Philips Lumify

The models are additions to a conventional imaging product rather than the reason it exists. What carries Lumify is Philips imaging engineering brought down from the cart line: Tissue Harmonic Imaging, SonoCT and xRes, three broadband transducers, fourteen presets and, since 2022, pulsed wave Doppler. The product page leads with image quality, workflow and Reacts tele ultrasound and does not mention artificial intelligence at all.

A clinician who never enabled a model would still be using a capable diagnostic ultrasound system. The two models that ship are real and each does work a clinician would otherwise do by hand: Auto B-line Quantification, a locked machine learning classifier cleared in K223771 that counts B lines and detects merged B lines per frame, and Auto EF, which returns an ejection fraction from a single apical four chamber view and is licensed from an outside developer. Both run only on Android.

The intended use keeps the models as aids for professionals who already scan. The K223771 indications place Lumify in environments where healthcare is provided by healthcare professionals, and Philips positions Auto EF as reducing subjectivity among users of varying experience rather than replacing expertise.

The lung feature is described in the clearance record as supporting evaluation of the severity of lung conditions together with clinical and radiological information, an explicit statement that the output is an adjunct. The design supports that framing: Auto B-line Quantification runs inside a guided 12 point protocol with zone labeling, so the count appears attached to documented anatomical coverage rather than as a free floating number, and a reviewer can see which zones were scanned.

What is not published is what training accompanies the models, what competency is expected, or the escalation path when an automated count or ejection fraction conflicts with the clinical picture; and the K223771 results show sensitivity falling to 0.74 on the linear transducer, a limit a user would need to know.

The regulatory record names what produces the outputs, with gaps on training data. The Auto EF clearance summary states that the feature is based on the LVivo EF algorithm developed by DiA, naming the outside developer of a shipped model, and describes what it computes: end diastolic and end systolic volumes and the ejection fraction derived from them, on a named transducer and preset.

K223771 describes the lung feature as a machine learning classifier that identifies B lines and merged B lines, running frame by frame inside a guided protocol, states that the algorithm is locked with no further adjustment once commercialized, and says its development data did not overlap its test data. Imaging technology is named specifically (Tissue Harmonic Imaging, SonoCT, xRes), and the Android only constraint is stated in the regulatory record.

What is missing is the size and provenance of either model's training data, and a version history: automatic application updates deliver new features as they are developed, and nothing says whether a stored result records the version that produced it.

Three upstream dependencies are identifiable by name. The cardiac model is licensed rather than built in house, and the clearance record says so, identifying it as the LVivo EF algorithm developed by DiA, so the supplier is named and open to evaluation on its own.

The lung model has a disclosed development lineage: it was built under a partnership with the US Biomedical Advanced Research and Development Authority, and K223771 describes it as a locked machine learning classifier with development data kept separate from its test data. The tele ultrasound platform, Reacts, is supplied by Innovative Imaging Technologies and integrated into the application, presented as a partnership rather than a native capability.

Missing are a training corpus for either model, a statement of whether the licensed cardiac algorithm is updated by its developer or by Philips and how a change reaches devices through automatic application updates, and a software bill of materials for this product that customers can inspect.

The device is well evidenced as an acquisition instrument, and the models on it rest on the company's own studies. Lumify appears widely in the point of care ultrasound literature as the scanner, including a published lung imaging cohort in which it supplied 94 percent of the scans; that shows the hardware produces research grade images, not that its shipped algorithms perform.

For the lung feature, the 510(k) summary (K223771, May 2023) reports a retrospective analysis of prospectively collected video from 157 adults with shortness of breath in a hospital setting, 416 loops, against clinician majority agreement, with an acceptance threshold set in advance (lower confidence limits of 0.746): merged B line sensitivity 0.83 (0.77 to 0.88), specificity 0.92 (0.88 to 0.96), count agreement ICC 0.91.

A prospective single center study of automated B line quantification on this device, at an academic medical center in patients presenting with breathlessness, with repeat exams over the admission and expert visual counts as reference, lists the manufacturer's research organization among its authors, so it is company associated. Philips cites no independent trial of Auto EF on this platform and no outcome study linking either feature to a change in management or diagnosis.

Security documentation is substantial, and the training data question is unanswered. The product page says Philips diligently safeguards information to ensure confidentiality, integrity and availability, and exams go by email, DICOM to PACS, a shared network or a local directory, so the default data path stays within the customer's own systems; the application runs on a phone or tablet the customer owns.

Nothing published addresses whether images acquired on Lumify may be used to develop or refine models, under what de identification standard, or with what customer control. The absence is conspicuous for two reasons: the application was launched in 2015 as cloud enabled, with exam data described as housed in Philips' cloud with access to research enhancing analytics, a training adjacent phrase never replaced by a clear current statement either way; and the Reacts tele ultrasound integration streams live exams through an outside platform whose data terms are not described.

The B lines classifier is stated to be locked once commercialized, and its development data were kept separate from its test data. Safety engineering, meaning guardrails and how a safety event is handled, is not published.

Product level data security documentation is published, and no business associate position is. Philips publishes a product specific system and data security document for Lumify and runs a manufacturer disclosure statement (MDS2) program for products that create, transmit or maintain electronic protected health information, with model specific answers including audit logging of access, creation, modification, deletion and export.

The product page says Philips diligently safeguards information for confidentiality, integrity and availability, and exams move to the customer's own archive, email, network share or local directory, which keeps much of the data path inside the customer's compliance boundary.

Philips does not state its status for the service: no business associate agreement is offered, described or scoped, and the question is live because the application has been described as cloud enabled and Reacts tele ultrasound is operated by a third party.

Philips runs a mature product security disclosure program, and no independent certification of it is named. The Philips security pages lead to three sections: coordinated vulnerability disclosure, product security and security advisories.

The disclosure policy publishes a submission procedure and a reporting address for researchers, and the advisory archive runs by year, with advisories that credit external researchers, give affected version ranges and mitigations, and state whether exploitation or patient harm has been reported.

Philips also describes an MDS2 program with model specific security answers, software bills of materials monitored for new vulnerabilities, security practices across the development lifecycle, and controls mapped to a federal control catalog; Lumify has its own system and data security document. The outside check is the external researchers the disclosure program invites and credits.

No audited certification such as ISO/IEC 27001, SOC 2 or IEC 62443-4-1 is named on the security landing page or the product page, and no attestation can be opened, though the product security section of the Philips site could name one.

A long, verifiable clearance record, with the AI lung feature cleared in its own submission. FDA records list Lumify clearances as K152899 (Philips Lumify Diagnostic Ultrasound System, 30 October 2015), K162549 (October 2016, the S4-1 phased array that opened cardiac use), K201012 (May 2020, a portfolio clearance covering EPIQ, Affiniti, CX50, Sparq and Lumify), K203406 (February 2021), K223771 (4 May 2023), K242519 and K242670 (May 2025), K252557 (December 2025) and K252558 (May 2026), all under 21 CFR 892.1550.

K223771 is the Lumify Diagnostic Ultrasound System with Expanded B-lines Software Feature: automatic B line counting and merged B line detection by a locked machine learning classifier, on the C5-2, L12-4 and S4-1 transducers, compatible only with Android, for use where healthcare is provided by healthcare professionals, with K203406 and K152899 as predicates.

Auto EF, which the clearance record attributes to the LVivo EF algorithm developed by DiA, carries its own clearance, whose number is not on the product page. The lung work was developed under a partnership with the US Biomedical Advanced Research and Development Authority.

The B lines clearance names its subgroups and does not show results for all of them. The 510(k) summary for the expanded B lines feature (K223771, May 2023) tested the locked classifier on 416 video loops from 157 adults with shortness of breath, against clinician majority agreement, with the acceptance criterion fixed in advance (lower confidence limits of 0.746 for sensitivity and specificity).

It reports results by transducer, and the spread is informative: merged B line sensitivity 0.89 on the C5-2, 0.81 on the S4-1 and 0.74 on the L12-4, whose point estimate sits below the prespecified lower limit for the pooled result. It names BMI bands (under 30, 30 and over, 40 and over) and image quality (poor or good) as analyzed subgroups; the summary gives no figures by BMI.

Body habitus is the obvious risk for both lung and cardiac features, since acoustic window quality drives artifact visibility and border detection. The automatic application updates that deliver new features sit beside a feature the summary describes as locked once commercialized.

A published error rate with its method and denominator is on the record for the lung feature, and nothing stands behind it contractually. The 510(k) summary for the expanded B lines feature (K223771, decision 4 May 2023) is Philips' own filing: a locked machine learning classifier tested retrospectively on prospectively collected lung ultrasound, 416 video loops from 157 adults with shortness of breath in a hospital setting, against clinician majority agreement.

Merged B line detection across transducers reached sensitivity 0.83 (95 percent CI 0.77 to 0.88) and specificity 0.92 (0.88 to 0.96), and agreement on the maximum B line count had an ICC of 0.91 (0.89 to 0.93); by transducer, sensitivity ran from 0.89 (C5-2) down to 0.74 (L12-4). It is a result on a test set for the cleared feature, not performance in use, and it covers the lung feature only: Auto EF, licensed from an outside developer, is not covered by it.

No warranty covering model output, accuracy commitment, indemnity or AI specific error route is published; the equipment warranty and service access cover a transducer that fails, not a number that misleads.

Imaging interoperability is included rather than gated, and the clinical record is untouched. The product page says exams are delivered by email, DICOM to PACS, a shared network or a local directory, and none of that sits behind a separate tier; the product material states plainly that sharing to a network is not available on Apple devices. Beyond imaging nothing is published: no HL7 interface, no FHIR support and no named EHR integration.

An automated ejection fraction or B line count reaches the archive and does not reach the note as a discrete coded result. Reacts tele ultrasound is interoperability of a different kind, moving a live exam and a conversation between two clinicians with the scanning clinician controlling what is shared; it solves a collaboration problem rather than a records problem.

The local path is lean, and the remote paths are undescribed. The transducer is cabled, so there is no pairing and no probe battery; the application comes from the ordinary app stores and runs on a phone or tablet the organization already owns and manages under its own mobile device policy, which is an advantage for a security team.

Exams go by email, DICOM to PACS, a shared network or a local directory, and sharing to a network is not available on Apple devices, so the data path differs by platform. The remote paths are where the questions sit. The application was launched as cloud enabled, with exam data described as housed in Philips' cloud, and nothing current states the hosting arrangement, region, residency options or retention. Reacts, the tele ultrasound service, is operated by an outside company, so a live exam and a camera feed cross into third party infrastructure whose location and terms are not described.

The buying model is stated, and the price is not. The US product page says there are no subscriptions and that a buyer purchases outright and gets access to all software and features, which settles a question older material left open: the product launched in 2015 on a subscription at an introductory $199 a month, and regional pages in some markets have described subscription orders bundling the transducer, the application, upgrades, warranty, service and education.

What the vendor publishes is the inclusion list: transducer and cable, the application, software upgrades, support and education, a carry case and the manufacturer's warranty, with service tickets through a customer portal at no cost, and both AI features included rather than sold as options. What it does not publish is a price for any of the three transducers or for the Reacts tele ultrasound service.

Coverage is broad on the hardware, and validation of the AI stands behind part of it. Three transducers cover the point of care range: a phased array for cardiac and lung work, a curved array for abdominal, obstetric and deeper imaging, and a linear array for vascular access, musculoskeletal and superficial imaging.

Fourteen presets are published, including abdominal, cardiac, lung, FAST, musculoskeletal, obstetric and ocular, and K223771's indications span fetal and obstetric, abdominal, pediatric, cardiac and lung applications. Settings run from prehospital and ambulance use through emergency departments, critical care, operating rooms, office practice and community programs, and a Philips Foundation program in Kenya trains village midwives to perform antenatal screening with remote radiologist support. The AI coverage is narrower: the lung feature is validated in adults with shortness of breath in a hospital setting (K223771), Auto EF on the phased array, and both are unavailable on Apple hardware.