Philips Lumify

The models are additions to a conventional imaging product rather than the reason it exists, which places this with the rest of the handheld cluster and above Vscan Air within the band.

What carries the product is Philips imaging engineering ported down from the cart line: Tissue Harmonic Imaging, SonoCT and xRes for artifact reduction and border definition, three broadband transducers, fourteen presets and, since 2022, pulse wave Doppler for hemodynamic quantification. A clinician who never enabled a model would still be using a capable diagnostic ultrasound system, and the marketing leads with image quality and 30 years of ultrasound heritage rather than with artificial intelligence.

The two models that do ship are real and each does work a clinician would otherwise do by hand. Auto B-line Quantification counts B lines per frame and reports the maximum per zone, replacing a manual count that is tedious and known to vary between operators. Auto EF returns an ejection fraction from a single apical four chamber view with no manual border editing.

This sits above Vscan Air in the band because both features are included rather than sold as a paid option, and below the top of it because they are Android only, cover two organ systems, and one of the two is licensed from an outside developer rather than built in house. The device is an imaging platform with models on it, not a model delivered through a probe.

The intended use statement is the most disciplined in this cluster and it is not an accident of drafting.

The clearance describes a device intended for use by trained professionals at various settings of patient point of care. That is a materially different posture from the guidance products in this cluster, which are designed and marketed on the premise that an inexperienced operator can be walked to a diagnostic image. Philips positions the models as standardising the work of clinicians who already know how to scan, and the framing on Auto EF says exactly that, reducing subjectivity among users of varying experience rather than replacing expertise.

The B lines indication carries the same discipline. The cleared feature is described as supporting evaluation of the severity of lung conditions together with clinical and radiological information, which is an explicit statement that the output is an adjunct rather than a determination. Naming the other inputs a clinician must weigh is a small piece of drafting that does real work.

The design supports the framing. Auto B-line Quantification runs inside a guided 12 point protocol with zone labelling, so the number appears attached to a documented anatomical coverage rather than as a free floating count, and a reviewer can see which zones were scanned.

Graded B rather than A because nothing published sets competency expectations, describes what training accompanies the models, or defines the escalation path when an automated count or ejection fraction conflicts with the clinical picture. The framing is right and the operational controls behind it are undocumented.

The best model transparency in the handheld cluster, and it sits in the regulatory record rather than in the brochure.

The clearance summary for the cardiac feature states that Auto EF is based on the LVivo EF algorithm developed by DiA, identifying the outside developer of a shipped model by name. Very few vendors in this index name the supplier of a model they sell as their own feature. The same summary describes what the algorithm computes, end diastolic and end systolic volumes and the ejection fraction derived from them, and which transducer and preset it runs in. The lung summaries describe the counting mechanism frame by frame, the merged B line computation and the guided protocol it runs inside.

Imaging technology is named specifically rather than gestured at, with Tissue Harmonic Imaging, SonoCT and xRes described by what each contributes, and the constraint that the models run only on Android is stated in both the regulatory record and the product material.

Two gaps hold this at B. Training set size and provenance are not disclosed for either model, so the reader knows who built the cardiac algorithm without knowing what it learned from. And there is no model versioning story, which matters more here than usual: the vendor advertises automatic application updates delivering new features as soon as they are developed, so the algorithm producing an ejection fraction can change without the customer initiating anything, and nothing states whether a stored result records the version that produced it.

The best supply chain disclosure in the handheld cluster, with three separate upstream dependencies identifiable by name.

The cardiac model is licensed rather than built in house, and the clearance summary says so, identifying it as the LVivo EF algorithm developed by DiA. A vendor naming the outside developer of a model it markets under its own feature name is rare in this index, and it lets a diligence team evaluate the supplier independently.

The lung model has a disclosed funding and development lineage. The work ran under a partnership with the United States Biomedical Advanced Research and Development Authority, and the resulting clearance was publicly counted within that authority's portfolio of supported authorisations. Public co funding is a form of provenance most commercial models do not have, and it comes with an external party that has visibility into the development.

The tele ultrasound platform is a third named dependency, supplied by Innovative Imaging Technologies and integrated into the application, with the partnership and its origin documented rather than presented as a native capability.

What is missing keeps this from an A. No training corpus is described for either model beyond the validation dataset construction in the clearance summaries. Nothing states whether the licensed cardiac algorithm is updated by its developer or by the manufacturer, or how a change propagates through the automatic application updates. And while software bills of material are described as continuously monitored internally, no bill of material for this product is published for customers to inspect.

The device is well evidenced and the models on it are not, and that distinction sets the grade.

Lumify appears constantly in the point of care ultrasound literature as the acquisition instrument, including studies where it supplied 94 percent of the scans in a published lung imaging cohort. That establishes the hardware produces research grade images. It is not evidence that the shipped algorithms perform, because in those studies the algorithm under test was somebody else's.

For the models themselves the strongest located evidence sits in the clearance summaries. The automated B lines submission describes a study built to a pre defined success criterion of agreement with clinicians, ground truth set by clinician majority, with the threshold expressed as lower confidence limits of 0.746 for both sensitivity and specificity. Publishing the mechanism and the threshold is genuine transparency and the threshold itself is a modest bar, which is worth saying plainly rather than reporting the clearance as though it settled accuracy.

One prospective single centre study of automated B line quantification on this device was located, run at an academic medical centre on patients presenting with breathlessness during the pandemic, with repeat examinations over the admission and expert visual counts as reference. Its author list includes the manufacturer's own research organisation alongside the university, so it is company associated rather than independent.

No independent trial of Auto EF on this platform was located, and no outcome study attaching either feature to a change in management, length of stay or diagnostic accuracy in practice was found. Graded C against a cluster where two records carry independent multicentre validation.

The security documentation is strong and the training data question is unanswered, which is the shape this axis is built to separate.

On the safety side the record is good. A product specific system and data security document is published, the manufacturer disclosure statement programme covers protected data audit logging in detail, and the application runs on customer owned hardware so the default data path stays inside the customer's boundary. Exams reach the customer's own imaging archive.

On stewardship nothing current was located. No statement addresses whether images acquired on this device may be used to develop or refine models, under what de identification standard, or with what customer control. The absence is more conspicuous here than it would be elsewhere for two reasons. The application is described as cloud enabled and the original launch positioning explicitly described data being housed in the manufacturer's cloud platform with access to research enhancing analytics, which is a training adjacent phrase that has never been replaced by a clear current statement either way. And the tele ultrasound integration streams live exams through an outside platform whose data terms are not described in any material located.

What pushes this to C rather than lower is that the customer largely controls the data path by default, so a cautious organisation can configure around the ambiguity. What keeps it from B is that a buyer should not have to infer a training data position from a 2015 launch announcement and an architecture diagram.

Published documentation on protected data handling is the most specific in this cluster, and the contractual layer is still unpublished.

The manufacturer publishes a product specific system and data security document for this device and points buyers to it directly from the product page, alongside a statement that the device works with current encryption and data security systems and can be configured to comply with enterprise data security policies. Above that sits a manufacturer disclosure statement programme covering products that create, transmit or maintain electronic protected health information, supplying model specific answers including audit logging of protected data access covering viewing, creation, modification, deletion and import or export. A privacy office reviewing this device can therefore obtain structured, comparable answers on data handling before purchase rather than after contracting.

The device architecture helps. The application runs on hardware the customer owns and controls, exams can move to the customer's own imaging archive, and sharing runs to email or a shared network. That keeps much of the data path inside the customer's own compliance boundary.

What is not published is the business associate position. Nothing located offers an agreement, describes its terms or names which services it attaches to, and the question is live because the application is described as cloud enabled and the tele ultrasound service is operated by a third party. A buyer must raise it in procurement. Graded B: unusually strong published privacy control documentation, no published contractual posture.

The most complete published security posture in this cluster, and it is complete in the dimension that is hardest to fake: history.

The manufacturer operates a global product security policy governing incident management and risk assessment, a coordinated vulnerability disclosure policy with a published submission procedure, a dedicated reporting address and a public encryption key for researchers, and a security advisory archive organised by year that runs continuously from at least 2018 through 2025. Reading that archive, a buyer can see advisories issued proactively for the manufacturer's own products, weaknesses attributed to external researchers and validated internally, affected version ranges, mitigations, remediation timelines and explicit statements about whether exploitation or patient harm has been reported. Several advisories concern products with no patient safety implication and were published anyway.

Below that sits product level material. A manufacturer disclosure statement programme supplies model specific security answers on handling of electronic protected health information, including audit logging detail, using the standardised form endorsed by the main clinical engineering and health information bodies. Software bills of material are described as continuously monitored for new vulnerabilities, security practices are integrated across the development lifecycle in both premarket and postmarket phases, and controls are mapped to a recognised federal control catalogue. This specific device has its own published system and data security document.

One qualification. No independent audited certification such as an information security management standard was located for this product line, so the grade rests on disclosure practice and documented process rather than on third party attestation.

A long and steadily maintained clearance record, unusually easy to verify because the submissions are traceable and the summaries are substantive.

The platform was introduced in 2015 and the phased array transducer that opened cardiac use cleared in October 2016. The lung work cleared in two stages, a B line counting algorithm and then an automated merged B lines feature in June 2023 extended across all three transducers. The Auto EF feature cleared in 2023 in its own submission. A portfolio wide clearance in 2020 covered use of the ultrasound range, this device included, for management of pandemic related lung and cardiac complications, which was an unusual authorisation in its own right.

Two features of the record raise it above a simple count. The clearance summaries carry real technical content, describing the algorithms, the transducers each feature is cleared with, the operating system constraint and the validation design, so a reviewer can check what was authorised rather than inferring it from a press release. And the lung development ran under a partnership with the United States Biomedical Advanced Research and Development Authority, with the 2023 clearance counted publicly as the 74th authorisation supported under that programme, which places an additional layer of public accountability around the work.

The platform constraint is disclosed inside the regulatory record rather than only in marketing: the summaries state that the artificial intelligence features are compatible with Android and not with the Apple operating system. Disclosing a capability limit in the regulatory filing is the correct place for it.

Methodological disclosure is better than the norm and demographic disclosure is absent, so this lands mid band.

What is disclosed comes through the clearance summaries and it is more than most vendors publish anywhere. For the automated B lines feature the summary describes how the validation dataset was assembled, with coverage across anterior, lateral and posterior lung zones on both sides and multiple loops per patient, states that ground truth was established by clinician majority agreement, and gives the pre defined success criterion as lower confidence limits of 0.746 for sensitivity and specificity. Naming the ground truth method and fixing the threshold before the study is proper practice and it is legible to anyone who reads the summary.

What is absent is any performance breakdown by patient characteristic. No results by sex, age, body habitus or ethnicity were located for either feature, and body habitus is the obvious risk in both applications, since acoustic window quality drives both cardiac border detection and lung artifact visibility. There is no model card, no statement of training population composition, and no description of post market performance monitoring or of how a model update is revalidated.

The update mechanism sharpens the last point. The vendor advertises automatic application updates that roll out new features as soon as they are developed, so model behaviour can change on a device without a procurement event, and no governance process around that is described.

Graded C. The pre registration style discipline in the clearance work is genuine and it stops short of the subgroup question.

Nothing published sets out what happens when an automated count or an automated ejection fraction is wrong. No performance warranty attaches to a model output, no indemnity is described, no adjudication or error reporting route specific to the artificial intelligence is offered, and no accuracy commitment appears in any commercial term located.

What exists is a manufacturer's warranty on the equipment, service access through a customer portal at no cost, and education and support bundled into the purchase. Those cover a transducer that fails, not a number that misleads.

The cleared indication provides the only genuine boundary on reliance and it is a good one, describing the lung feature as supporting assessment of severity together with clinical and radiological information, and the device as intended for trained professionals. Read strictly that allocates interpretation, and the consequences of interpretation, to the clinician. It is the strongest implied position in this cluster and it is still implied, arriving through a regulatory indication rather than through a term a buyer negotiates.

One feature of the product makes the silence more consequential than it would otherwise be. The vendor advertises automatic application updates that deliver new features as soon as they are developed, so the model behaviour a clinician relies on can change without a purchasing decision, and nothing published describes notification, opt out, revalidation or recourse if a change degrades performance in a particular setting.

Graded D on the same basis applied across this cluster: the absence is of a published position, not of a reasonable one.

Imaging interoperability is included rather than gated, which is a better commercial posture than the neighbouring Vscan Air record, and the clinical record is untouched in the same way.

The device connects to a patient imaging archive and can share images, video and notes by email or over a shared network, and none of that sits behind a separate tier. One platform limitation is disclosed plainly on the product page: sharing exams to a shared network is not available on Apple devices. Disclosing a capability that works on one platform and not the other, in the product material rather than in a support note, is the right behaviour and it is worth crediting because the same vendor could easily have left a buyer to discover it.

Beyond imaging there is nothing. No health level seven interface, no fast healthcare interoperability resources support and no named electronic health record integration were located for this product. An automated ejection fraction or B line count reaches the archive and does not reach the note as a discrete coded result, which is the same ceiling every device in this cluster hits.

The tele ultrasound integration is a genuine interoperability capability of a different kind, moving a live exam and a conversation between two clinicians in real time with the scanning clinician controlling what is shared. That solves a collaboration problem rather than a records problem, and it does not lift this grade, but it is the capability most likely to matter day to day in the settings this device is sold into.

The local path is the leanest in this cluster and the remote paths are undescribed.

Locally there is very little to deploy. The transducer is cabled, so there is no pairing and no probe battery, the application comes from the ordinary consumer application stores, and it runs on a phone or tablet the organisation already owns and already manages under its own mobile device policy. That is a genuine advantage for a security team, because the endpoint is one they already control rather than a new managed device class, and the manufacturer states the device can be configured to comply with enterprise data security policies.

The remote paths are where the questions sit. The application is described as cloud enabled and the original launch positioning described exam data being housed in the manufacturer's own cloud platform, and nothing located states the current hosting arrangement, the region, whether residency options exist, or what retention applies. The tele ultrasound service is operated by an outside company, so a live exam and a patient facing camera feed cross into third party infrastructure whose location and terms are not described in any product material found.

One configuration detail deserves recording because it changes the data path by platform: sharing exams to a shared network is unavailable on Apple devices, so an organisation standardised on that hardware routes exams differently from one on Android.

Graded C. Excellent local deployment story, no published answer on where anything goes once it leaves the handset.

This is the least legible pricing in the handheld cluster, and the reason is not silence but inconsistency.

The vendor publishes what a purchase contains, which is more than nothing: transducer and cable, the application, software upgrades, support and education, a carry case and the manufacturer's warranty. Regional sites in some markets carry a subscription pricing page describing orders that bundle the transducer, the cloud enabled application, upgrades, warranty and access to service and education, with a purchase alternative available on request. The original 2015 launch was built on subscription at an introductory $199 a month, a genuinely unusual model for capital medical equipment at the time.

What a United States buyer can find today is contradictory. One commercial analysis checked in July 2026 reports that the current United States position is outright purchase with no software subscription, with tele ultrasound plans handled separately, and cites a public distributor listing near $12,027 for a single transducer. A second analysis published the same year puts new units at $5,995 to $9,995 including subscription, with refurbished units at $2,500 to $5,000. Those two accounts differ by roughly a factor of two on the same product in the same market, and neither is a vendor price.

The divergence is itself the finding. A buyer cannot determine from published material whether this device costs six thousand dollars or twelve, whether a subscription applies in their market, or how the answer changes across the three transducers. Both competitors that publish rate cards, Clarius and Butterfly Network, make that determination trivial. Graded C rather than lower because the inclusion list, warranty and subscription structure are documented and a regional pricing channel exists.

Coverage is the broadest in the cluster on the hardware and it is matched by a deliberately wide preset library rather than by marketing claims.

Three transducers cover the full point of care range: a phased array for cardiac and lung, a curved array for abdominal, obstetric and deeper work, and a linear array for vascular access, musculoskeletal and superficial imaging. Fourteen presets are published, including abdominal, cardiac, lung, FAST, musculoskeletal and obstetric, with ocular added recently, which is a preset most handhelds omit and one that matters in emergency and trauma assessment.

Settings run from prehospital and ambulance use through emergency departments, critical care, operating rooms, office practice and community programmes. The plug into an existing phone or tablet design lowers the deployment barrier further than any competitor here, because a service can equip clinicians without buying them hardware, and battery endurance of roughly 7.5 hours makes a full shift plausible without a charge.

The global health deployment is real rather than aspirational, with a foundation programme in Kenya training village midwives to perform antenatal screening supported remotely by radiologists. That is a use case the specification supports and most competitors describe only in principle.

The artificial intelligence coverage is narrower than the device coverage, spanning lung across all three transducers and cardiac on the phased array, and it is unavailable on Apple hardware entirely. That restriction is recorded on the model transparency and centrality axes rather than here, because the device itself covers the settings it claims.