Brainomix

The algorithms are the product and the regulatory record proves it, since what was cleared is the software's ability to interpret an image rather than a workflow around one.

Both platforms exist to produce a reading. Brainomix 360 Stroke generates real time interpretation of acute stroke imaging, and e-Lung quantifies features on computed tomography of the lung using a proprietary imaging biomarker. Remove the models and there is no residual product, because the delivery layer, returning results to the picture archiving system and to web and mobile applications, exists only to carry model output to a clinician.

The biomarker is worth naming as evidence of depth rather than assembly. The weighted reticulovascular score is a purpose built measure developed and published by the company rather than an off the shelf classifier applied to a new dataset, which places the intellectual work in the model itself.

A network performance dashboard exists alongside, reporting how the system is performing across sites and tracking clinician engagement, and that is operational tooling rather than intelligence. It is a thin layer on a product that is otherwise wholly inference.

Graded A without qualification. Ten regulatory clearances covering the algorithms across two disease areas is the clearest possible demonstration that the models, not the surrounding software, are what the company sells.

Decision support with the clinician reading every study, delivered fast enough to matter, with the operating point unpublished.

The position is sound and appropriate to the domain. Output is interpretation presented to a clinician, returned into the picture archiving system as annotated series alongside the original images, so the reader sees the model's finding next to the evidence for it rather than instead of it. In imaging that is the strongest available form of reviewability, because the clinician can inspect exactly what the algorithm marked. Radiologists and stroke physicians read the study regardless, so the model advises a decision that a person was always going to make.

Delivery design supports the time critical use case. Results reach web and mobile applications and email, which means a stroke team member away from a workstation is reachable, and speed is the entire point in a pathway where treatment eligibility expires.

What is absent from public material is calibration. No sensitivity, specificity or threshold is published in customer facing material for any module, so a hospital cannot state the miss rate it is accepting. Regulatory submissions will contain performance data, and that is not the same as publishing it. Nothing describes expected behaviour when the algorithm and the reading clinician disagree, or whether disagreement is captured for audit.

The network dashboard tracking clinician engagement is a genuine oversight instrument and its purpose is adoption monitoring rather than safety.

Ask for sensitivity and specificity at the deployed threshold per module, and how algorithm and reader disagreement is recorded.

A named, published proprietary biomarker and a real methods literature, with deployed operating points still absent.

The lung biomarker is the clearest disclosure. The weighted reticulovascular score is named rather than referred to as proprietary technology, and it is documented in a peer reviewed journal through a study with a pharmaceutical partner showing it stratifying patients at risk of pulmonary fibrosis. Naming a biomarker and publishing the method behind it lets an independent reader evaluate the construct rather than accepting a claim, and it is the kind of transparency that costs something because it exposes the idea.

The wider publication portfolio supports the same point, addressing accuracy, reproducibility and clinical validity across both imaging domains, so a technical reader has somewhere to go.

The delivery layer is described concretely too, including the formats results return in and the deployment options available.

What is missing is the deployed operating point. No sensitivity, specificity or threshold appears in customer facing material for any module, so the figures that determine what a hospital actually experiences are not published even though they exist in regulatory submissions and in the literature. Architecture is not described, no model card exists, and the composition of the training data is not characterised.

Graded B because a named and published biomarker with a peer reviewed methods base is genuine transparency, and rather than higher because the numbers a deploying hospital needs are not gathered anywhere a buyer can find them.

Ask for a performance summary per module at the deployed threshold, and the composition of the development data.

The architecture answers the question this axis usually has to leave open, and the components remain unnamed.

The structural answer is genuine. Inference runs on an appliance inside the hospital in the on premise and virtual deployments, and patient data is stated not to leave the building, so no third party model provider sits in the processing path and no external service receives patient imaging. For most vendors in this index that is the single largest unanswered supply chain question, and here the deployment model settles it rather than a policy statement doing so. Certification against the international information security management standard additionally implies documented supplier management controls, since that standard requires them.

What is not disclosed is the composition. No cloud or hosting provider is named for the cloud deployment option, no sub processor register was located, no machine learning framework or third party component is identified, and no hardware supplier is named for the physical appliance.

Training provenance is the more consequential gap. Models of this kind are developed on imaging datasets drawn from named institutions, often through research collaborations, and nothing published describes which datasets underpin the stroke or lung models, from which populations, or under what consent and governance. For a company with a substantial academic publication record, that information partly exists in the literature and is not gathered anywhere a buyer would find it.

Whether imaging processed at customer sites contributes to development is also unstated.

Ask for the development dataset provenance and governance, the cloud provider and sub processor register, and whether customer imaging feeds model updates.

The strongest evidence record encountered in this session, because it reports what changed for patients rather than how accurate a model was.

A study published in The Lancet Digital Health in December 2025 examined 15,377 patients whose scans were reviewed using the tool from January 2022 onward, finding clots identified more than an hour earlier. That is real world, at scale, in a journal with genuine review, and in acute stroke an hour is not an efficiency metric. Reperfusion outcomes degrade steeply with time, so earlier identification converts directly into preserved brain tissue.

Separately, real world evaluation has been associated with an increase of more than 50 percent in mechanical thrombectomy rates, and the company positions itself as the only stroke imaging tool with a demonstrated effect on treatment rates. Treatment rate is an outcome, not a performance statistic. This index has recorded vendor after vendor publishing discrimination figures while measuring nothing about whether care changed, and has written that gap into several records as the missing study. Here the missing study exists.

On the lung side, a collaboration with a pharmaceutical partner published in the American Journal of Respiratory and Critical Care Medicine showed the proprietary biomarker stratifying patients at risk of pulmonary fibrosis, which extends the evidence base into a second disease area with a different endpoint.

Deployment corroborates it: more than 70 hospitals in the English health service, alongside United States and continental European sites.

Two honest qualifications. The thrombectomy figure is an association from real world evaluation rather than a randomised comparison, so confounding by concurrent pathway improvement cannot be excluded. And the strongest claims are surfaced through company communications, though the underlying journals are independent.

Ask for the Lancet analysis directly, and for the design behind the treatment rate evaluation.

A complete and specific stewardship picture, published where a prospective customer can read it before contacting sales.

Every element a buyer would ask for is stated. Patient data is stored on the Brainomix server inside the hospital and is not transmitted externally. Data is pseudonymised. Encryption applies in transit and at rest. Retention is configured by the customer rather than fixed by the vendor. Authentication runs through the hospital's own enterprise directory using LDAP, Microsoft Active Directory or single sign on, with multi factor authentication supported throughout, so access governance stays inside the identity system the hospital already audits. Email notifications carry pseudonymised results only, which closes the most common accidental disclosure path in imaging workflows. The whole is backed by certification against the international information security management standard.

That combination is qualitatively different from a policy commitment. Data that never leaves the building cannot be breached at the vendor, and retention the customer sets cannot quietly extend.

One gap remains and it is the same one this index asks of every imaging vendor. Nothing states whether scans processed at customer sites contribute to model development, under what consent, or whether a hospital can decline. An on premise architecture makes silent collection harder, which is reassuring structurally, and it is not a published position.

Graded A because what is disclosed is complete, specific and verifiable, and because the architecture does the protective work rather than merely describing it. Ask whether customer imaging trains models, and what the cloud deployment changes.

European regime compliance is stated with a named certification behind it, and the architecture removes much of what an agreement would ordinarily need to cover.

The strongest fact is structural rather than contractual. Patient data is stored on the Brainomix server inside the hospital and is not transmitted externally, and email notifications carry pseudonymised results only. A vendor that does not receive identifiable patient data has a materially smaller obligation than one that does, and this is stated plainly in a customer facing trust centre rather than buried. Retention is configured by the customer, which places the disposal decision with the party that owns the legal duty.

European data protection compliance is claimed alongside certification against the international information security management standard, and the certification is named precisely rather than by family, which this index treats as the difference between a checkable credential and a gesture.

What is missing is the United States picture specifically. No statement addressing federal health privacy law was located, no business associate agreement template or execution requirement is published, and nothing describes how the posture differs across the three deployment models. A cloud deployment plainly creates obligations that an on premise appliance does not, and the published material does not separate them.

Given active United States expansion with operations in Chicago and ten clearances, that gap will be closed by procurement pressure regardless.

Ask for the agreement template, how obligations differ by deployment model, and what leaves the hospital in a cloud deployment.

A public trust centre, a certification named by its standard rather than by its family, and specific controls published beside both.

The trust centre exists as a standing customer facing resource with a frequently asked questions section addressing data location, transmission, encryption and retention in plain terms. A buyer can begin diligence before speaking to sales, which is the openness this axis exists to reward.

The certification is stated precisely. Certification against the international information security management standard is named by that standard, not as certified against information security generally, and this index has repeatedly recorded vendors blurring exactly that distinction by referencing a family without the number or the tier that makes a claim assessable. European data protection compliance is stated alongside it.

Controls are published rather than summarised: encryption in transit and at rest, pseudonymisation, customer configured retention, enterprise directory authentication with multi factor support, and an architecture in which patient data does not leave the hospital. Support channels are published with named offices and telephone numbers in two countries.

Three things are absent. No service organisation controls report is claimed, which United States health system procurement increasingly expects and which matters given active expansion there. No assessor, certificate number or audit period is published. And no penetration testing statement or vulnerability disclosure policy was located.

Graded A on the strength of a public trust centre and a precisely named external certification. Ask for the certificate number and scope, whether a controls report is planned for the United States market, and the vulnerability disclosure route.

Ten United States clearances across two platforms, plus European marking, which is the strongest regulatory position recorded in this index.

The count and the shape both matter. Eight clearances cover the stroke platform and two cover the lung platform, with the first stroke clearance obtained in 2023, the lung clearance in May 2024 and a further stroke clearance in April 2025. Eight clearances for one platform indicates modules cleared individually for distinct indications rather than a single authorisation stretched across a product line, which means each capability has been examined against its own intended use with its own performance data.

The contrast with the rest of this lane is the point. Vendor after vendor in this index positions itself as clinical decision support, publishes no determination and leaves a buyer to guess. Others present establishment registration, which involves no review of safety or effectiveness, as though it were a quality mark. This company submitted, was reviewed and was cleared, repeatedly, over three years.

Clearance also brings obligations that function as accountability. A cleared device carries post market surveillance duties, adverse event reporting and change control on the algorithm, so material modification is not something the vendor can make silently. That is a form of continuing oversight no unregulated competitor is subject to.

What is not public is the substance. Indications for use, the specific claim cleared for each module, and the performance data submitted are not enumerated in company material, though clearance summaries are matters of public record.

Ask for the clearance numbers and the indication for use of each module.

A stated ethical position and an extensive validation record, with no subgroup or scanner stratified performance published.

What exists is more than a slogan. The company states a focus on ethical artificial intelligence, patient safety and responsible innovation, and it is backed by a substantial portfolio of peer reviewed studies addressing accuracy, reproducibility and clinical validity. Reproducibility in particular is a fairness relevant property, since a model that produces consistent results across conditions is less likely to fail selectively.

What is absent is stratification. No performance broken down by sex, age or ethnicity was located, no fairness testing, no calibration analysis and no external algorithmic audit.

The bias axis that matters most in imaging is technical rather than demographic and is equally unaddressed. Computed tomography output varies by scanner manufacturer, model, slice thickness, reconstruction kernel and acquisition protocol, and a model trained predominantly on modern scanners can degrade on older equipment. That distribution is not random: newer scanners concentrate in better resourced hospitals, so accuracy that tracks scanner generation is accuracy that tracks institutional wealth, and in acute stroke the affected hospitals are often the peripheral units where fast interpretation is most needed because no neuroradiologist is on site.

A network performance dashboard exists and reports operational metrics rather than stratified accuracy.

Ask for performance stratified by scanner manufacturer and generation, by acquisition protocol, and by patient sex, age and ethnicity.

No contractual allocation is published, in the one record in this session where a parallel accountability regime genuinely exists.

A dedicated pass located no service level agreement, no accuracy warranty, no uptime commitment, no indemnity and no remediation position. Support channels and a support portal are published, and no service commitment behind them was found.

The mitigating structure is real and belongs on the record even though it does not move the grade. Cleared devices carry post market surveillance obligations, adverse event reporting duties and change control on the algorithm, which means material modification cannot be made silently and failures have a reporting path that does not depend on the customer's contract. That is more accountability than any unregulated competitor in this index is subject to, and a hospital has a regulator to escalate to.

What that does not resolve is the commercial question. Uptime matters acutely here because the pathway is time critical: an imaging interpretation tool unavailable during a stroke call is unavailable exactly when it was bought to help, and no availability figure or commitment is published. Nor is there any statement about what happens when the algorithm misses a large vessel occlusion that a reader, relying on it, also misses.

That second scenario is the one a hospital's counsel will raise, and automation bias makes it more than theoretical when a tool is trusted and fast.

One pre emptive note: further clearances or publications cannot move this grade. Only contractual terms or a published availability commitment will.

Ask for the availability commitment, what is warranted on performance, and the process when a missed finding is identified.

Integration is specified where it matters for imaging, with the standard, the direction and the format all named.

Results return to the picture archiving system as annotated DICOM series or as PDF reports. That single sentence contains what this axis usually has to ask for: the standard is named, the direction of flow is stated, and the format is specified. Writing back as an annotated series is the correct choice for the domain, because the finding arrives inside the study the radiologist is already opening rather than in a separate application requiring a second login during a time critical pathway. Web and mobile applications and email provide alternative delivery for clinicians away from a reading workstation.

Installation is handled rather than delegated. Company engineers configure scanner and archive connections, protocol setup and go live testing with hospital technology staff, which is a meaningful commitment for a product that must sit inside imaging infrastructure.

Identity integration is equally concrete, running through LDAP, Microsoft Active Directory or single sign on with multi factor authentication.

What is missing keeps this below the top of the axis. No electronic health record is named and no clinical messaging standard such as HL7 or FHIR is described, so whether an interpretation reaches the patient record as a discrete result, or stops at the imaging archive, is unstated. In stroke the finding drives a treatment decision documented elsewhere, and no marketplace or validated integration listing was located.

Ask whether results post to the record system, through what standard, and what integration exists beyond the imaging archive.

Three deployment options offered, two of which keep data entirely inside the hospital, with the data movement position stated explicitly rather than implied.

The options are a dedicated physical server, a virtual server running on infrastructure the hospital already owns, or a secure cloud environment, delivered as a managed appliance with vendor engineers performing installation and connection to scanners and the imaging archive.

The accompanying statement is what lifts this to the top of the axis. Patient data is stored on the Brainomix server inside the hospital and is not transmitted externally. That is a residency answer rather than a residency description: an organisation with data location obligations, whether from European data protection law, national health service requirements or its own policy, satisfies them by construction rather than by contract. Most vendors in this index answer this axis by describing present practice, which is a statement about today. Retention is configured by the customer, which extends the same principle to disposal.

For imaging that architecture is also practical rather than merely compliant, since transmitting full computed tomography studies off site introduces latency in a pathway measured in minutes.

What remains unstated is thin. No provider or region is named for the cloud option, and nothing describes whether capability, model version or support differs across the three models, which is the practical question facing a buyer choosing between them.

Ask whether capability differs by deployment model, which provider hosts the cloud option, and what the vendor can access remotely under each.

Cost is absent from every published surface, and it is the single weak point in an otherwise unusually open record.

A dedicated pass located no pricing page, no unit of charge, no range, no module tiering, no implementation fee position and no minimum commitment. Nothing indicates whether licensing runs per module, per site, per scan, per bed or per network.

The module structure makes the unit question substantive rather than routine. Company material refers to licensed modules and to training covering how to interpret results for each licensed module, which establishes that the platform is sold in parts rather than whole. A hospital licensing stroke imaging and a hospital adding lung analysis are buying different things, and no published material indicates how the second is priced against the first.

Deployment choice adds another dimension. A dedicated physical server, a virtual server on hospital infrastructure and a cloud environment carry different cost bases, and nothing states whether the option chosen affects the fee or what hardware the hospital supplies.

The contrast with the rest of the record is stark and worth stating. This vendor publishes a trust centre, names its security certification precisely, states where patient data lives and lets the customer configure retention. A company this willing to be examined on every other dimension has no obvious reason to withhold a pricing basis.

Ask for the unit of charge, how modules price relative to one another, whether deployment model affects cost, and the contract term.

Narrow by clinical scope and deep and international within it, which is the reverse of the usual pattern in this index.

Two disease areas are addressed and no more: acute stroke, and interstitial lung disease including idiopathic pulmonary fibrosis. That is a deliberate constraint rather than a limitation, and the depth within stroke is evidenced by the regulatory record, since eight separate clearances covering one platform indicates modules addressing distinct parts of the imaging pathway rather than a single algorithm.

Geographic reach is genuinely multinational and evidenced rather than asserted. More than 70 hospitals in the English health service, United States operations based in Chicago, an office in Ireland serving Europe, a published customer account from a university hospital in Czechia, and user documentation maintained in multiple languages. Satisfying several regulatory and procurement regimes at once is itself a coverage credential.

What is not evidenced is distribution by setting. Stroke care is organised into networks with comprehensive centres performing thrombectomy and smaller units transferring to them, and the value of imaging artificial intelligence differs sharply between those roles. Nothing states how the installed base divides, or how many sites run the lung module rather than stroke alone.

Nothing addresses paediatric stroke, which uses different imaging and different thresholds.

Ask how the installed base divides between comprehensive and referring centres, how many sites run the lung module, and whether paediatric use is supported.