Iterative Health

For SKOUT the computer vision model is the device. It is the thing FDA cleared, and without it there is no product, only an endoscope. The clinical trial optimization line is likewise machine learning applied to endoscopic image scoring and patient identification.

Structurally sound by design: the model flags suspicious tissue in real time and the gastroenterologist decides and acts during the procedure. The AI cannot resect or diagnose, only draw attention, which is the appropriate autonomy ceiling for a detection aid and is reflected in the cleared indication.

The clinical evidence for this device is genuinely strong and is graded on the evidence axis where it belongs. Model level transparency is a different question and it is thin.

No architecture, no training corpus description, no data provenance and no performance breakdown by polyp size, morphology or anatomical location. The omission that matters most for a detection aid is the false positive rate. Every incorrect flag costs procedure time and operator attention, and the practical objection to computer aided detection in endoscopy has always been the burden of spurious alerts rather than the ability to find lesions. A trial reporting an increase in adenomas detected per procedure establishes benefit. It does not tell a purchaser how often the system will interrupt a clean withdrawal, and that number is what determines whether endoscopists keep it switched on.

On the research side there is less again. The eligibility screening models have no published validation, no sensitivity or specificity for identifying eligible patients, and no description of the features used. The published network figures on activation speed and enrollment rate measure operational performance rather than model performance, and the two should not be read as the same claim.

Ask for the false positive or per procedure alert rate for the device, and for validation of the screening model against a known eligible population.

Nothing identifies any party in the chain and no retention period, training use position or de identification standard was located, while the company now holds three distinct data estates rather than one. The first is endoscopic video, which is imagery of the interior of a patient's body captured during a procedure. The second is clinical records screened across partner practices to identify trial candidates. The third is trial data held for dozens of sponsors across several regions.

One question connects them and it is the one to press. The detection model had to be trained on endoscopic video, and the company also operates and partners with the sites that generate exactly that video, so a business holding both the research network and the device has an obvious potential feedback loop.

Nothing published describes whether procedural video or records from network sites inform model development, whether that use is disclosed to the patient, or what basis would support it. Where a company sits on both sides of a data flow, a buyer should assume the question is live until the vendor answers it.

De identification of video also deserves a specific ask rather than a general one, because faces are not the identifying feature in this material: an internal recording carries no face and can still be linked through procedure metadata, timing and the report it accompanies. Ask for retention by data type, the standard applied to video, and a plain statement on site data in model development.

A randomized controlled trial published in Gastroenterology reported a 27 percent relative increase in adenomas detected per colonoscopy, and the company states the device was evaluated in the largest US based multicenter study completed for a computer aided detection device in this category. An RCT with a published effect size on a clinically meaningful endpoint is the strongest evidence standard in this index. The vendor also reports the device does not extend procedure or withdrawal time, which addresses the practical objection to detection aids.

No retention period, no training use position and no de identification standard was located, and the company now holds three distinct data estates rather than one.

The first is endoscopic video, which is imagery of the interior of a patient's body captured during a procedure. The second is the clinical records screened across partner practices to identify trial candidates. The third is trial data held for more than forty sponsors across four regions.

One question connects them and it is the one to press. The detection model had to be trained on endoscopic video, and the company also operates and partners with the sites that generate exactly that video. A business holding both the research network and the device has an obvious potential feedback loop, and nothing published describes whether procedural video or records from network sites inform model development, whether that use is disclosed to the patient, or what the basis for it would be. Where a company sits on both sides of a data flow, a buyer should assume the question is live until the vendor answers it.

Ask for retention by data type, for the de identification standard applied to video specifically since faces are not the identifying feature in that material, and for a plain statement on whether customer or site data is used in model development.

No business associate agreement terms are published and no compliance statement was located. More importantly, this axis needs re scoping for what this company has become, because asking about a business associate agreement is only half the right question.

As a device vendor selling into provider organisations, ordinary business associate analysis applies. As the operator of a clinical research network, it does not. Screening patient records to identify people who might be eligible for a trial happens before anyone has consented to anything, and the instruments governing it are different: institutional review board oversight, the Common Rule, and either an authorisation for research use or a waiver of authorisation, rather than a services agreement. The company embeds research into partner practices and applies its models to their records, so the question a provider organisation must answer is under whose authority that screening occurs, who holds the review board approval, and whether the vendor is acting as the practice's agent or on its own account.

A second dimension is jurisdictional. The network spans Europe, India and Australia, so European data protection law, India's data protection statute and Australian privacy law reach parts of this estate. Nothing published addresses any of them, or how trial data moves between regions.

Graded C because two distinct regimes plainly apply, the research one is the larger part of the business today, and neither is addressed. Reuse this scoping for any vendor operating research sites or performing pre consent eligibility screening.

No independent attestation was located and no security page or trust centre exists. The site's own footer carries terms and conditions, a cookies policy, a privacy policy and two California notices, and nothing else. That was checked against the footer directly rather than inferred from search results, which is the reliable method when a vendor's own pages are the only place a security posture would live.

The absence is notable rather than routine because of what this company now holds. It sells a cleared device that processes live endoscopic video inside the procedure room, and it operates a research network of more than one hundred sites across North America, Europe, India and Australia, handling trial data on behalf of more than forty pharmaceutical, biotech, device and contract research organisations. Pharmaceutical sponsors run mature vendor assurance programmes and commonly require independent attestation before a supplier touches patient level trial data. The material very likely exists in a sales package. Being past early stage, and past a substantial funding round, is precisely why nothing published is worth remarking on rather than excusing.

Ask for the attestation, its type and its scope, and establish whether it covers the device, the site network operations, or both, since those are different systems serving different customers.

SKOUT holds FDA 510(k) clearance obtained in 2022 with a stated indication: adults undergoing colorectal cancer screening or surveillance. The company states the indication precisely rather than describing the product as broadly cleared, and separately identifies its clinical trial products as research services rather than regulated devices. Clear per product scoping is what this axis rewards.

No subgroup analysis, no fairness evaluation and no bias assessment was located for either product line. Three mechanisms make that worth stating rather than noting.

Detection performance in colonoscopy depends heavily on bowel preparation quality, and preparation quality is not randomly distributed. It tracks health literacy, language, time and support available to follow a preparation regimen, and comorbidity. A detection aid that performs less well on poorly prepared bowels may therefore perform least well for the patients already least well served by screening, and nothing published tests that.

Second, whether routine reliance on a detection aid changes an endoscopist's unassisted performance over time is an open question for this product class, and it is the governance question a detection vendor is best placed to answer. Nothing indicates the company monitors unassisted detection rates across its installed base.

Third, and now the larger business, eligibility screening determines who is offered access to an investigational therapy. A model that under identifies particular populations reduces their access to trials directly, and improving representation in research is part of how this network is positioned. A vendor making an access or representation claim carries a stronger duty to publish evidence for it than one selling efficiency. Ask for screening model performance broken down by the populations the diversity claim is about.

The clinical evidence for the device is genuinely strong and is graded on the evidence axis where it belongs, so this record turns on what the evidence does not cover. The omission that matters most for a detection aid is the false positive rate, and the reasoning is specific to endoscopy rather than general.

Every incorrect flag costs procedure time and operator attention, and the practical objection to computer aided detection has always been the burden of spurious alerts rather than any doubt about finding lesions. A trial reporting an increase in adenomas detected per procedure establishes benefit; it does not tell a purchaser how often the system will interrupt a clean withdrawal, and that number is what determines whether endoscopists keep it switched on.

A tool disabled by its users has an effectiveness of zero regardless of its trial result, which makes the alert burden an adoption question rather than a nuisance. Nothing published gives a per procedure alert rate, and no architecture, training corpus description, data provenance or performance breakdown by polyp size, morphology or location was located.

On the research side there is less again: the eligibility screening models have no published validation, no sensitivity or specificity, and no feature description, and the published network figures measure operational performance rather than model performance. Ask for the per procedure alert rate and screening model validation.

Real integration depth on both sides of the business, with one dependency a buyer should understand.

The device reaches the procedure room through an exclusive United States partnership with a gastroenterology documentation vendor reporting more than 3,500 customer facilities, which places it inside the system endoscopy units already document in rather than beside it. The research side embeds into partner provider organisations and applies screening models to their clinical records at scale across more than one hundred sites, which is not possible without substantive record level access.

The dependency is that the workflow integration is the distribution partner's, not this vendor's. An exclusive channel through another company's documentation product means the buyer inherits that partner's integration footprint, its supported platforms and its release cycle, and it means the commercial and technical counterparty for many buyers is the partner rather than the vendor. That is the same structure this index tracks when a product runs inside a host platform: the capability is real, but it belongs to somebody else and does not travel if the relationship ends.

Held off A because no electronic health record platforms are named on either side, no interoperability standard is stated, and the mechanism by which screening models reach practice records is not described.

Nothing is published on hosting, region, residency or architecture for either half of the business, and both halves raise the question in different ways.

The device is an unusual case. Computer aided detection running on live video during a colonoscopy has to return a result inside the operator's visual reaction time, and that latency budget effectively rules out sending video to a distant data centre and back. So a buyer can reason toward local or edge processing in the procedure room. Inference is not disclosure. The company should state plainly whether video leaves the endoscopy suite at all, whether any is retained after the procedure, and where it goes if so, because the answer determines whether this is a self contained appliance or a networked service.

The research network raises the opposite problem. It spans North America, Europe, India and Australia, so trial and screening data crosses borders by construction, and cross border transfer is a regulated act in three of those four regions. No residency commitment, no regional segregation statement and no description of where screening models run against patient records.

Ask both questions separately. They are different systems and the answers will not be the same.

No public pricing. Contact the vendor. SKOUT is distributed in the US through an exclusive partnership with Provation, which means the commercial counterparty for many buyers is the distributor rather than Iterative Health, a structural fact worth establishing early in procurement.

Precisely stated on both sides of the business, which is what this axis rewards, but the scope is now considerably wider than a reader of an older description would expect.

The device remains narrow and exactly bounded: colonoscopy, in adults undergoing colorectal cancer screening or surveillance, which is the cleared indication stated as such rather than described loosely.

The research network is not narrow. The company names four therapeutic areas explicitly, gastroenterology, hepatology, obesity and cardiology, having entered the last two during 2026, and operates more than one hundred sites across North America, Europe, India and Australia. Buyer types are named on both sides, provider organisations and trial sponsors, and the two are served by different parts of the business.

The thing to establish before purchase is that these two scopes are unrelated. The detection device covers one procedure in one specialty. The research network covers four specialties and does not involve the device at all. A provider organisation evaluating one is not evaluating the other, and the company's precision about each individually is not matched by any statement of how they relate.