Clarius Mobile Health

A deep portfolio of cleared models sitting on top of a hardware business that would work without them.

The model count is the strongest part and it is unusual. By mid 2024 the company had cleared eight artificial intelligence models, two of which it described as first of their kind: an application for musculoskeletal ultrasound automatically identifying and measuring tendons, and a fetal biometry tool placing calipers and estimating age, weight and growth. A prostate model follows the same pattern. These are not guidance overlays, they are measurement models producing numbers a clinician acts on, and each went through its own regulatory examination.

What holds the grade is what the company sells. The scarce asset is a line of wireless scanners with image quality approaching cart based systems at a fraction of the cost, engineered down to roughly 300 grams and sealed for immersion disinfection. Remove every model and a working, differentiated ultrasound business remains, which is the test this index applies.

The commercial structure states the same conclusion plainly. The probe functions without a membership, and the models are among the things the membership unlocks. A vendor whose hardware works without the artificial intelligence has told you where the intelligence sits in the value stack.

Graded C on the same reasoning applied to Butterfly Network, and the two records should be read together. Clarius has more cleared models; Butterfly holds silicon that Clarius does not.

Ask which models are included at which membership tier, and what each was validated against.

The models measure and the clinician decides, with the same competence question every capable acquisition aid in this category raises.

The workflow is described concretely, which helps. A model activates while the clinician is scanning, highlights the relevant anatomy, places the calipers and produces measurements that appear on screen and can then be added to the patient record from the application. The clinician sees what was measured and where the calipers landed before anything is recorded, which is a meaningful form of oversight: an incorrect caliper placement is visible on the image in a way an opaque numerical output would not be.

That visibility is worth crediting specifically. The located clinical testimony concerns exactly this, a clinician teaching midwives observing that caliper placement was accurate across multiple scans and multiple measurement types. Whether or not that is systematic evidence, it describes a clinician doing the review the design intends.

The competence question sits underneath and is unavoidable. The fetal biometry model is aimed explicitly at midwives, nurses and new ultrasound users in resource limited settings, so the operator may not have the training to recognise a plausible measurement from an implausible one. That is the point of the tool and it is also the reason the usual safeguard is weakest exactly where the tool is most used.

What is not described is behaviour at the edges: what the model does with an image that does not contain the anatomy it expects, whether it declines to measure rather than measuring badly, and what confidence signal the operator sees.

Graded B.

Ask what happens on an inadequate image and what confidence information is shown.

Above the category norm on description and still silent on performance.

The disclosure that exists is real and specific. Each model is named individually rather than bundled under a general artificial intelligence claim, so a buyer knows there is a musculoskeletal model, a fetal biometry model and a prostate model rather than a vague capability. The fetal model's training corpus size is published. The technique is characterised as deep learning rather than left unstated. Device specifications are published in detail, including frequency ranges, maximum imaging depth, scan and charge times, and the encryption used in transit and pairing. A technical evaluator gets further here unaided than with most records in this index.

What is missing is every measure of how well the models work. No model card, no accuracy or agreement figure against a reference standard, no error distribution, no inter operator variability data, no validation population and no retraining cadence were located.

For measurement models this is the material gap. A tool that places calipers automatically is competing against a trained sonographer placing them manually, and the only meaningful question is how closely the two agree and where they diverge. That comparison exists, because clearance required it. It has simply not been published in a form a buyer can read.

A version question sits alongside it. Models are delivered through the application to devices already in the field, so which model version produced a given measurement is a real record keeping matter, and no published version history was located.

Graded C.

Ask for agreement against manual measurement for each model, the validation population, and model version history.

One disclosure about training data and nothing about anything else.

The credit is the fetal biometry corpus, stated at more than 30,000 de identified fetal ultrasound images. That is a genuine data provenance disclosure and it is more than most records in this index offer, though it names scale without naming source, so a buyer still cannot tell whether the images came from partner institutions, licensed archives, research collaborations or the company's own customer base.

Everything else in the chain is closed. No cloud or infrastructure provider is named for the storage service, no sub processor register was located, no subcontractor list exists, and no component or manufacturing partner is identified for the scanners themselves. For a device business shipping hardware into more than 90 countries, the manufacturing and component chain is a real dependency and none of it is described.

The model development chain is equally unstated. Whether the models were built entirely in house, developed with academic or clinical partners, or licensed in from elsewhere is not addressed for any of the eight, and no external model provider is named or excluded.

One consequence is worth spelling out. Because models arrive through application updates to devices already in the field, a change in an upstream dependency propagates to installed scanners without the customer initiating anything, and nothing describes how such changes are communicated.

Graded D.

Ask where the training images came from, who provides cloud infrastructure, whether any model was developed or licensed externally, and how model changes are notified.

Regulatory depth is real and the customer evidence layer is absent.

The clearances carry genuine weight and are counted here rather than on this axis: eight models by mid 2024, two of them described as first of their kind. Each required performance data examined by a regulator, which is a materially stronger foundation than a marketing claim.

Adoption is stated in units that can at least be reasoned about. Scanners are reported in use across more than 90 countries, and three million high definition scans were reported performed. Both are real signals of installed base.

The gap is that no named institutional deployment with a quantified outcome was located. That is the specific distinction from Butterfly Network, graded a step higher on this axis in this index, which publishes a named university health system reporting a 116 percent increase in point of care ultrasound revenue after systemwide deployment and files audited financial statements as a listed company. Clarius is privately held, so revenue, growth and installed base cannot be verified independently, and the scan and country figures locate to 2023 with no more recent update found.

Clinical testimony that was located is individual rather than systematic, taking the form of a clinician describing caliper placement accuracy while teaching midwives. That is a useful signal about usability and it is not an outcome study.

No peer reviewed publication and no accuracy figure outside the regulatory process was located for any model.

Ask for named deployments with measured results, current scan and installed base figures, and any published accuracy data.

Better stewardship disclosure than peers, with the one question that matters most still open.

Three things are published that comparable vendors do not publish. Encryption parameters for both the wireless data channel and device pairing appear on product pages. A protected health information removal capability is named as a product feature rather than described vaguely, which matters for a device whose images are routinely used in teaching and case sharing. And local storage is offered alongside cloud storage, so a customer can decide whether clinical images enter the vendor's estate at all. That last point is the substantive one: the strongest privacy control is an architecture where the data does not arrive, and offering it as a supported option rather than an enterprise exception is a real design choice.

Model development disclosure is also above the norm. The fetal biometry model is stated to have been developed on more than 30,000 de identified fetal ultrasound images, which names both the scale of the corpus and the fact that it was de identified. Most vendors in this index name neither.

What remains unanswered is whether customer scans feed anything. Nothing located states whether images acquired by customers are used to train, tune or evaluate models, whether that differs between locally stored and cloud stored studies, or whether a customer can decline. For a company shipping new cleared models regularly, where the training data for later models comes from is a fair question and the published material does not address it.

Graded B.

Ask whether customer images train any model, and whether cloud storage changes that answer.

Technical handling is published in more detail than almost anywhere in this index, and the contractual layer is absent.

The published specifics are genuinely unusual. Product pages state the transport encryption used for the wireless data channel and the encryption used for device pairing, name a protected health information removal capability as a product feature, and state that such data is stored in a compliant manner. Vendors in this category normally assert compliance at the level of a badge; specifying the actual cryptographic parameters on a public product page is a different standard of disclosure and lets a security reviewer form a view before any conversation.

Storage choice supports the posture. Studies can be held locally or in the company's cloud, both unlocked by membership, so an organisation with a policy against clinical images leaving its control has an option that does not require abandoning the product.

What is missing is the agreement. No template business associate agreement, breach notification window, liability cap position, audit rights statement or data return provision was located.

The individual purchase channel raises the same question recorded against Butterfly Network. Scanners are bought by individual clinicians and small practices at published prices, in a transaction where no agreement is negotiated and standard terms govern, and nothing describes what those terms provide for a sole practitioner holding patient images.

Graded C.

Ask for the template agreement, the breach notification window, and what governs individual purchasers.

Unusually specific technical disclosure standing in place of a named certification.

What is published is concrete. Product pages state the transport encryption protocol for the wireless data channel and the symmetric and asymmetric algorithms used for device pairing, alongside a protected health information removal feature and a statement about compliant storage. Publishing actual cryptographic parameters rather than a compliance badge is a different and in some ways more useful kind of disclosure, because a reviewer can assess whether the choices are current rather than trusting that someone else assessed them.

What is absent is independent attestation. No service organisation control report at either assurance level, no information security management certification, no healthcare specific assurance framework certification, no penetration testing cadence and no vulnerability disclosure programme were located, and no trust centre or portal exists where a buyer could retrieve current artefacts without asking. Self published parameters are a statement of intent; an audited report is a third party saying the controls operate.

The comparison inside this index is instructive. Butterfly Network holds a named information security management certification and publishes no cryptographic detail. Clarius publishes the detail and no certification. Both sit at C for opposite reasons, and a buyer should ask each for what the other has.

The device layer is a further surface. These are connected medical devices pairing with personal phones and tablets across many countries, and firmware update integrity and lost device handling are not addressed.

Ask for an audited security report, firmware update and device authentication design, and penetration testing cadence.

A sustained programme of device clearances for the models themselves, not merely for the hardware they run on.

The depth is what earns the grade. By mid 2024 the company had taken eight artificial intelligence models through clearance, and it has continued adding to the line since. Two were described as first of their kind at the time: an application cleared for musculoskeletal ultrasound, automatically identifying and measuring tendons in the foot, ankle and knee, and later a fetal biometry tool. Clearing a model that measures anatomy and outputs a number a clinician acts on is a different undertaking from clearing an imaging device, and doing it eight times is a programme rather than an event.

The hardware carries its own approvals across markets, including regulatory clearance in Canada for the current generation line and availability across more than 90 countries.

One detail deserves particular credit because it is the kind of thing vendors usually blur. The prostate model is stated to be available in approved regions only, marked as such on the public product pages. That is a company telling prospective buyers that a capability shown in its materials may not be lawfully available to them, rather than leaving it to be discovered later. Regional regulatory divergence is real in this industry and disclosing it at the point of marketing is the honest practice.

Graded A, alongside Butterfly Network. That two handheld ultrasound vendors hold the strongest regulatory positions in this index is itself a finding: this is the corner of healthcare artificial intelligence where the device pathway is unavoidable and the vendors have accepted it.

Ask for the clearance numbers and intended use statements for each model.

One useful disclosure about the training corpus, and nothing about who is in it.

The credit first. The fetal biometry model is stated to have been developed on more than 30,000 de identified fetal ultrasound images. Naming the corpus size at all is more than most vendors in this index manage and it lets a reader judge whether the scale is plausible for the task.

What is absent is composition, and for this particular model composition is the whole question. Fetal biometry converts measurements into an age and weight estimate through growth references, and growth patterns differ across populations, with an established literature on whether a single international standard or population specific references should be applied. A model trained predominantly on one population carries that population's assumptions into every estimate it produces. Nothing located states where the 30,000 images came from, which reference the model applies, or how it performs across maternal characteristics.

The deployment strategy sharpens this rather than softening it. The tool is aimed explicitly at resource limited settings and at midwives and nurses rather than sonographers, so it will be used most heavily on populations plausibly least represented in a development corpus, by operators least equipped to notice a systematically wrong answer. Gestational age drives decisions about delivery timing and growth restriction, so a dating error is a clinical error.

The musculoskeletal and prostate models raise a quieter version of the same question, since tendon and prostate measurement norms vary by body habitus and by sex specific anatomy respectively.

A dedicated pass located no subgroup performance data, no fairness review and no model documentation.

Ask where the training images came from, which growth reference the model applies, and how performance varies across populations.

Hardware terms are published in detail and model output carries none, which makes the asymmetry unusually visible.

The company publishes a three year warranty extendable to five and a sixty day return policy, and states device environmental operating limits. A buyer therefore knows precisely what happens if the scanner fails.

Nothing equivalent covers the models. A dedicated pass located no indemnification position, no warranty covering measurement accuracy, no accuracy guarantee, no service credit regime and no described route to dispute a result. The same document set that tells a buyer the probe is covered for three years is silent on what happens if the fetal biometry model produces a materially wrong estimate.

The exposure is clinical and specific. Fetal age and weight estimates drive decisions about delivery timing and growth restriction assessment. The tool is aimed at midwives, nurses and new users, so the operator is less likely to catch an implausible result, and in resource limited settings there may be no second opinion available. A tendon measurement driving an orthopaedic decision and a prostate measurement driving a urological one carry smaller but real versions of the same risk.

The purchase channel compounds it. Many scanners are bought at published prices by individual clinicians where no agreement is negotiated, so whatever standard terms provide is the entire allocation of risk, and those terms were not located.

Regulatory clearance does not settle this. Clearance establishes that a regulator accepted the evidence for a stated use; it does not say who bears the loss when a cleared model is wrong within that use.

Ask what warranty attaches to measurement accuracy and what the standard terms allocate for individual purchasers.

Interoperability is included rather than sold separately, which is the distinguishing fact on this axis in this category.

Imaging interchange in the standard medical format is supported and included with membership, alongside archive integration, and studies can be stored locally or in the company's cloud. Measurements produced by the models appear on screen and can be added to the patient record directly from the application, so the output of an examination reaches the chart rather than remaining in a device gallery.

The contrast within this index is the point. Butterfly Network, graded a step lower on this axis, does not include imaging interchange in any membership tier and sells it separately, which means a clinician who has bought the device and the subscription still cannot get studies into the hospital archive by default. Here it comes with the membership a buyer is already paying for. For an imaging device, whether examinations reach the record determines whether they can be billed, audited or seen by the next clinician, so this is not a feature comparison but a workflow one.

Local storage adds a second path, letting an organisation move images without routing them through the vendor's cloud.

What holds it below the top grade is the absence of specifics. No record system integrations are named, nothing describes whether ordering and results flow in both directions or how an examination reconciles to an order, and archive configuration is described as requiring organisational setup, which implies work that is not quantified anywhere.

Ask which record systems are integrated, whether the connection is bidirectional, and what archive configuration involves.

A genuine architectural choice for the customer, which is rare in this category and worth more than a policy statement.

Studies can be stored locally on the acquisition device or in the company's cloud, both enabled through membership. That means an organisation whose policy prevents clinical images leaving its control has a supported configuration rather than an exception to negotiate, and a clinician working where connectivity is unreliable has a path that does not depend on upload. For a product sold into more than 90 countries and aimed partly at resource limited settings, working offline is a deployment requirement rather than a preference.

The acquisition architecture is described concretely. Probes connect directly to a phone or tablet over their own wireless link or a short range pairing, so the imaging chain does not depend on facility network infrastructure at all. Device operating characteristics are published, including scan time on a charge, charge time and standby duration, which matter for planning in exactly the settings where this product is deployed.

What is not stated is where cloud data lives. No infrastructure provider is named, no region selection is described and no residency position is published, which is a meaningful gap for a company operating across 90 countries including jurisdictions with strict data protection regimes and cross border transfer rules.

No recovery objective or availability commitment for the cloud service was located, though the local storage path reduces the clinical consequence of an outage compared with a cloud only design.

Graded B.

Ask where cloud data resides by market, whether region selection exists, and what happens to locally stored studies when a membership lapses.

The most complete commercial disclosure encountered in this index, and it covers terms rather than only prices.

A prospective buyer can assemble the whole picture without contacting anyone. Scanner prices are published from roughly 2,995 dollars, with regional equivalents stated. Membership is listed at 595 dollars a year, with a lifetime licence alternative offered and its regional variation acknowledged rather than hidden. Accessory pricing is published down to the level of a spare battery. A three year warranty extendable to five is stated. A sixty day return policy is offered.

Warranty and return terms are what lift this to the top grade. Price without terms is half a disclosure, because the questions that decide total cost of ownership are what happens when the device fails in year four and whether a buyer who has misjudged the fit can send it back. Publishing both is rare enough in healthcare technology to be worth naming.

The structural choice matters as much as the numbers. The probe functions without a membership, so the recurring fee buys storage, advanced models, teleguidance and imaging interchange rather than the right to use hardware already paid for. Set against Butterfly Network in this index, where a lapsed membership drops the probe to live viewing with no study saving and no artificial intelligence, this is a materially different bargain and a buyer comparing the two should weigh it directly.

What is not published is enterprise and volume pricing, and how membership scales across a department buying many scanners. Those are negotiated and no rate card was located, which is the only material gap on this axis.

Ask how membership prices at department scale and what enterprise agreements look like.

Coverage achieved by building a probe per purpose, which is the opposite strategy to a single whole body device and produces a different trade.

The current line runs to six models spanning full body scanning, high frequency imaging for musculoskeletal and small parts work, curved array abdominal and obstetric imaging, a dual array model and an endocavity scanner for urology and women's health. Cleared indications reach across eye, fetal and obstetric, abdominal, adult and paediatric cardiac, small parts, musculoskeletal and vascular imaging. Buyers include physiotherapists and orthopaedic clinicians, midwives and nurses, urologists, and general practice, in more than 90 countries.

Depth per indication is the advantage of this approach. A high frequency linear probe designed for tendon imaging will outperform a general purpose device on tendons, and the cleared musculoskeletal model exists because that hardware makes it possible. An endocavity scanner reaches examinations a surface probe cannot perform at all.

The cost is the trade and a buyer should price it explicitly. An organisation needing musculoskeletal, obstetric and endocavity coverage buys three scanners where a whole body device would be one, and if membership attaches per user or per device the recurring cost multiplies alongside. That comparison, run against Butterfly Network in this index, is the central purchasing decision in this category and neither vendor frames it neutrally.

What holds it at B is the absence of evidence about where the boundary sits against cart based systems by indication.

Ask which indications are supported to diagnostic rather than triage standard, and how membership prices across multiple scanners.