effectiveness should not be used as an escape hatch from the endpoint problem. Areia and colleagues’ microsimulation modeling in The Lancet Digital Health suggested a threshold of about $19 per procedure at current pricing assumptions.[6] That figure is useful for budget sensitivity discussions, but it inherits assumptions about how ADR improvement translates into colorectal cancer reduction.
For procurement, this means cost modeling is conditional rather than conclusive. If additional adenomas removed by CADe meaningfully reduce interval cancer, the model improves. If the incremental findings are skewed toward lower-risk lesions, if SSL detection remains weak, or if non-neoplastic resection rises without proportional benefit, the value case changes.
Departments with low baseline ADR may still see a practical quality-improvement argument. But that is a local performance-management argument, not proof that routine adoption across all colonoscopy settings is already justified as a colorectal-cancer-prevention intervention.
What This Means for a Procurement Committee
A defensible CADe proposal should state the outcome hierarchy plainly. The system is being purchased to improve polyp and adenoma detection during colonoscopy. Any claim about reducing interval colorectal cancer should be labeled as unproven or under active investigation unless supported by future peer-reviewed causal evidence.
- Ask vendors to separate regulatory clearance, RCT detection evidence, real-world implementation data, and modeled economic value.
- Require local monitoring of ADR, sessile serrated lesion detection, non-neoplastic resection, pathology volume, withdrawal behavior, and alert burden.
- Avoid device-to-device superiority claims unless supported by head-to-head evidence; pooled CADe evidence does not automatically rank individual systems.
- Predefine whether the deployment is a quality-improvement pilot, a research implementation, or a routine standard-of-care purchase.
- Treat interval colorectal cancer reduction as a future validation endpoint, not as a benefit already proven by ADR improvement.
For device-specific due diligence, the category-level evidence here should be paired with narrower appraisals. ClinicalMind’s GI Genius evidence review examines the first FDA-cleared CADe system in procurement terms, while the RCT-focused CADe appraisal drills further into randomized trial design and detection endpoints.
Evidence Scorecard Through Mid-2026
| Question | Current answer | Procurement stance |
|---|---|---|
| Does CADe improve adenoma detection? | Yes. Multiple RCT meta-analyses show a consistent ADR increase, including a pooled RR of 1.20 in Makar 2024. | Defensible evidence-supported claim. |
| Does CADe reduce adenoma miss rate? | The research brief reports an approximately 55% reduction in adenoma miss rate. | Supportive detection evidence, still surrogate-level. |
| Does CADe improve sessile serrated lesion detection? | Not clearly. The pooled RR of 1.10 with a 95% CI of 0.93 to 1.30 is not statistically significant. | Major mechanistic uncertainty. |
| Does CADe reduce interval colorectal cancer? | Not yet demonstrated in full peer-reviewed causal evidence. DDW 2026 TriNetX data are provocative but observational. | Do not treat as proven. |
| Does CADe increase low-value intervention? | Yes, pooled evidence reports increased non-neoplastic resection. | Include pathology and workflow burden in value assessment. |
| Is CADe cost-effective? | Potentially, depending on modeling assumptions and local pricing. | Conditional, not a stand-alone justification. |
| Should routine adoption be framed as cancer prevention? | No, not through mid-2026. | Frame as detection support or monitored quality-improvement deployment. |
The cleanest verdict is also the most defensible one: CADe has a real evidence base for improving adenoma detection, but routine adoption as a colorectal-cancer-prevention intervention remains insufficiently proven through mid-2026.
References
- Computer-aided detection during colonoscopy: a systematic review and meta-analysis of randomized controlled trials, Gastrointestinal Endoscopy, 2024.
- Computer-Aided Detection During Colonoscopy: A Systematic Review and Meta-analysis, Annals of Internal Medicine, 2024.
- The Use of Artificial Intelligence for Detection of Colorectal Adenomas During Colonoscopy: A Systematic Review and Meta-Analysis, Journal of Gastrointestinal Cancer, 2025.
- New guideline: AI technology in colonoscopy not yet certain to further prevent colorectal cancer beyond the human eye, American Gastroenterological Association, March 2025.
- AI-assisted colonoscopy linked to higher adenoma detection, GI & Hepatology News, May 2026.
- Cost-effectiveness of artificial intelligence for screening colonoscopy: a modelling study, The Lancet Digital Health.