EchoNous

The models measure and guide, and the device works without them. EchoNous says Kosmos leverages proprietary deep learning algorithms to help novice users learn ultrasound quicker, and the AI features are substantial: AI FAST labels anatomy and identifies views in real time, Auto Preset switches the exam preset as the operator moves between organ systems, Auto Doppler places the sample gate on the valve being interrogated, and an assisted ejection fraction workflow returns a left ventricular measurement without manual tracing.

The company's own case against cart based systems rests on hardware: continuous wave Doppler, proprietary piezoelectric transducers and proprietary application specific circuits. Remove the models and a capable diagnostic ultrasound system remains.

The device is a supervised instrument, and the company describes it that way. Automated tracings and measurements appear on screen where the operator can review and adjust them, and the company's own validation protocol was built around that behavior, with echocardiographers scanning and free to correct the automated left ventricular tracing before it was recorded. Image quality grading is surfaced to the user, which gives a novice a signal that an acquisition was poor.

The pressure on that design comes from the marketing: the stated benefit is that proprietary models help novice users learn ultrasound quicker, and named user groups include nurses and pre hospital staff. Where the selling point is that an inexperienced operator can rely on the number, the supervision the design assumes is thinnest, and nothing published sets a competency threshold for the automated features or says what a feature does with an inadequate image.

Hardware transparency is detailed, and model transparency is absent. On the hardware the company is specific: proprietary piezoelectric transducers, proprietary application specific circuits, ingress protection ratings for probes and tablet, battery endurance with and without AI and continuous wave Doppler running, and why two variants of the same phased array probe are not interchangeable.

Each AI feature is described by what it does, and the home page says the models are proprietary deep learning. Nothing is published on architecture, training set size or composition, validation set separation, update cadence or model versioning.

The device takes over the air updates, so the model producing an ejection fraction today may not be the one that produced it last quarter, and nothing says whether a result carries the version that generated it; for serial measurements across chemotherapy cycles, a use the company markets, that is a reproducibility question.

A model provider is named on the platform, and the terms behind the names are not published. Two outside services are identified by company and function: Us2.ai supplies automated echocardiography report generation, and 19Labs supplies tele ultrasound streaming of the live exam, device camera and audio. The company's own deep learning models run on the device.

What is not disclosed is what happens to a study streamed through the tele ultrasound service or submitted for automated reporting, which is where patient data leaves the local architecture the privacy policy is built on, and whether any pretrained or licensed component sits inside the labeling, view identification or ejection fraction models. The ruggedized tablet is built with an industrial hardware partner, and the transducers and circuits are described as proprietary.

Most of the evidence for the automated ejection fraction was produced outside the company, and some of it reports against the product. A prospective multicenter study across four Japanese centers, 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, and reported that the automated system tended to underestimate ventricular volumes even where correlation was good.

A study in npj Digital Medicine assessed diagnostic accuracy against echocardiographer reads, its authors noting convenience sampling that may have excluded the sickest patients. A single center Italian cohort of 49 patients compared the automated result with cardiac magnetic resonance, reporting a correlation of 0.99 and a bias of 1.1 percent. A 115 patient cardio oncology cohort had oncology staff without ultrasound expertise perform the scans, reaching accuracies of 89 to 94 percent.

The company also publishes its own prospective 78 patient study, run at a single clinic and adjudicated blind by core laboratories at two academic centers; that one is the vendor's own evidence. The evidence covers the ejection fraction workflow; the other AI features (AI FAST, Auto Preset, Auto Doppler) have none published.

The products privacy policy answers the question every health system asks a model vendor, whether patient data feeds the next version of the model, architecturally: the data never reaches EchoNous. 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 deletion of technical data can be requested from a named address. Two gaps remain. The policy commits to commercially reasonable safeguards and encryption in transit for technical data, and says nothing about encryption of patient studies stored on the device, the control that matters for a tablet carried between wards and into field settings.

And nothing published states where the images used to build the labeling, view identification and ejection fraction models came from, or under what de identification standard or consent basis; local processing answers what happens to a customer's data going forward, not the corpus the shipped models were built on. The tele ultrasound and automated reporting partner services are the exception to the local architecture and are described as features, not as data flows.

The company states plainly that it does not handle patient data, and explains the scope in a service privacy document. The products privacy policy, dated 23 March 2026, separates two data classes. Device identifiers and crash logs are collected for license verification, support and post market surveillance, and the policy says this class contains no protected health information.

Patient data (images, names, dates of birth, record numbers and measurements) is accessed only locally on the device, and the policy states that EchoNous does not access, collect or disclose it. A Chief Privacy Officer is named with a contact address, and the home page describes the product as HIPAA compliant.

Two things sit awkwardly beside the blanket statement: the tele ultrasound streaming and third party reporting services move patient data to parties other than the customer, under terms the policy does not describe, and the company does not say whether a business associate agreement covers them.

A SOC 2 compliance claim is made on the home page, and nothing behind it is published. The home page describes the product as HIPAA and SOC2 compliant for data protection and privacy. No SOC 2 report type, period, scope or owner is stated; the site footer links privacy policies, terms of use and a cookie policy but no security or trust page; and no penetration testing statement, software bill of materials, vulnerability disclosure policy or incident notification commitment is published.

The products privacy policy commits to commercially reasonable safeguards and encryption in transit for technical data. The local processing architecture reduces the blast radius without removing it: the device registers over the internet, receives over the air updates, exports to enterprise imaging destinations and streams live exams through a partner service, and update integrity is where a medical device security review begins.

The device clearances are identifiable; the clearance status of the AI features is less clear. EchoNous holds five 510(k)s as applicant: K181574 (Uscan, a bladder scanner, 10 July 2018, 21 CFR 892.1560), K193518 (KOSMOS, 26 March 2020, 892.1550), K212100 (Kosmos, 6 August 2021), K233826 (Kosmos, 29 August 2024) and K242321 (Kosmos, 25 April 2025).

The 2020 summary covers cardiac, thoracic, abdominal and peripheral vascular assessment and needle guidance by qualified and trained professionals, with no AI feature and no clinical data. The 2021 summary states that the AI assisted ejection fraction workflow and the Trio tool were not yet cleared and were being offered under FDA's COVID-19 enforcement policy for imaging systems.

The 2024 and 2025 clearances followed, and a February 2026 release was announced as adding cleared obstetric biometry and vascular presets. Which AI features each later clearance covers, and with what indication, is not stated.

One self critical study exists, and it addresses a different question. The company's validation library hosts a study that evaluates the system as an anatomy education tool and explores whether gender bias appears in its use, a real step for a vendor to publish. It is adjacent to, and not the same as, whether the models perform equally across patients.

Nothing published breaks out automated ejection fraction performance by sex, body habitus, age or ethnicity, and body habitus is the obvious risk because acoustic window quality varies with it and the product is aimed at operators least equipped to recognize a poor window. The 2020 and 2021 clearance summaries (K193518 and K212100) contain no AI performance data or subgroup analyses. The nearest published limitation comes from outside the company: an independent multicenter Japanese study reported that the automated system tended to underestimate ventricular volumes.

Nothing published says what happens when an automated measurement is wrong. No warranty attaches to a model output, no indemnity is described, and no error reporting or correction route is offered. What exists covers the hardware: a five year warranty and a US based service team, which protect a buyer whose probe fails and say nothing about an ejection fraction that reads 55 percent when the true value is 40.

The 2020 and 2021 clearance summaries contain no AI performance data: K193518 (March 2020) states that clinical performance data were not required, and K212100 (August 2021) states that the AI assisted ejection fraction workflow and the Trio tool were not yet cleared and were offered under FDA's COVID-19 enforcement policy for imaging systems. The company's prospective 78 patient study, adjudicated by echocardiography core laboratories, is where a vendor published error rate would sit. The gap weighs more because the marketed benefit is that novice operators reach reliable results, and named users include nurses and pre hospital staff.

Imaging interoperability is real, and the clinical record is not reached. The device exports through DICOM, supports a modality worklist through a configured imaging profile, offers direct file export over cable and screen casting, and publishes a software development kit and API page for partners; a June 2026 release added DICOMweb, the modern route into enterprise archives.

No HL7 interface, FHIR support or named EHR integration is published, and nothing describes an automated ejection fraction landing as a discrete, coded result in a chart. What the device produces reaches the imaging archive; whether it reaches the note depends on whether the customer has wired that archive into the chart. The gap costs more here because the output is a number a clinician wants in the record, for trending, quality reporting and billing.

The deployment is stated plainly: the models run on the device and studies stay on it. A network connection is needed for three named things, transducer registration, export to an imaging destination and software updates; scanning works offline, 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 there is no vendor cloud tenancy to question, and offline use makes field and global health deployment credible. Two things are open. The partner services offered on the platform, tele ultrasound streaming and automated report generation, are network services, and nothing published says where they are hosted or under what terms. And the technical data the company does collect (device identifiers and crash logs) has no stated storage location or retention period beyond what support and regulatory record keeping need.

One figure is published, and it is a ceiling rather than a price. The company states that its Kosmos Plus configuration (a full system with probes, AI, advanced Doppler and a 12.9 inch tablet) is available for less than $20,000, and its frequently asked questions route every other pricing question to a quote, explaining that price depends on configuration and location.

Alongside it the company publishes a five year warranty, US based service, ingress protection ratings and availability in more than 70 countries, which bear on total cost of ownership. Nothing published gives a probe level price or a configuration ladder, or says whether a recurring fee exists and what it unlocks, which decides five year cost for a handheld system.

Coverage is broad and specific, and validation stands behind part of it. Twelve clinical areas carry their own material: anesthesiology, cardiology, critical care, emergency medicine, family medicine and primary care, global health, hospital medicine, musculoskeletal, nephrology, nursing (bladder assessment and ultrasound guided IV access), obstetrics and gynecology, and vascular access.

The hardware supports the claim: a phased array probe for cardiac and abdominal imaging, a linear probe for lung, vascular access, musculoskeletal work and nerve blocks, and a bladder probe. Three platforms are supported (iOS, Android and a dedicated tablet, including a ruggedized Android tablet), the device is sold in more than 70 countries, and offline scanning makes resource limited use credible.

Validation is published for the automated ejection fraction in cardiology, oncology and hospital settings; the other areas are named without published validation of the AI features used there. The company's frequently asked questions page, last modified in June 2026, still said there was no obstetric preset four months after the release announcing cleared obstetric presets.