EchoNous

The models here measure and guide, they do not decide. That places this record with Butterfly Network and Clarius rather than with Exo, and the line drawn on the Exo record holds: an automated ejection fraction is a quantification a clinician would otherwise perform, whereas a cleared indicator that detects pneumonia is a finding a clinician would otherwise have had to make.

The distinction matters more here than elsewhere because the artificial intelligence is genuinely substantial. AI FAST performs real time anatomical labelling and view identification, Auto Preset switches exam presets as the operator moves between organ systems, Auto Doppler places the sample gate on the valve under interrogation, and the assisted ejection fraction workflow returns a left ventricular measurement without manual tracing. Kosmos AI is sold as its own product line rather than as a feature list.

What holds the grade at C is where the company itself puts the argument. The pitch against cart based systems rests on continuous wave Doppler, proprietary piezoelectric transducers and proprietary application specific circuits, which is a hardware case. Strip the models out and a capable diagnostic ultrasound system remains, which is not true of a vendor whose product is the model. The artificial intelligence makes the device easier to use and faster to document with, and it is not the reason the device works.

The device is a supervised instrument and the company designs it that way. Automated tracings and measurements appear on screen where the operator can review and adjust them, and the vendor's own validation protocol was built around exactly that behaviour, with four echocardiographers scanning and free to correct the automated left ventricular tracing before it was recorded. Describing the tool as it is actually used, rather than as a fully automatic result, is the honest construction and it is not the common one.

The pressure on that model comes from the marketing rather than the design. The stated benefit is that proprietary models help novice users learn ultrasound faster, and the evidence base leans on a study where oncology staff without ultrasound expertise produced accurate assessments. Named user groups include nurses and pre hospital staff. Where the selling proposition is that an inexperienced operator can rely on the number, the supervision the design assumes is thinnest precisely where the design most depends on it, and nothing published addresses that tension or sets a competency threshold for the automated features.

Graded B rather than lower because the outputs are quantifications a clinician can see, sanity check against the image in front of them and override, and because image quality grading is itself surfaced to the user, which gives a novice a signal that the acquisition was poor. That is a meaningfully better oversight surface than a system that returns a finding without showing its working.

Hardware transparency is well above average and model transparency is well below it, which makes this record a clean illustration of the difference.

On the hardware the company is specific in a way that can be checked: proprietary piezoelectric transducers described as chosen for image quality and durability, proprietary application specific circuits, the processor doing the work, ingress protection ratings for probes and tablet stated separately, battery endurance quoted by platform and by whether the artificial intelligence and continuous wave Doppler are running, and an explicit engineering explanation of why two variants of the same phased array probe are not interchangeable across platforms. Each artificial intelligence feature is described individually in terms of what it does.

On the models themselves nothing is published. No architecture, no training set size or composition, no validation set separation, no update cadence and, most consequentially, no model versioning. The device takes over the air updates, so the model producing an ejection fraction today may not be the model that produced one last quarter, and nothing states whether a result carries the version that generated it. For a measurement that may be compared serially in the same patient across chemotherapy cycles, which is a use case the company markets, silent model change is a real reproducibility question and it is unaddressed.

Third party components are the exception and are handled well, with the report generation and tele ultrasound partners named rather than described generically.

Third party dependencies are named openly, which is more than most records in this index manage. Two outside services are identified by company name and by function: one supplying automated echocardiography report generation for interpretation speed and reproducibility, and one supplying tele ultrasound streaming of the live exam, device camera and audio. Naming a model partner rather than describing an unnamed integration is the disclosure that lets a buyer run their own diligence on it, and the company does that without being asked.

Beyond the names nothing is disclosed. No contractual terms are described for either partner, and specifically nothing states what happens to a study streamed through the tele ultrasound service or submitted for automated reporting, which is the point at which patient data leaves the local architecture the privacy policy is built on. Those two integrations are the exception to the vendor's central data claim and they are documented as features rather than as data flows.

The first party models are equally opaque upstream. Whether any pretrained or licensed component sits inside the labelling, view identification or ejection fraction models, and what corpus those were built on, is not addressed. The hardware supply chain by contrast is partly visible, with the ruggedised tablet built with a named industrial hardware partner and transducers and circuits described as proprietary.

Graded C: the naming is genuine and the terms behind the names are not.

This is the strongest evidence base encountered in the handheld ultrasound cluster, and the reason is that most of it was not produced by the company.

Four independent lines exist for the automated ejection fraction alone. A multicentre prospective study across four Japanese centres, with separate image acquisition, analysis core laboratory and statistical core laboratory, detected reduced ejection fraction below 50 percent with sensitivity of 85 percent and specificity of 81 percent against high end equipment. A study in npj Digital Medicine assessed diagnostic accuracy against echocardiographer reads. A single centre Italian cohort of 49 patients compared the automated result against cardiac magnetic resonance, the reference standard, and reported a correlation of 0.99 with a bias of 1.1 percent. A 115 patient cardio oncology cohort published in a cardiology journal had oncology staff without ultrasound expertise perform the scans, reaching accuracies of 89 to 94 percent, which tests the claim that actually matters commercially.

The independent work also publishes against the product. The Japanese study reported that the automated system tended to underestimate ventricular volumes even where correlation was good, and the npj authors noted convenience sampling that may have excluded the sickest patients. Findings of that kind are what separate a validated product from a well marketed one.

The company additionally publishes a prospective 78 patient study run at a single clinic with four echocardiographers scanning, adjudicated blind by core laboratories at two academic medical centres using cleared third party calculation software. That is a well designed study and it is the vendor's own, recorded here as company evidence rather than independent evidence. The grade does not depend on it.

The question every health system asks a model vendor is whether patient data feeds the next version of the model. This record carries the strongest available answer to it, which is that the data never reaches the vendor. Processing is local, the transmitted class is limited to serial numbers, software versions and crash logs, and the policy states that class carries no protected health information. Retention and deletion sit with the user, who can clear studies through the application or by removing it, and technical data deletion can be requested from a named address.

Two gaps hold this below A. The first is encryption at rest. The policy commits to commercially reasonable safeguards and specifies encryption in transit for technical data, and says nothing about encryption of patient studies stored on the device. For a tablet carried between wards, taken to a dialysis clinic and used in resource limited field settings, at rest protection of locally held studies is the control that matters most, and it is the one not described.

The second is provenance of the training data. Nothing published states where the images used to build the labelling, view identification and ejection fraction models came from, under what de identification standard, or with what consent basis. A local processing architecture answers what happens to a customer's data going forward. It says nothing about the corpus the shipped models were built on, and that is a stewardship question in its own right.

The posture is architectural and it is stated with unusual precision. The products privacy policy, last reviewed 23 March 2026, distinguishes two data classes and treats them differently. Device identifiers and crash logs are collected and transmitted, for licence verification, support and mandated post market surveillance, and the policy states explicitly that this class contains no protected health information. Patient data, meaning ultrasound images, names, dates of birth, medical record numbers and clinical measurements, is accessed only locally on the device to display imaging, perform calculations and generate reports, and the policy states that EchoNous does not access, collect or disclose it. A Chief Privacy Officer is named with a contact address, which is more accountability than most records in this index carry.

If that description is accurate then EchoNous may not be a business associate at all for the imaging workflow, and the absence of a business associate agreement would be correct rather than a gap. The problem is that the company never says so. Nothing published addresses the agreement question in either direction, which leaves a hospital privacy office to infer the answer from an architecture description written for an application store policy.

Two things also sit awkwardly against the blanket claim. The policy states patient data is not disclosable by EchoNous or any third party, while the platform offers a tele ultrasound integration that streams the live exam and a third party report service, both of which move patient data to parties other than the customer under terms the policy does not describe. Graded B because the stated posture is strong, plainly written and recently reviewed, and because the two questions a covered entity would actually ask are unanswered.

Nothing published describes a security programme. No audited certification of any kind was located, no trust centre or security page exists on the site, no penetration testing statement, no software bill of materials, no coordinated vulnerability disclosure policy and no incident notification commitment. The only security language found anywhere is a single line in the products privacy policy committing to commercially reasonable safeguards and encryption in transit for technical data.

This is the sharpest contrast in the handheld cluster. Exo publishes multiple externally audited credentials and a portal where a buyer can request the audited report rather than negotiate for it. Clarius specifies transport and pairing encryption on its product pages. Here a hospital security review would find nothing to review.

The local processing architecture reduces the blast radius of that absence and does not remove it. This is a networked device that registers over the internet, receives over the air software updates, exports to enterprise imaging destinations, pairs with hospital owned tablets and streams live exams through a third party service. Every one of those is an attack surface, and update integrity in particular is the control a medical device security review begins with. Regulators have expected premarket cybersecurity documentation for devices of this class for several years, so material almost certainly exists in the submissions. None of it is public.

Graded D as an absence of published evidence rather than an assertion about the underlying engineering.

The regulatory position is unambiguous and repeatedly refreshed. The company operates as a medical device manufacturer rather than as a software vendor arguing it sits outside a pathway, and clearances span nearly a decade: the vein guidance product cleared in 2018, the Kosmos platform cleared through the 510(k) route in March 2020, the linear probe cleared in 2021, and the February 2026 release described as adding cleared presets for obstetric biometry and vascular workflows. The bladder scanning line predates all of it under the original Signostics identity.

The February 2026 expansion is the one that matters most for a buyer, because it is framed around billable studies. Moving from visualisation to comprehensive measurement changes what the device is used for and what a clinician may claim for it, and the company describes those presets as cleared rather than as capabilities. That is the correct representation and a common place for vendors to overstate.

International posture is consistent with the same discipline, with distribution across more than 70 countries and geography specific availability documentation, which implies conformity assessment well beyond the United States.

One documentation gap keeps this from being complete rather than merely strong. No consolidated public list of clearance numbers with their dates and indications was located, so a buyer verifying a specific claim must reconstruct it from press releases and the public device database. That is a publishing choice rather than a regulatory shortfall, and it does not change the underlying status.

The company hosts a public validation library and one entry in it is unusual enough to be worth naming: a study designed to evaluate the system as an anatomy education tool and to explore whether gender bias appears in its use. Vendors rarely publish work that goes looking for bias in their own product, and hosting it openly is a real act rather than a gesture.

It does not, however, answer the question this axis asks. A study of bias in how the system is used as a teaching aid is adjacent to, and not the same as, whether the diagnostic models perform equally across patient groups. Nothing published breaks out automated ejection fraction performance by sex, body habitus, age or ethnicity, and body habitus is the obvious risk here because acoustic window quality varies systematically with it and the product is aimed at operators least equipped to recognise a poor window. No model card, no intended use limitation statement and no statement of training population composition were located.

The closest thing to a published performance limitation comes from outside the company, in the multicentre Japanese study reporting a tendency to underestimate ventricular volumes. That is useful and it is not the vendor's disclosure.

Graded C because one genuine and self critical artefact exists and the substantive governance disclosures do not. Removing that artefact would put this record at the floor.

Nothing published addresses what happens when an automated measurement is wrong. No performance warranty attaches to a model output, no indemnity is described, no error reporting or adjudication path is offered, and no service level commitment covers accuracy.

What exists covers the hardware. A five year warranty is published, unusually long for this category and a real commitment, and a United States based service team is reachable by phone and email. That protects a buyer whose probe fails. It says nothing about a buyer whose ejection fraction reads 55 percent when the true value is 40, which is the failure mode the artificial intelligence introduces.

The gap carries more weight here than it would on a record where the tool serves an expert. The marketed benefit is that operators without ultrasound training reach reliable results, the evidence base leans on studies of exactly those operators, and the named user groups include nurses and pre hospital staff. A design that invites reliance and a contract silent on the consequences of reliance is an asymmetry a buyer should price.

The conventional answer, that the clinician remains responsible because the output is reviewable, is available to the company and is not made anywhere located. Graded D because the position is undocumented rather than because it is unreasonable.

Imaging interoperability is real and was deepened this year. The device exports through the standard medical imaging protocol, supports a modality worklist through a configured imaging profile, offers direct file export over cable, supports screen casting on all three platforms, and publishes a software development kit and application programming interface page for partners. The June 2026 release added support for the web based variant of the imaging standard, which is the modern route into enterprise imaging archives and cloud connected image exchange, and that is a genuine advance over cable and file transfer.

The clinical record is the missing half. No health level seven interface, no fast healthcare interoperability resources support, and no named electronic health record integration were located. Nothing describes an automated ejection fraction landing as a discrete, coded result in a patient chart in Epic or Oracle Health. What the device produces reaches the imaging archive; whether it reaches the note depends entirely on whether the customer has already wired that archive into the chart.

That gap costs more here than it would for a pure imaging device, because the output is a number a clinician wants in the record rather than a picture a radiologist will read. It is the difference between an exam being stored and a measurement being usable downstream for trending, quality reporting or billing support.

Graded C: solid and improving on the imaging side, absent on the record side.

Deployment is the simplest arrangement available in this index and the company states it plainly rather than leaving it to be inferred. The models run on the device. Studies stay on the device. A network connection is required for exactly three things, all of them named: initial transducer registration, export to an imaging destination, and software updates. Scanning itself works with no connectivity at all, and updates can be applied over a wireless connection or from a flash drive.

For residency that is a strong answer. Patient data comes to rest wherever the customer's own archive sits, so a buyer operating under a data localisation rule inherits the vendor's compliance rather than negotiating for it, and there is no vendor cloud tenancy to ask questions about. The offline capability also makes the marketed global health and field deployment use credible in a way that a cloud dependent device would not be.

Two things keep this from an A. The third party integrations offered on the platform, tele ultrasound streaming and automated report generation, are network services by definition, and nothing published states where they are hosted, under what terms, or which of them are available in which jurisdictions. And the technical data the company does collect, device identifiers and crash logs, has no stated storage location or retention period beyond a general statement that it is kept as long as needed for support and regulatory record keeping.

One real number is published and the rest is a quote request. The company states that its Kosmos Plus configuration, a full system with probes, artificial intelligence, advanced Doppler and a tablet, sits under $20,000. A ceiling for a named configuration is genuine disclosure and it is more than nothing, which is why this does not sit at the floor. Alongside it the company publishes a five year warranty, United States based service, an ingress protection rating, battery endurance figures and availability across more than 70 countries, all of which bear on total cost of ownership.

What is missing is the structure. There is no probe level price, no configuration ladder, and no statement anywhere located of whether a recurring fee exists. That last omission is the material one. A third party device listing describes new systems in a range of roughly $7,495 to $11,995 with an annual subscription of about $720 described as required, and the vendor's own materials do not mention any subscription at all. The index cannot corroborate that listing against a second independent source and does not treat it as established, but the question it raises is the right one: both direct competitors in this cluster charge an annual fee, and the terms of that fee decide five year cost. A buyer cannot answer it from anything the company publishes.

Set against the cluster this is the weakest of the four. Clarius publishes scanners, membership, accessories, warranty and returns. Butterfly Network publishes device and tier pricing. Exo published a launch device price. Here a buyer knows only that one configuration costs less than $20,000.

Coverage is broad and specific rather than broad and vague. Twelve clinical areas are addressed with their own material: anaesthesiology, cardiology, critical care, emergency medicine, family medicine and primary care, global health, hospital medicine, musculoskeletal, nephrology, nursing for bladder assessment and ultrasound guided intravenous access, obstetrics and gynaecology, and vascular access.

The hardware supports the claim rather than trailing it. A phased array probe covers cardiac and abdominal imaging, a linear probe covers lung, vascular access, peripherally inserted central catheter placement, musculoskeletal, nerve blocks and foreign body work, and a separate bladder scanning probe carries the original Signostics business. The February 2026 release added obstetric biometry including gestational sac, crown rump length and femur length, which moves the device from viewing to dating and viability assessment.

Deployment settings are equally wide. Three platforms are supported, Apple iOS, Android and a proprietary tablet for users who want neither, plus a ruggedised Android tablet built with an industrial hardware partner and a small cart configuration. The device is sold in more than 70 countries and is marketed explicitly into resource limited clinics as well as tertiary hospitals. Offline scanning makes that credible rather than aspirational.

One inconsistency is worth recording plainly. The company's own frequently asked questions page, last modified in June 2026, states that there is no obstetric preset, four months after the press release announcing cleared obstetric presets. That is a documentation currency failure rather than a coverage failure, and it is noted here because a buyer researching obstetric capability would find the contradiction.