Viz.ai

Detection models sit under everything, with more than 50 FDA cleared algorithms analyzing CT, ECG, and echocardiography. But the product's distinguishing move is what happens after detection: the system automatically alerts and connects the treating specialist rather than only marking a study. Both halves are essential, and neither works without the model identifying the finding first.

The regulatory category defines the boundary precisely and conservatively. The original LVO product received FDA De Novo authorization as a computer aided triage and notification platform, meaning it notifies a specialist earlier rather than rendering a diagnosis, and the FDA's own framing at authorization was that earlier notification could decrease time to treatment.

In practice the alert reaches a physician's phone with the images attached, so the human sees the underlying evidence and makes the call. For time critical stroke care, compressing notification while preserving physician judgment is the correct design.

The original authorization came with disclosed performance: an AUC of 0.91 in a 300 patient study, 90 percent sensitivity and specificity, and a median scan to notification time under six minutes. Subsequent clearances specify indications concretely, such as quantifying intracranial hyperdensities, lateral ventricles, and midline shift on non contrast CT. Architecture is described as deep learning without further detail, and per algorithm performance across the 50 plus cleared products is not published in aggregate.

Nothing identifies any party in the chain: no model or model family, no hosting arrangement and no sub processor list was located in two passes, and no retention, de identification or training position was found beyond a general compliance statement. The architecture creates an exposure that is deliberate rather than incidental and should be understood as such, because it is the source of the product's value rather than a flaw in it.

The coordination workflow pushes patient imaging to physicians' mobile devices so that a specialist can see a study within minutes wherever they are, which is a materially broader surface than software resident on a hospital server: images reach personally owned handsets, over consumer networks, on devices the hospital does not manage and cannot wipe, belonging to physicians who may cover several institutions.

That is the trade a buyer is making for speed, and speed in stroke is a real clinical good, so the question is not whether to accept it but what governs it. Nothing published states what is cached on a device, for how long, what happens when a physician leaves, whether images persist after a case closes, or what device management is required. Ask all five, plus a sub processor list and the training position on submitted studies.

Among the strongest outcome evidence in imaging AI, and unusually it measures care delivery rather than detection accuracy. Peer reviewed studies in Interventional Neuroradiology and the Journal of NeuroInterventional Surgery document reduced transfer times and length of stay and improved stroke workflow in hub and spoke networks, with the company citing decreases in door to transfer, door to puncture, and door to recanalization alongside improved Modified Rankin Scores, the standard disability outcome measure.

Independent analysis has reported higher LVO detection rates with AI than without. Deployment spans over 1,700 hospitals. Moving the outcome metric from image reading to patient disability is what separates this from most of the category.

The company states HIPAA compliant image review within its coordination platform, which matters because the workflow deliberately pushes patient imaging to physicians' mobile devices, a broader exposure surface than PACS resident AI. No detailed data governance, retention, or training data disclosure was located beyond that compliance statement.

HIPAA compliant image review and care plan discussion are stated for the mobile coordination platform. No explicit business associate agreement commitment was located, which is what would move this higher given the volume of identifiable imaging moving through the system.

Converted from Not Rated. No independent attestation was located, and the prior note is right that publication would be a reasonable expectation at this scale and architecture.

No SOC 2 of either type, no HITRUST, no ISO 27001, no trust centre and no penetration testing statement was retrieved.

The combination of scale and endpoint model is what makes this stand out rather than the absence itself. The platform is reported across roughly 1,700 hospitals and its core function is distributing patient imaging and clinical alerts to clinicians' mobile devices. That means the security perimeter extends past the vendor's infrastructure to a large population of personal phones, across many organisations, under many different device management regimes.

An attestation would not by itself resolve the endpoint question, but it is the mechanism by which a hospital's security review gets any independent view of how the vendor handles the parts it does control: key management, session handling, transmission security, access revocation and breach detection. Without one, a health system evaluating this is relying on the vendor's own description of a distributed system it cannot inspect.

Several imaging peers in this index hold ISO 27001 or a named SOC 2 and are credited accordingly, so this is achievable in the category.

Ask for the attestation, its type and period, and specifically for the mobile endpoint controls it covers.

Historically significant and currently the broadest in imaging AI. The company received the first ever FDA De Novo authorization for computer aided triage and notification software in February 2018 under DEN170073, creating the regulatory category that many competitors now clear into, followed by CE marking in Europe. It now reports more than 50 FDA 510(k) clearances spanning stroke, intracranial hemorrhage quantification, pulmonary embolism, and aortic disease. Buyers should confirm which specific modules and indications their contract covers, since clearances are per algorithm.

Converted from Not Rated. The prior analysis frames the equity question structurally rather than demographically, which is the right frame for this product and worth keeping.

No governance framework, bias evaluation, subgroup performance analysis or monitoring commitment was located.

The mechanism is allocation rather than classification. The platform detects a suspected large vessel occlusion and accelerates communication between a spoke hospital and a hub with thrombectomy capability, compressing the interval before transfer. Thrombectomy capacity is finite and time dependent, so a system that speeds some transfers is, in aggregate, deciding sequence. Whether that acceleration reaches all catchment populations evenly depends on which hospitals run the platform, and hospital technology adoption tracks institutional resourcing rather than patient need.

The detection layer carries the conventional question alongside it. Stroke imaging findings and presentation patterns vary across populations, and no performance breakdown by any patient characteristic is published.

The stakes are unusually concentrated. This is a time critical pathway where minutes map to disability outcomes, so an uneven acceleration produces uneven neurological results rather than an inconvenience.

Ask for detection performance by demographic group, and whether transfer acceleration has been analysed by originating site.

The original authorisation published a real evidence package and the portfolio has since outgrown it. That package was good: a stated area under the curve in a defined patient cohort, sensitivity and specificity both given rather than a single accuracy figure, and a median time from scan to notification, which is the operationally decisive number for this product since the entire value proposition is shortening the interval before a specialist sees a stroke.

Subsequent clearances also specify indications concretely, naming the quantities measured rather than describing capability, which lets a radiologist know what is and is not covered. What is missing is scale of evidence matching scale of product.

The portfolio now spans more than fifty cleared products and no per algorithm performance is published in aggregate, so a hospital buying a module released last year is relying on the reputation established by a module authorised years earlier, and the two may differ in evidence, indication and error profile. Architecture is described as deep learning without further detail, and no warranty, indemnity or remediation commitment was located.

The failure direction worth pressing is the missed activation, since a case not flagged simply proceeds at ordinary speed and no artefact records that the system should have caught it. Ask for per algorithm sensitivity and specificity on the modules you are buying, and the false negative rate.

Integration is imaging and communication rather than chart based, and the communication layer is the differentiator. The platform connects directly to hospital CT scanners via DICOM and pushes findings through a proprietary mobile application that serves as the coordination surface for neurologists, interventional radiologists, and stroke nurses across primary and comprehensive stroke centers. Spanning multi hospital networks is the harder version of this problem. No EHR integration was located.

Converted from Not Rated. No hosting, region or tenancy terms were located, and the architecture makes the gap more consequential than for a workstation bound product.

The design requires imaging to leave the hospital. Alerts and image views reach clinicians on personal mobile devices, which is the point: a neurointerventionalist at home at three in the morning sees the study without going in. That is a genuine clinical advance and it necessarily means patient imaging traverses public networks to endpoints the hospital does not own or manage.

Nothing published addresses what follows from that. No cloud provider, no region, no statement of whether studies rest in vendor infrastructure or stream transiently, no retention period, no tenancy model and no subprocessor list. Nor is the endpoint side described: what is cached on the device, whether it persists after the case, what happens when a clinician's phone is lost, and how access is revoked when someone leaves the organisation.

Deployment scale, reported across roughly 1,700 hospitals, makes these operational questions rather than theoretical ones.

Ask where studies rest, what persists on the mobile endpoint, and how device level access is revoked.

No published pricing, and third party analysis describes enterprise only contract based licensing that varies by hospital size, module set, and term, typically structured either per indication or as a platform wide subscription with multi year commitments.

One genuinely useful commercial fact does exist publicly: reimbursement through the Medicare New Technology Add on Payment pathway has applied in stroke care, which materially changes the return calculation and is worth a buyer investigating for their own case mix.

Broad and deep simultaneously. Clinical coverage spans neurovascular disease including LVO stroke and intracranial hemorrhage, pulmonary embolism, aortic disease, and cardiac indications, across CT, ECG, and echocardiography. Setting coverage is the more distinctive dimension: the platform is built for hub and spoke networks connecting primary stroke centers to comprehensive centers, which is a system level rather than department level deployment, across more than 1,700 hospitals globally.