Vscan Air

The artificial intelligence here is severable from the product in a way it is not on any other record in this cluster, and the vendor's own documentation says so. Support material states plainly that Vscan Air artificial intelligence tools are optional and can be enabled if purchased, that cardiac guidance appears in the tools menu only if bought, that it is supported on the SL configuration only, and that it is unavailable in guest mode. A buyer can own this device, scan all day and never touch a model.

What remains without the software is a competent wireless handheld with B mode, M mode, colour Doppler and pulsed wave Doppler, all shipping with the probe and working with no subscription. That is a real ultrasound system on its own terms.

When the software is enabled it does substantial work. Caption Guidance issues turn by turn instruction toward standard echocardiographic views and captures automatically once diagnostic quality is reached, and AutoEF returns an ejection fraction from up to three views without manual border tracing. This is a genuinely capable pair of models and it sits at the bottom of the C band rather than the top, because it is a purchased option on one of two configurations in some of the markets the device is sold in. Ranked against Butterfly Network, Clarius and EchoNous, whose models ship with the device, this is the least central artificial intelligence of the four.

The oversight boundary here is set by a regulator rather than by the vendor, and that is worth more than a policy statement. The De Novo authorisation covers acquisition of echocardiograms for limited diagnostic use, and the limitation is part of the authorised indication rather than a caveat added afterwards. A device authorised for limited diagnostic use is one whose output is expected to be read by someone qualified to read it, and the pivotal study was built that way, with nurses acquiring and cardiologists interpreting.

The design is more directive than anything else in this cluster. The guidance does not suggest, it instructs, giving turn by turn prescriptive movements and then capturing automatically once its own quality threshold is met. That last step is the one worth pausing on: the model decides when the image is good enough, and the operator it was designed for is the operator least equipped to disagree. AutoEF likewise returns a number without the border tracing step that would have shown a clinician what the calculation was drawn on.

Balanced against that, the design gives the user real signal. Views are labelled in the stored gallery, the ejection fraction report records which views it was calculated from, and the vendor states accuracy improves as views accumulate, which tells the operator that a single view result is weaker than a three view result.

Graded B. Nothing published sets a competency threshold, describes training requirements beyond two remote sessions bundled with the subscription, or defines the escalation path when the automated number and the clinical picture disagree.

Functional description is precise and mechanical description is absent, which is the common shape on this axis but with better than usual detail on the functional side.

The two software components are separated and defined. Cardiac guidance assists acquisition of standard echocardiographic views with real time prescriptive movement instruction and automatic capture. AutoEF calculates left ventricular ejection fraction from one or any combination of three named views, parasternal long axis, apical two chamber and apical four chamber, and the vendor states the calculation increases in accuracy as each view is acquired. That last statement is unusually concrete: it tells a clinician that a one view result is a weaker result and gives them a reason to acquire more, which is model behaviour disclosed in a form that changes what a user does.

Beyond function, nothing. No model architecture, no training set size or composition, no validation set separation and no accuracy figure published on the current platform. Most consequentially, no model versioning is described. Software updates arrive through the subscription, so the model calculating an ejection fraction after an update may not be the model that calculated one before it, and nothing states whether a stored AutoEF report records the version that produced it. For a measurement that clinicians compare serially in the same patient, silent model change is a reproducibility problem and it is unaddressed.

The supplementary material to the pivotal publication describes the training dataset, expert quality labelling and algorithm optimisation, which is genuine transparency and sits in an academic paper rather than in anything a buyer would be handed.

Model provenance is more traceable here than on most records in this index, and it is traceable through public history rather than through vendor disclosure.

The models did not originate with the manufacturer. Caption Health was founded in 2013, built the guidance and interpretation software, took it through the De Novo authorisation in February 2020 under its own name, published the pivotal trial in 2021, and was acquired by GE HealthCare in 2023, after which the software reached this device in April 2024. A buyer can therefore follow the models from their origin, through the regulatory record and the peer reviewed literature, to their current owner. That is a supply chain a diligence team can actually audit.

What is not disclosed is everything since. Nothing states whether the models have been retrained, refined or revalidated under the new owner, what corpus any retraining used, or whether third party or pretrained components sit inside them. The academic supplement describing the original training set is the only provenance document located and it predates the acquisition and the port to this hardware.

The hardware and platform supply chain is partly visible through the manufacturer disclosure statements available on the security portal, which cover third party components and record whether a software bill of materials exists. No bill of materials specific to this product was located, so the mechanism is available and the artefact was not found.

Graded C: unusually clear origin, no disclosure of the four years since.

The pivotal study is the best designed piece of evidence in this cluster. Published in JAMA Cardiology in 2021 from a registered trial, it was prospective and paired: eight nurses with no prior echocardiography experience each scanned 30 patients across two hospitals using the guidance software, and the same patients were scanned by sonographers on the same hardware without it. Five level three trained echocardiographers evaluated every acquisition independently and blind.

The results carry the claim. Acquisitions were of diagnostic quality for left ventricular size and function in 98.8 percent of patients, right ventricular size and function in 92.5 percent, and presence of pericardial effusion in 98.8 percent, with at least 92.5 percent agreement between nurse and sonographer scans on the corresponding diagnostic assessments. More than 90 percent of the enrolled patients were found to have cardiac abnormalities on their scheduled full echocardiograms, so this was not a healthy volunteer population. The primary endpoint was selected in consultation with the regulator as part of the De Novo submission, which means the trial was designed against an agreed definition of diagnostic adequacy rather than one chosen after the fact.

Two qualifications belong on the record. Several of the authors were employees of the company whose product was under test, and the company announced the publication, so the study is company associated even though the readers were blinded and independent. And the trial is now five years old, run on the predecessor hardware rather than on Vscan Air, with the software subsequently ported to this device and acquired by a new owner. No equivalent trial on the current platform was located.

Graded A because the design, the blinding, the registration, the regulator agreed endpoint and the abnormal patient population together clear a bar nothing else in this cluster reaches.

The vendor publishes a data rights clause on the product page, which is more than most do, and the clause is broad. It states that GE HealthCare may collect, prepare derivatives from and use non personal data related to its products, services and software services for purposes including training, demonstration, research and development and continuous product improvement, that it will own the property rights resulting from that activity, and that it will not sell the data or use it to identify the customer without consent.

Take the commitments first, because they are real. A commitment not to sell and not to re identify is meaningful, it is stated in the terms rather than in marketing copy, and putting the clause on the product page where a buyer will actually meet it is better behaviour than burying it in a contract annexe.

The problem is the undefined term the whole clause rests on. Non personal data is not defined anywhere located, and in imaging the boundary is exactly where the value sits: a de identified echocardiographic loop is the raw material for training an acquisition guidance model, and whether it falls inside or outside the term decides whether a customer's studies feed the next model version. No de identification standard is named, no opt out is described, and no distinction is drawn between telemetry and clinical images. Asserting ownership of derivatives while leaving the input class undefined puts the customer in the position of agreeing to something whose scope only the vendor can determine.

Graded C. Disclosure without definition is better than silence and worse than the local processing architecture EchoNous documents on the neighbouring record, where the question does not arise because the images never leave the device.

Published privacy documentation is the strongest in this cluster and the contractual question is still unanswered.

The parent operates a product security portal that supplies device privacy and security manuals and manufacturer disclosure statements for medical device security to customers. Those statements are the standardised form a hospital technology management team uses to evaluate how a device handles protected data, mapped to recognised control frameworks, and they cover ports, protocols, storage, transmission and the presence of a software bill of materials. A buyer here can obtain a structured, comparable answer on data handling before purchase. No other record in this cluster offers that.

What could not be confirmed is whether a Vscan Air specific statement is published, as opposed to the portal existing at the parent level with coverage across the wider device catalogue. That distinction matters to a buyer evaluating this product rather than the manufacturer.

The business associate question is unaddressed in anything located. It is not academic here, because the cloud storage and imaging interchange tier holds patient studies on the vendor's infrastructure, which places the vendor squarely in the business associate role for that service. No public statement offers an agreement, describes its terms or names the tier it attaches to. Enterprise vendors of this size routinely execute such agreements and this index does not credit assumed practice.

Graded B: unusually good published privacy control documentation, no published contractual posture.

This is the best published security posture in the handheld cluster and one of the better ones in the index, and it earns the grade on behaviour rather than on badges.

A product security portal supplies device privacy and security manuals, manufacturer disclosure statements for medical device security, proactive security advisories, vulnerability notifications and patch provisions. A coordinated vulnerability disclosure statement is published, applies to all commercially available products, sets out what researchers may expect and what is expected of them, and names a reporting channel. A documented secure development process governs product design, with security architecture decisions taken at the point where intended use and operating environment are defined.

The advisories are what separate this from a compliance page. The vendor has publicly participated in coordinated disclosures affecting its own ultrasound line, including one describing a command injection and path traversal weakness reachable with physical access, published with the external researchers and their employer named, and with the residual risk decision stated openly rather than left implied. Publishing an advisory that says a weakness exists in your own ultrasound devices, crediting the people who found it, and explaining why the remaining risk is acceptable is the behaviour a hospital security team can actually evaluate.

Two qualifications. The portal is operated at the manufacturer level and a Vscan Air specific disclosure statement could not be confirmed as published, so a buyer may still have to request it. And no independent audited certification such as an information security management standard was located for this product line, which is why this rests on disclosure practice rather than on third party attestation.

The regulatory position is the strongest in this cluster and the reason is the pathway rather than the volume. Caption Guidance was authorised in February 2020 through the De Novo route, which is the route taken when no predicate device exists. It was the first artificial intelligence authorised to guide the acquisition of medical images, and clearing it meant creating a device classification rather than demonstrating equivalence to something already on the market. Every guidance product that follows it does so against the special controls written for that classification.

The surrounding record is consistent. Caption Interpretation AutoEF is a separately described cleared feature. An expedited clearance followed in May 2020 to support frontline pandemic use. The parent states approximately 60 artificial intelligence enabled device authorisations, around 30 of them in ultrasound, which is context rather than evidence for this product but does describe an organisation that runs submissions as routine practice.

The disclosure discipline is good in a specific way that matters. The vendor states explicitly that the artificial intelligence is subject to government regulation and may not be available in all countries, and publishes an indications reference guide. Vendors commonly market a cleared capability globally and leave the buyer to discover the regulatory limit; naming it on the product page is the correct behaviour.

One gap keeps this short of complete. No consolidated public list of clearance and authorisation numbers tied to the Vscan Air platform specifically was located, so verifying which software version on which configuration is covered by which authorisation requires reconstruction from press material and the public device database.

This record carries the only published subgroup consistency reporting located in the handheld cluster. The pivotal trial reported that results held across body mass index, sex and race, and body mass index is the confounder that matters most in echocardiography because acoustic window quality varies systematically with body habitus. A guidance model that works on easy windows and fails on difficult ones would produce exactly the pattern of harm this axis exists to detect, and the study went looking for it and reported the answer.

The patient population strengthens it further. More than 90 percent of enrolled patients had cardiac abnormalities on their scheduled full echocardiograms, so the subgroup result was obtained on genuinely abnormal hearts rather than on a clean cohort.

Three things keep this from an A. The subgroup analysis comes from a 2021 cohort of 240 patients on predecessor hardware, and nothing published refreshes it for the current platform or for the model versions shipping under the subscription. Several authors were employees of the company whose product was tested, which does not invalidate blinded independent reads but does mean the bias analysis was performed by an interested party. And there is no model card, no statement of training population composition, no documented post market performance monitoring and no description of how a model update is revalidated against the same subgroups.

Graded B, and it is the highest grade on this axis in the cluster by a clear margin.

Nothing published addresses what happens when the artificial intelligence is wrong. No performance warranty attaches to an automated ejection fraction, no indemnity is described, no adjudication or error reporting route is offered, and no service commitment covers accuracy of a model output.

What is published cuts the other way. The product terms state that a customer's failure to maintain system software and enabling hardware compatibility for the artificial intelligence is not grounds for non payment. That is an allocation of risk, it is the only allocation located, and it runs toward the customer. It also creates a live obligation: the guidance software runs only on validated handsets, handset vendors update operating systems on their own schedule, and the customer carries both the compliance duty and the payment obligation through any resulting gap. Subscriptions renew automatically for the same term, which compounds it.

The regulatory indication provides the only genuine boundary on reliance, limiting authorised use to acquisition for limited diagnostic use. That defines what the tool may be relied on for and says nothing about who bears the consequence when a clinician relies on it correctly and it is still wrong.

Graded D on the same reading applied to EchoNous on the neighbouring record. The absence is identical: a product designed so that inexperienced operators can trust the output, and no published position on the consequences of that trust. The conventional answer, that the interpreting clinician remains responsible, is available and is not stated anywhere located.

Interoperability here is a purchase decision rather than a property of the device, and that is the finding that sets the grade.

Out of the box the probe streams to a validated phone or tablet and stores exams in the application. Imaging interchange in the standard medical format, cloud storage and the wider measurement toolset sit behind a separate optional subscription tier described by commercial analyses at roughly $430 a year. Without that tier a buyer has a scanner whose images are difficult to move into an enterprise archive, which one analysis states plainly, noting that the device is functionally limited where archive integration or compliant sharing matters. Charging separately for the ability to get images out of a device is a defensible commercial choice and it is a choice a buyer should price before signing.

The clinical record is untouched in either configuration. No health level seven interface, no fast healthcare interoperability resources support and no named electronic health record integration were located for this product. An ejection fraction produced at the bedside reaches the imaging archive if the tier is bought and the archive is wired; it does not reach the note as a discrete coded result.

The parent operates enterprise imaging and cardiology reporting products that would close much of this gap, and they are separate purchases with their own contracts rather than capabilities of Vscan Air. This record grades the product, so they do not lift the grade.

The architecture is split and only half of it is documented.

The local half is clear enough. The probe pairs wirelessly to a phone or tablet drawn from a published compatibility list, the application runs on that device, and imaging and the artificial intelligence execute against it. A clinician can scan without touching a network, and the compatibility list is a genuine operational disclosure because it tells a hospital which handsets will actually run the guidance software.

The cloud half is where the questions sit and none are answered. The optional tier provides cloud storage and image sharing, so patient studies come to rest on vendor operated infrastructure. Nothing located states the hosting region, whether residency options exist for customers under localisation rules, which subprocessors are involved, what retention applies, or how a customer exports and deletes on exit. For a globally sold device that is a substantial gap, and it is larger here than for a vendor selling into one jurisdiction.

A second geographic wrinkle compounds it. The artificial intelligence is stated as possibly unavailable depending on regulatory approval in a given country, so the same product deployed across a multinational health system will have different capability and, plausibly, different data paths in different territories, with no published map of which is which.

Graded C. The local path is well described, the cloud path holds the patient data and is not described at all.

The purchasing structure is the clearest in this cluster after Clarius, and most of the clarity comes from third parties rather than from a rate card.

Two independent commercial analyses describe the same shape and broadly agree on the figures. The probe is bought outright, around $4,999 for the dual probe SL in one account and a range of roughly $4,995 to $7,995 across configurations in another. On the SL a subscription described as required runs about $500 a year and carries software updates, advanced presets and the artificial intelligence features, along with two remote training sessions. A second optional tier at roughly $430 a year unlocks cloud storage, imaging interchange and additional measurement tools. One analysis puts five year cost at about $9,645, with recurring fees adding roughly 93 percent to the purchase price. Used units are described in a secondary market range.

What the vendor itself publishes is terms rather than prices, and those terms are real. Subscription auto renewal is stated, the artificial intelligence is documented as a purchased option, country availability is qualified explicitly, and a direct purchase channel with a basket exists on regional sites. That is materially more than a quote request with nothing behind it.

What is missing keeps this from an A. No price for the artificial intelligence option was located anywhere, which matters because it is the feature that distinguishes this product. No United States rate card was found, prices vary by geography without a published table, and the difference in subscription obligation between the CL and the SL, which one analysis describes as optional on one and mandatory on the other, is not stated in vendor material.

The hardware covers the point of care range and the intelligence covers one organ. That gap is what sets the grade.

On hardware the coverage is broad and deliberately split. The SL pairs a sector phased array with a linear transducer for cardiac, lung, vascular access and musculoskeletal work, and is aimed at cardiologists, intensivists and emergency physicians. The CL pairs a convex with a linear transducer for abdominal, obstetric and deeper imaging. Between them the named settings run across cardiology, critical care, emergency medicine, musculoskeletal, primary care and vascular access, with bladder assessment and obstetric use supported through the additional measurement tools tier. The device is verified for limited use outside professional healthcare facilities, which is an explicit and unusually honest statement of where it may and may not be taken.

The artificial intelligence coverage is much narrower and the restrictions are stacked. Caption AI is cardiac only. It runs on the SL and not the CL. It requires a tablet or phone from a validated compatibility list. It is a purchased option rather than an included capability. And the vendor states it may be unavailable depending on the regulatory position in a given country, which for a globally distributed product means the same catalogue number offers different capability in different markets.

Graded B rather than A because a buyer choosing this platform for its intelligence is choosing a cardiac tool on one of two probe configurations, while a buyer choosing it for breadth is buying a very good handheld with no models in most of the settings the device covers.