GuideAI Health

The only product in market is VascularAssist Occlusion Triage (VAOT), and it is a model. The FDA summary for K260729 describes a pipeline of several deep neural networks that reads a CT angiography study and decides whether a possible occlusion should be flagged. Take the models away and nothing is left to sell.

The website now describes a wider company: Elevate, a workflow layer that routes and tracks patients after a finding; segment level stenosis and plaque characterization; carotid and coronary modules; and an executive health offer. GuideAI labels every one of these as in development or investigational and not FDA cleared. If Elevate reaches market, the company would also be selling coordination software that has value without the model.

What the device does alone is stated precisely in the K260729 summary. It runs in parallel with normal reading and sends a notification, with a compressed preview image, to the PACS or RIS when a CT angiography study shows a pattern consistent with total occlusion from the aorta to the popliteal arteries. It removes nothing from the worklist, does not mark locations on the diagnostic images, and the interpreting clinician still reads every full study. The summary also says what it may not do: no primary diagnosis, screening, disease detection or patient management, and it was not tested on children.

What is missing is the threshold. The device flags a study "when predefined criteria are met", and the operating point behind those criteria is not published. The website also sells routes the clearance does not cover, among them population screening in primary care and segment level characterization for executive health, each labeled investigational.

The approach is named. The K260729 summary describes several deep neural networks behind a gateway that listens on the PACS. The gateway de-identifies each study before sending it to the GuideAI cloud and gets a preview slice back when an occlusion is flagged. The cleared version is v1.0. The pivotal test cases were kept apart from the training cases. The website says the models were developed on curated imaging from more than 2,100 hospitals, drawn from a model development agreement with Virtual Radiologic (vRad).

Neither the filing nor the website gives the network architectures, the training set's size and makeup, or how updates are released and announced to customers. The website quotes "up to 95% patient level sensitivity". The filing reports 91.5 percent on the pivotal set, and the site does not say which study produced 95 percent.

The chain is structurally short. The model is GuideAI's own, a cleared v1.0 pipeline of deep neural networks that cannot be swapped without a new FDA review. No outside model provider sits in the path. The K260729 summary says only de-identified imaging leaves the hospital, going from the on site gateway to the GuideAI cloud. An on premises option, per the website, removes that transfer altogether.

What is not named is the cloud host and any subprocessors behind it. The training imaging came through a model development agreement with vRad, and the terms governing that data, and any customer data that might join it, are not published.

The evidence is one pivotal study, published in the K260729 summary with enough method to test it. It was retrospective, blinded and multicenter, on 317 CT angiography cases from US sites (165 with occlusion, 152 without). Ground truth was a majority vote of three US board certified radiologists, and the cases were kept apart from training. Sensitivity was 91.5 percent (95% CI 86.2 to 95.3) and specificity 90.1 percent (84.2 to 94.4), against a performance goal of 80 percent. Time from study to notification was under five minutes. The set was enriched for aortic occlusions (21 cases).

This is the company's own standalone test. No peer reviewed paper, reader study, prospective deployment or outcome evidence is published. Nothing yet shows that earlier notification shortens time to treatment or prevents limb loss. The website says academic pilots are run to produce that evidence.

One commitment is stated specifically enough to check. The K260729 summary says a gateway at the hospital takes a copy of each CT angiography study, de-identifies it, and only then sends it to the GuideAI cloud. What comes back is one axial preview slice, and the link to the patient is restored inside the hospital. On that design, identified imaging stays at the site. The website also offers an on premises deployment, and says HIPAA documentation plus model and data governance documentation are available on request.

The company does not say how long deidentified studies are kept in the cloud, whether customer studies are used to train or retrain the model, or what the deidentification removes. The models were built on imaging from vRad's hospital network. Nothing says whether a customer's studies would join that corpus. No guardrails or safety event process are published for this product.

The platform page says HIPAA documentation and a Business Associate Agreement are available, so business associate status is stated. Nothing stands behind it in public. The website has no privacy policy, terms or legal page in its navigation or footer, and the agreement is not published.

The design in the FDA summary matters here. The hospital gateway de-identifies a study before it goes to the cloud, and a preview slice is re-linked to the patient on site. Even so, the gateway software handles identified imaging inside the hospital, so an agreement is still expected.

The listed parent, GuideAI Health Corp., is a British Columbia company; the operating business, GuideAI Health, Inc., is in Boston.

No security certification, attestation or trust center is published, and the website has no security page. The platform page offers security questionnaire responses on request.

The K260729 summary says GuideAI did its own cybersecurity risk assessment and testing under FDA's 2025 premarket cybersecurity guidance. That is the manufacturer's testing, filed for device review. It is not an outside audit of how the company runs security, such as SOC 2, HITRUST or ISO 27001.

GuideAI Health Corp. is a reporting issuer in Canada. Its June 2026 prospectus on SEDAR+ carries risk factors that may cover data security.

VascularAssist Occlusion Triage v1.0 was cleared through 510(k) K260729 on 18 June 2026. It is Class II radiological computer aided triage and notification software (21 CFR 892.2080, product code QAS). The predicate is Aidoc BriefCase LVO Triage (K220709). The indication is adult CT angiography, flagging patterns consistent with possible total occlusion in the aorta, common and external iliac, common femoral, superficial femoral and popliteal arteries, for workflow prioritization only.

The indication is narrow. It covers total occlusion only, not stenosis at any grade, and the filing says the device does not detect, measure or diagnose vascular disease. Segment stenosis quantification, plaque characterization, Elevate and the carotid and coronary modules are not cleared, and the company labels them that way. CarotidAssist clearance is targeted for the first quarter of 2027. The site's "up to 95%" sensitivity does not match the filing's 91.5 percent, and its source study is not named.

The K260729 summary discloses the make up of the pivotal set with counts. By scanner maker: GE 192, Siemens 73, Canon/Toshiba 37, Philips 14, other 1, of 317. By sex: 180 male, 137 female. Occlusion sites run from the aorta to the popliteal, with 21 aortic occlusions added by enrichment. The negative cases include look alikes: non occlusive disease, post operative and device changes, out of scope vessels and poor quality scans. What comes with those groups is only the overall conclusion. No sensitivity or specificity is given by scanner maker, sex, age, occlusion site or confounder.

Peripheral arterial disease presents differently with diabetes, kidney disease, race and smoking history, and none of these is reported. Governance documentation is offered on request, not published.

GuideAI's own FDA filing publishes an error rate for the model, with method and denominator. In the K260729 pivotal study, the deep learning pipeline was tested on 317 cases kept apart from training, with 165 positive and 152 negative by three-radiologist majority vote. It missed 8.5 percent of occlusions (sensitivity 91.5 percent, 95% CI 86.2 to 95.3). It flagged 9.9 percent of negative studies (specificity 90.1 percent, 84.2 to 94.4). A buyer can hold the product to those figures.

Nothing stands behind them beyond that. There is no warranty, indemnity or remedy tied to a missed or false flag. The filing names the two failure modes: a false positive moves another patient down the queue, and a false negative leaves the case to normal reading.

The integration path is DICOM, and the K260729 summary describes it in detail. A gateway acts as a PACS listener and receives a copy of each CT angiography study. The result comes back as a DICOM Secondary Capture carrying the preview image, which the PACS attaches to the original study and uses to notify radiologists. The website says it fits existing PACS and reporting infrastructure with no parallel viewer.

No PACS, RIS or record system is named, and no live deployment is named to show the path running in production. The filing refers to a desktop application for notifications without describing it.

Two options are stated: the platform page offers cloud or on premises. The K260729 summary describes the cloud path. A gateway inside the hospital de-identifies each study and uploads it to the GuideAI cloud, and a preview slice comes back.

Where that cloud runs is not stated: no provider, region or country, and no word on tenant isolation. That matters here because the listed parent is a British Columbia company and the operating business is in Boston. The on premises option is named but not described.

No price is published, but the shape of the charge is. For hospitals and specialty practices buying directly, the homepage describes a platform fee plus a per study component. For teleradiology, the model is bundled into the service layer the hospital already buys, through vRad. Neither fee has a figure. The route in is a time boxed evaluation on the customer's own studies, with success metrics agreed first; the company says it will tell a site plainly if the economics do not work.

The validated setting is named and supported. It is adult CT angiography of the aorta through the popliteal arteries, read in hospital radiology workflows, tested on 317 cases from US sites across four scanner makers. It was not tested on children. The validation covers total occlusion only, and only for triage.

The website sells into five channels: teleradiology through vRad, academic centers, community hospitals and specialty practices, primary care population health, and executive health. Only the first three use the cleared triage function in the setting the filing tested. The primary care channel is screening, which the clearance excludes, and executive health is sold on investigational segment characterization. Both are labeled as in design or building.