Eko Health

The algorithms are unambiguously the product of SENSORA, and they carry their own clearances, but they are inseparable from Eko's hardware business and that is the buyer relevant fact. SENSORA analyses data captured by an Eko digital stethoscope; it is not device agnostic. The stethoscopes themselves sell as functional products without any AI, offering digitally enhanced audio, active noise cancellation, waveform display and exam storage, so there is a real non AI business underneath.

That coupling is the sharpest contrast with AccurKardia, indexed here, where the algorithm is deliberately device agnostic and the model is the whole product. Same structural distinction this index draws in pathology between scanner coupled and platform independent clearances. Graded B: the AI is genuinely the differentiator and the reason a health system buys SENSORA, and it also cannot be bought on its own.

The clearest oversight disclosure on this sourcing list and one of the better ones in the index, because Eko publishes three things most vendors publish none of. First, a mandatory human gate stated plainly: physician or qualified healthcare professional review and interpretation is required for every SENSORA analysis prior to clinical use, and the software is intended to support the physician rather than serve as a sole means of diagnosis.

Second, an explicit NEGATIVE scope statement, that the software does not identify arrhythmias other than atrial fibrillation, which tells a clinician what the absence of a flag does not mean. That is the disclosure most detection vendors omit and the one that prevents a false sense of coverage.

Third, and unusually candid, the company states that billing under CPT 0962T requires a documented provider over read, which ties the oversight requirement to the economics rather than leaving it as guidance. A buyer cannot claim reimbursement without performing the review. Publishing the condition that constrains your own revenue is the behaviour this axis is meant to reward.

Materially more disclosed than most cleared algorithm vendors and short of the top grade only on method. Eko publishes the training basis for its low ejection fraction algorithm, more than 100,000 paired electrocardiogram and ultrasound examinations, and it publishes the resulting performance: approximately 74.7 percent sensitivity and 77.5 percent specificity.

Those are honest rather than flattering numbers, meaning roughly one in four cases with reduced ejection fraction will not be flagged, and publishing them rather than a rounded accuracy figure puts Eko on this index's roster of vendors that publish the unflattering half alongside AZmed and VUNO. The company also states which signals each output derives from and what intervals are computed.

Held at B because no model card, architecture description or evaluation protocol is published, and the performance disclosure is detailed for low ejection fraction but not equivalently stated for the murmur and atrial fibrillation algorithms.

The company states openly that its cardiac foundation model was trained on more than four million de identified heart sound and electrocardiogram recordings gathered across diverse care settings, so patient recordings demonstrably feed model development at very large scale and the company says so rather than leaving it to inference.

Most vendors in this position disclose nothing, and recordings are also stored in the company's cloud with sharing for second opinions marketed as a feature, so retention is a product capability rather than an incidental by product. What the statement surfaces rather than settles is everything a buyer needs: whose recordings, whether routine clinical use at customer sites contributes, whether a health system can decline, and what de identification means for this signal type.

That last question is not trivial, because a heart sound and electrocardiogram pair is a physiological signature and the cardiac cycle carries individually distinctive features, so removing a name from a file is not obviously the same as making the recording non identifying. One practical feature of the workflow compounds it.

Capture happens during intake by medical assistants and nurses rather than by the interpreting clinician, so the person collecting the recording is often not the person who could explain to a patient what becomes of it, and the consent conversation, if it happens, happens with the wrong staff member. Ask whether customer recordings enter training corpora and under what agreement, the de identification method, retention, and what a patient is told.

Among the strongest evidence records in this index, and the only one on this sourcing list with a randomised controlled trial. Three separate publications in major journals, none of which is a vendor white paper. A randomised controlled trial spanning more than 200 NHS primary care practices was published in The Lancet and is described as the largest randomised trial of AI in cardiology conducted to date, which is the right study design for a screening tool because it tests whether earlier detection actually happens in routine practice rather than whether an algorithm scores well retrospectively.

An independent clinical validation in Circulation found the structural murmur algorithm doubled sensitivity for structural heart disease compared with an analogue stethoscope, a direct head to head against the incumbent method. A Lancet Digital Health study of more than 1,000 patients addressed low ejection fraction detection in primary care specifically. Real world deployment corroborates the research: more than 100 UK clinics through the NHS and Imperial College London.

The low ejection fraction algorithm was developed with Mayo Clinic and a pulmonary hypertension algorithm with Brown University Health System's Cardiovascular Institute. The one caution worth recording: several of these studies involve institutional partners who are also collaborators, which is normal in device development but means the evidence is not wholly independent of the company.

Converted from Not Rated after a second search, which answered part of the question and sharpened the rest.

What is now known. Recordings are stored in the company's own cloud, and the platform markets storage and sharing of exams for second opinions as a feature, so retention is a product capability rather than an incidental by product. More significantly, the company states that its cardiac foundation model was trained on more than four million de identified heart sound and electrocardiogram recordings gathered across diverse care settings. Patient recordings therefore demonstrably feed model development at very large scale, and the company says so openly, which is more than most vendors in this position disclose.

What that surfaces rather than settles. The statement establishes that de identified recordings train models; it does not state whose recordings, whether routine clinical use at customer sites contributes, whether a health system can decline, or what de identification means for this signal type. That last question is not trivial. A heart sound and electrocardiogram pair is a physiological signature, and the cardiac cycle carries individually distinctive features, so removing a name from a file is not obviously the same as making the recording non identifying.

Capture also happens during intake by medical assistants and nurses rather than by the interpreting clinician, so the person collecting the recording is often not the person who would explain to a patient what becomes of it.

Ask whether customer recordings enter training corpora and under what agreement, what de identification is applied, the retention schedule, and what a patient is told.

Converted from Not Rated. No business associate statement, scope description or contracting detail was located for the United States, and nothing equivalent for the United Kingdom. The prior note's framing holds: the company sells into United States health systems and operates across the National Health Service, so it is inside two regimes at once and publishes a position on neither.

The two regimes ask different questions and both are live for this product.

In the United States the arrangement is straightforward in principle. Recordings are captured during a clinical encounter and stored in the company's cloud, which makes the company a business associate of the practice or health system, with the usual questions about permitted uses, subprocessors and termination. The permitted use question carries unusual weight here because the company has stated that millions of de identified recordings train its models, and whether the agreement authorises that conversion is the single term a health system should read most carefully.

In the United Kingdom the framework is different in kind rather than in detail. Health data is special category data requiring an identified lawful basis, the common law duty of confidentiality applies alongside the statute, and use of patient data for product development typically requires either consent or a formal approval route. A large randomised trial across primary care practices means substantial exposure to that regime.

Ask which entity contracts in each territory, what the agreement permits by way of model development, and what approvals cover any secondary use of recordings collected in the United Kingdom.

Converted from Not Rated. No commercial attestation, trust centre or security page was located. The prior note correctly separated device quality system obligations from information security, and a further distinction now applies that lifts this above the absence.

Since 2023, a submission for a networked software device must include a cybersecurity plan, a software bill of materials and a post market security update process. This company's portfolio straddles that date, which produces a result worth stating precisely rather than generically. Its algorithms for atrial fibrillation and structural murmur detection were cleared in 2020 and carry no such review. Its low ejection fraction algorithm was cleared in 2024, and its cardiac foundation model in 2025, both after the requirement took effect, so those submissions will have carried cybersecurity documentation examined by the regulator.

That is a useful general rule for this index. A vendor's clearances can sit on both sides of the cybersecurity requirement, so which algorithm a customer is buying determines whether that review sits behind it. Asking whether a company is cleared is not sufficient. Ask when, and for which component.

What remains absent is any examination of the organisation rather than the products. The platform stores heart sound and electrocardiogram recordings in the company's cloud, and that estate, its access controls and its development practices are not covered by a device submission.

Ask for the cybersecurity documentation for the recently cleared components, and separately whether any organisational attestation covers the cloud platform.

A decade of stacked clearances rather than a single milestone, each tied to a specific stated indication. The digital stethoscope was first cleared in 2015, followed by algorithms for structural heart murmur and atrial fibrillation, the Duo combined ECG and stethoscope device, and in 2024 the Low Ejection Fraction tool, developed in collaboration with Mayo Clinic and cleared for detecting an ejection fraction at or below 40 percent in roughly fifteen seconds during a routine examination.

What lifts this to A rather than simply recording the clearances is the quality of the indication language Eko publishes alongside them: the detection thresholds are numeric, the computed intervals are named, the limitation that other arrhythmias are not identified is stated, and the requirement for professional over read before clinical use is explicit. Read the individual algorithm clearances rather than treating SENSORA as a single cleared entity, since the platform is a collection of separately cleared functions.

No fairness, subgroup or population performance disclosure was located, and the published sensitivity figure makes the gap concrete rather than theoretical. At roughly 74.7 percent sensitivity for low ejection fraction, approximately one in four affected patients will not be flagged, and the question that matters is whether those misses are distributed evenly.

They plausibly are not: heart sound acquisition and ECG morphology vary with body habitus, chest wall thickness, breast tissue and comorbidity, and auscultation quality has documented variability across patient populations. No breakdown of performance by sex, body mass index, ethnicity or age is published for any of the three algorithms. This matters more than usual because of where Eko aims the product.

The company explicitly positions SENSORA for underserved communities whose primary care offices lack ready access to echocardiography, which is a genuinely good use of a low cost screening tool and also means any systematic underperformance would land on exactly the population the product is meant to reach.

The company publishes the training basis for its ejection fraction algorithm at over a hundred thousand paired examinations and publishes the resulting performance, and the numbers are honest rather than flattering: the stated sensitivity means roughly one in four cases with reduced ejection fraction will not be flagged.

Publishing that rather than a rounded accuracy figure puts this vendor on the short roster in this index that discloses the unflattering half, and the disclosure is more useful than a better number would have been, because a clinician who knows the miss rate keeps their own suspicion active while one told the tool is accurate does not. The company also states which signals each output derives from and what intervals are computed, so a reader can tell what the algorithm is looking at.

Held below the top grade on two points. No model card, architecture description or evaluation protocol is published, and no warranty, indemnity or remediation commitment attaches. And the disclosure is uneven across the portfolio: performance is detailed for one algorithm and not equivalently stated for the murmur and rhythm algorithms, which matters because a buyer who has seen the rigour applied to one will reasonably assume it extends to the others.

Rigour in one place is evidence of capability, not of coverage. Ask for equivalent sensitivity and specificity on the murmur and rhythm algorithms, with their training bases and validation populations.

No electronic health record integration, standard or named platform was located. SENSORA is positioned as an enterprise platform operating within the patient intake workflow and generates results that must reach the record and, given the CPT code, the billing system, but no integration method, EHR partner or reference deployment is documented publicly. For a health system evaluating deployment at intake across many sites, how results and the required provider over read are captured in the record is a first order question with no published answer.

A hardware plus cloud platform deployed at the point of care, usable in physical and virtual settings including telehealth. No hosting provider, region, tenancy model or data residency commitment is published. The gap is worth closing given the company operates at scale in both the United States and the United Kingdom, where NHS deployment across more than 100 clinics would ordinarily carry explicit data residency requirements.

Better than most because Eko discloses the reimbursement side, which is the half of the equation that usually stays hidden. The company publishes that the AMA granted a Category III CPT code, 0962T, for its AI detection algorithms in November 2024, and states the condition attached to billing it, namely a documented provider over read.

Naming the code and its condition lets a practice model the revenue side and understand the work required to claim it, and it follows the pattern this index credits in HeartSciences for naming its APC code. Held at B rather than A because no product pricing is published at any level: neither the stethoscope hardware, the SENSORA platform licence, nor whether pricing is per device, per clinician or per analysis. Note also that Category III is a temporary tracking code with no established payment rate, so a buyer should not read the code as revenue.

Deliberately narrow clinically and genuinely broad operationally. Clinical scope is cardiac detection only, covering low ejection fraction, atrial fibrillation and structural murmur, with a pulmonary hypertension algorithm in development with Brown University Health System.

Operationally the reach is wider than most cleared devices manage: front line primary care is the primary setting, capture can be performed by medical assistants and nurses rather than requiring a physician, the platform supports both in person and telehealth encounters, and deployment spans enterprise United States health systems and more than 100 NHS clinics in the United Kingdom. The workflow point is the strategic one, since a screening tool that only a physician can operate does not scale to every encounter and this one is designed not to require that.