Eko Health

The AI is the reason a health system buys SENSORA, and the stethoscopes sell without it. Eko's digital stethoscopes offer amplified audio, noise cancellation, waveforms and recording on a free software tier; the cleared models for structural murmur, atrial fibrillation and low ejection fraction come with the paid Eko+ membership for individual clinicians and the SENSORA enterprise license. SENSORA runs only on Eko hardware (the CORE 500 and the Littmann CORE). The models are the engine of the detection product, and the company carries a hardware business underneath that stands without them.

What the software may and may not do is published, with the stop conditions and the human gate. The SENSORA page states that physician or qualified professional review is required for every analysis before clinical use, that the software is not a sole means of diagnosis, and that it does not identify arrhythmias other than atrial fibrillation, so a clear result does not mean a normal rhythm.

The cleared software grades its own inputs: a signal quality classifier withholds murmur analysis on a poor recording, and the rhythm model returns unspecified rhythm or poor ECG signal instead of forcing a call. Murmur results come as high or moderate likelihood, each with a published positive predictive value, and atrial fibrillation detection is limited to adults. Billing the Medicare code requires a documented provider over-read, so the review step is tied to payment.

The models and their validation data are described; the update practice is not. The low ejection fraction model is a convolutional neural network over single-lead ECG and heart sound (JACC Advances, 2025), and its clearance summary characterizes the training data (1,852 patients from seven US sites and one in India) and the separate test set. The 2025 analysis software is described in its clearance summary as deep neural networks with transformers, with five named components from a signal quality classifier to heart rate calculation, and a head-to-head against the model it replaced.

Two things are unresolved. The SENSORA page says the low ejection fraction model was trained on more than 100,000 ECG and echocardiogram pairs, which the clearance summary's training description does not match, and nothing reconciles the two. And the software runs in Eko's cloud, where the company describes a newest AI model without a version number, release notes or a statement of when customers are moved to it.

The models are Eko's own and cleared, so the chain is short at the model layer. The analysis software, the low ejection fraction tool and the murmur and rhythm models are Eko products with their own clearances, and the 2025 clearance summary names the storage layer: recordings sit in Eko's cloud on Amazon Web Services S3 and are read by the analysis server.

No subprocessor list is published; the security page says Eko holds business associate agreements with subcontractors and hosting providers without naming them, and the privacy policy mentions third parties that may process data outside the United States.

A randomized trial is published, and on its primary endpoint the tool did not increase detection. TRICORDER (The Lancet, 2026) randomized 205 UK primary care practices; by intention to treat, heart failure diagnoses did not differ between arms after twelve months (incidence rate ratio 0.94, 95 percent CI 0.86 to 1.02). Use of the stethoscope was associated with more heart failure, atrial fibrillation and valve disease diagnoses, and those are the figures on the SENSORA page.

An independent evaluation at Oregon Health and Science University (International Journal of Cardiology, 2025) found the earlier murmur model detected valve disease with 39.3 percent sensitivity and 82.3 percent specificity against echocardiography in 1,029 patients. A prospective primary care study (European Heart Journal Digital Health, 2026) found 92.3 percent sensitivity for moderate or worse valve disease against 46.2 percent for clinicians' own auscultation, with lower specificity. The low ejection fraction model was validated in 2,960 patients at four US networks (JACC Advances, 2025; area under the curve 0.85, positive predictive value 20.3 percent). A Mayo Clinic-led randomized trial in Nigeria (Nature Medicine, 2024) detected more left ventricular dysfunction in pregnancy with an AI-enabled digital stethoscope (24 of 587 against 12 of 608). A 2026 veterinary study concerns animals rather than patients.

The data terms are specific, and they run toward the company. The enterprise terms and conditions give Eko a perpetual, irrevocable, royalty-free license to use customer data in de-identified or aggregated form for any purpose, naming the development of its algorithms, and assign it all rights in aggregated statistics. The business associate agreement lets Eko de-identify any and all protected health information under the HIPAA standard, after which the agreement no longer applies, and the privacy policy says recordings are used to train the algorithms and that anonymized data is kept indefinitely. Identifiable data is kept for the life of an account and up to five years after, and is encrypted at rest.

A heart sound and ECG pair is a physiological signal, and whether removing identifiers from it prevents re-identification is not addressed. Eko publishes nothing on safety engineering, meaning guardrails, hallucination mitigation or how a safety event is handled; the nearest thing is a signal quality check that returns a poor signal result rather than a reading.

The business associate agreement is published and readable before contracting. It applies to any United States customer that is a HIPAA covered entity and takes effect on the customer's acceptance, and it binds Eko to flow down substantially similar restrictions to any subcontractor that receives protected health information; the security page adds that Eko holds such agreements with its subcontractors and hosting providers.

Two terms deserve reading before signing: Eko may de-identify any and all protected health information, after which the agreement stops applying, and it may amend the agreement on notice. The agreement does not apply outside the United States, and no equivalent terms are published for the UK, where the company's largest trial ran.

A recognized attestation is named without the report. The company's security and HIPAA page states that Eko is SOC 2 certified, with no type, period, auditor or report, and describes AES-256 encryption with managed keys, daily encrypted backups and a security contact; the telehealth page repeats SOC 2 compliance. No trust portal or downloadable report is offered.

The device side has outside review: the 2024 and 2025 software clearance summaries record cybersecurity testing as part of FDA review. Nothing says whether the SOC 2 scope covers the cloud that stores recordings.

Each model is cleared separately and identifiable in the FDA database. K192004 (January 2020) cleared the original analysis software for murmur and atrial fibrillation in adults; K213794 (June 2022) the structural murmur software for adults and children; K233409 (March 2024) the low ejection fraction tool for adults at risk of heart failure, with Anumana's ECG algorithm as predicate; and K251494 (August 2025) the transformer-based analysis software that now covers structural murmur and timing for adults and children and atrial fibrillation for adults. Stethoscope hardware clearances run from 2015 to the CORE 300 in July 2026. The 2025 software clearance carries no predetermined change control plan.

The SENSORA page's description still repeats the 2020 indication wording (adults only, QRS duration and EMAT), so it does not show which clearance covers the software in use.

Pre-specified subgroup results are published for the cleared models, and they show where performance drops. The low ejection fraction clearance summary (K233409) reports results across four age bands, sex, six body mass index bands, race and ethnicity, and conduction disorders: specificity fell to 70.7 percent over age 70 against about 81 percent at 31 to 70, was 73.6 percent in men against 81.8 percent in women, and was lower in patients with pacemakers, atrial fibrillation or wide QRS complexes, with no differences by body mass index or race. The 2025 murmur and rhythm clearance summary (K251494) reports murmur specificity of 76.0 percent in children against 93.9 percent in adults, and no differences by sex or race.

The race comparisons for the murmur and rhythm models rest on validation sets that were 90 and 87 percent White, and results by race are stated as no significant difference rather than shown. An independent evaluation at Oregon Health and Science University (International Journal of Cardiology, 2025) found the earlier murmur model's sensitivity varied widely by valve lesion.

Error rates are published with their method and denominators, and nothing in the terms answers for them. The clearance summaries give sensitivity, specificity and predictive values with confidence intervals against echocardiogram or expert reference: 74.7 percent sensitivity and 77.5 percent specificity for low ejection fraction in 3,456 patients, 83.4 and 86.0 percent for structural murmur, and 94.7 and 94.1 percent for atrial fibrillation. A buyer can hold the product to those figures.

The enterprise terms disclaim any warranty that results are accurate, require the customer to indemnify Eko, and cap Eko's liability at three months of fees; the individual terms cap it at $100. The detection guarantee on the website is a refund on the stethoscope, not a commitment on the output.

EHR integration is listed as a SENSORA feature with no system named. The plan comparison and the SENSORA page list EHR integration and EHR customization for enterprise customers, and the telehealth product offers SDKs and APIs for streaming into partner video platforms; no record system, marketplace listing, interface standard or reference deployment is named.

The product's Medicare billing needs a documented overread, and nothing published says how results and that review reach the chart.

Residency is stated, with isolation and the onward path left open. The privacy policy says the services are hosted in the United States, and the 2025 clearance summary describes the analysis software as cloud based, reading recordings from Eko's cloud storage on Amazon Web Services S3 and writing results back. SENSORA licenses are offered in the United States only.

The privacy policy also says data may be processed by third parties in other locations, without naming them, and tenant isolation is not described. There is one deployment option, the hosted cloud, reached through the mobile and desktop apps and an API.

Prices are published for the individual route and withheld for health systems. The plan comparison lists a free Eko Standard tier, Eko+ at $119.99 a year per clinician with murmur and atrial fibrillation detection, and a veterinary tier at $180, all bought online; the CORE 500 stethoscope is sold in the store at $449 to $499. SENSORA, which adds low ejection fraction detection, administration, EHR work and on-site training, is an annual enterprise license priced through sales.

On the payment side, the company states that CPT code 0962T covers SENSORA analyses and that CMS set outpatient payment under APC 5734 in December 2025, with physician fee schedule coverage varying by carrier; that is payer context rather than what a health system pays.

Coverage is named, with validation behind the primary care use. The clinical scope is cardiac detection: structural murmur for adults and children, atrial fibrillation for adults, and low ejection fraction for adults at risk of heart failure, with capture by medical assistants and nurses at intake. Primary care is validated in the UK trial and the prospective US valve study, and each cleared model has its own validation population.

The company also markets SENSORA for emergency departments, citing one rural hospital, and for telehealth examination, neither with published validation. SENSORA licenses are sold in the United States only.