Ultromics

The model is the product, and it reports something a reader does not otherwise extract from the image. EchoGo Heart Failure applies a convolutional neural network to a single apical four chamber view and returns whether the echocardiogram shows heart failure with preserved ejection fraction, a form that is hard to diagnose. EchoGo Amyloidosis screens routine echocardiograms for cardiac amyloidosis, whose usual route to diagnosis runs through scintigraphy, cardiac MRI or biopsy.

The company also holds clearances for EchoGo Core, automated echocardiography measurement, and EchoGo Pro. Remove the models and there is nothing left to sell.

A diagnostic and screening aid with the clinician keeping responsibility, and with its uncertainty shown. EchoGo Heart Failure is indicated for adults over 25, on an apical four chamber view with an ejection fraction of 50 percent or more. It returns a binary result with an EchoGo Score from 0 to 100 percent plotted against a reference population, and returns no classification when a study is inadequate, as it did for 7.4 percent of the validation cohort. The labeling states that patient management decisions should not rest on the device alone and that diagnostic responsibility stays with the interpreting clinician. EchoGo Amyloidosis is cleared as a screening aid that flags patients who may benefit from further testing.

What is not published is the score threshold behind the binary result, or how a negative should be weighed against clinical suspicion, which matters for a test whose miss is a delayed diagnosis of a progressive disease.

The approach is named and the version history is visible in filings; the training data is not characterized. The K240013 summary describes EchoGo Heart Failure as a convolutional neural network producing a binary classification and a probability score, and lists what changed from version 1.0: an expanded training dataset, added preprocessing and data augmentation, new input and output interfaces, distributed deployment and a larger validation cohort. Six clearances across four products show versioning through FDA filings.

What the training data is, where it came from and its demographic makeup are not published.

A short chain at the model layer. EchoGo's models are the company's own, cleared under its own 510(k)s, so they cannot be swapped without a filing, and no outside model provider sits between the echocardiogram and the result.

The K240013 summary says the software is hosted on the company's platform or on third party infrastructure, and neither the hosting provider nor any subprocessor is named in the summary or the trust center.

Peer reviewed, external and multi site. The amyloidosis model was validated in a study published in the European Heart Journal, run with Mayo Clinic and University of Chicago investigators across more than 18 sites, on an external cohort of 2,719 patients with a 22 percent prevalence of the disease: an area under the curve of 0.93, 85 percent sensitivity and 93 percent specificity. For the heart failure product, the K240013 summary reports a retrospective case control study of 1,578 patients at multiple US sites, and supporting evidence is cited in the American Society of Echocardiography's 2025 diastology guideline. The company cites more than 25 peer reviewed studies across the platform.

Two qualifications. The heart failure validation is retrospective and case control, with a prevalence of about 50 percent, so it describes discrimination rather than performance in a screening population. And a guideline citing supporting evidence is not the same as one recommending the test by name. The company names Mayo Clinic, City of Hope, University Hospitals and UChicago Medicine among its users.

General assurances that do not answer the questions AI raises. The company states HIPAA compliance, ISO/IEC 27001 certification and GDPR alignment, and the K240013 summary says the software is hosted on the company's platform or on third party infrastructure, so echocardiograms leave the hospital for analysis. No retention period for studies, de-identification standard or position on using customer studies to train or improve models is published.

The company has described building its models on a large echocardiography dataset, which makes the source of that data a live question: whether it was assembled under research arrangements separate from commercial use, or whether customer studies contribute, is unstated in either direction.

The company states HIPAA compliance on its home page, and its trust center lists HIPAA with a HIPAA report available on request. No business associate status or agreement terms are published, so the compliance claim stands without the underlying document.

The trust center also lists GDPR and the NHS Data Security and Protection Toolkit, reflecting the company's UK base.

A trust center exists and names its certifications, with the artifacts behind a request. The company's trust center, linked from its site footer, lists ISO/IEC 27001:2022 and ISO 13485:2016, both available on request, along with a HIPAA report and a penetration test report on the same terms, and lists GDPR and NHS Data Security and Protection Toolkit status.

Six FDA 510(k) clearances, all in Ultromics' name. EchoGo Heart Failure: K222463 (November 2022) and version 2.0 as K240013 (September 2024), for detecting heart failure with preserved ejection fraction in adults over 25 from an apical four chamber view with an ejection fraction of 50 percent or more (product code QUO, 21 CFR 870.2200). EchoGo Amyloidosis 1.0: K240860 (November 2024), cleared as a screening aid (product code SDJ, 870.2200). EchoGo Core, automated echocardiography measurement: K191171 (2019) and version 2.0 as K213275 (2021). EchoGo Pro: K201555 (December 2020).

The company states that both disease detection products carry Breakthrough Device designation, and holds ISO 13485:2016 certification.

The amyloidosis validation reports its results across the groups that matter most clinically, as conclusions rather than a full table. The study published in the European Heart Journal describes a multi ethnic external cohort from more than 18 sites and reports that performance held across the major subtypes of the disease, and that the model distinguished amyloidosis from the conditions that mimic it on an echocardiogram: hypertensive heart disease, heart failure with preserved ejection fraction and hypertrophic cardiomyopathy. That discrimination decides whether a screen raises false alarms in the patients who present most similarly.

The K240013 summary for EchoGo Heart Failure reports no subgroup results, and no performance is published by ethnicity, sex or site for either product. That matters because the hereditary transthyretin variant most common in the United States is carried disproportionately by people of West African descent, so a model whose sensitivity varied by ancestry would miss disease where one subtype concentrates.

The company's own clearance filing for its lead product publishes error rates with method and denominator, including the less flattering version. The K240013 summary for EchoGo Heart Failure 2.0 (September 2024) reports a retrospective case control study at multiple US sites, 1,578 patients (793 with heart failure with preserved ejection fraction, 785 controls): sensitivity 90.3 percent (95 percent confidence interval 88.5 to 92.4) and specificity 86.1 percent (83.4 to 88.3) excluding studies the device declined to classify, and 84.9 and 78.6 percent counting them, with 7.4 percent of studies (116) returned as no classification. The home page states the first pair and, for EchoGo Amyloidosis, 84.5 percent sensitivity and 89.7 percent specificity.

These are test set results in a cohort with a prevalence of about 50 percent, far above what a screening laboratory sees, so positive predictive value in use will be lower. No warranty, indemnity or remediation commitment is published.

Built to sit in the existing reading workflow. EchoGo accepts DICOM echocardiograms, and version 2.0 added application interfaces for input and output. Studies are analyzed with no change to acquisition, and the company states that results return into existing clinical systems, typically within 20 minutes, so the finding is available during routine interpretation rather than after the report is signed. The company names Mayo Clinic, City of Hope, University Hospitals and UChicago Medicine as users.

No named record system, echo reporting system or image archive integration is published.

A hosted service with its location unstated. The K240013 summary says EchoGo is hosted on the company's platform or on third party infrastructure, with distributed deployment added in version 2.0 for scalability and robustness. Studies are sent from the echocardiography laboratory for analysis, and the company states that results return typically within 20 minutes.

Which third party, in which region, and how long studies are kept are not published, and the trust center lists no hosting or residency detail.

The payer's side is published and the vendor's is not. The company names CPT Category III code 0932T for EchoGo Heart Failure and payment under APC 5743, and has described a new technology add on payment for inpatient use. Those are what an insurer pays the provider, not the vendor's price.

No license fee, per study charge or platform fee is published, and whether EchoGo Amyloidosis has an equivalent payment pathway is not stated.

Deliberately narrow, and validated in what it covers. Ultromics appears among imaging AI vendors because its products read routine echocardiograms. Coverage is one modality for two conditions: heart failure with preserved ejection fraction, in adults over 25 with a preserved ejection fraction, and cardiac amyloidosis as a screening aid, used in echocardiography laboratories and cardiology services.

The heart failure validation spans multiple US sites and the amyloidosis validation more than 18 sites; performance by setting, such as hospital laboratories against outpatient practices, is not published.