Anumana

The algorithm is the entire company. There is no device, no laboratory and no service layer, only neural networks trained on raw electrocardiogram signal and sold as software as a medical device.

What makes the centrality unusually clean is that the models are not automating a task a human already performs. A cardiologist reading a trace is not attempting to estimate ejection fraction or detect amyloid deposition, because the signal that supports those conclusions is not visible to human reading. The product is therefore not a faster version of human interpretation but a different capability applied to the same input, which is the strongest form this axis measures.

Low autonomy by design, and correctly so. Each algorithm returns a risk flag on a trace that a clinician then acts on, typically by ordering the confirmatory test the flag suggests, an echocardiogram for suspected low ejection fraction or the workup pathway for suspected amyloidosis. Nothing is treated or diagnosed by the software.

The oversight detail that would raise this grade is not in public material. Operating points and performance characteristics for a cleared device live in its labelling, so the information exists in a public regulatory document, but the company does not surface the threshold, the expected false positive rate at that threshold, or what a clinician is instructed to do with a positive flag in a patient with no symptoms.

Better than most of this index on the dimension that matters here, which is training data provenance and scale. The company publishes that the platform draws on more than 22 million patient records spanning over 20 years from academic medical centre partnerships, and that the flagship model was developed from roughly 2.9 million paired electrocardiogram and echocardiogram studies across more than 676,000 patients. Publishing the denominator of a training set is rare and it is exactly what lets an outside reader judge whether a claim is plausible.

Held below A because the architecture is described only as neural networks, because per algorithm performance characteristics are not collected in one accessible place, and because the phrase used about the training population, that it was diverse, is an assertion rather than a breakdown.

The corpus source is named at institution level and its scale is published, which is more than most, and the arrangement behind it is the most consequential fact about this product. The models were built on more than twenty two million patient records held by academic medical centres, and one of those institutions co founded the company and holds equity in it.

So the patients whose traces trained these algorithms were receiving care rather than participating in a commercial venture, and the resulting product is sold back into healthcare at a price, with the institution that supplied the data holding a stake in the proceeds. That arrangement is lawful, increasingly common, and arguably the best available use of data already collected, and it is also the thing a reader most needs to see stated.

Nothing published describes what those patients were told, what governance sat over the transfer, whether an ethics committee reviewed it, or whether any patient could have declined. No specific position on customer data handling, retention or model improvement was located either, so a health system deploying the product cannot establish what happens to the traces it generates, which is the same question one institution has already answered in its own favour. Ask what governed the original transfer, and whether traces from deploying customers contribute to future model development.

The strongest evidence position of any electrocardiogram artificial intelligence vendor in this index, and one of the strongest anywhere in it, because it rests on randomised evidence rather than on retrospective performance.

The underlying low ejection fraction algorithm was tested in a randomised trial in routine practice which found roughly 32 percent more diagnoses of low ejection fraction than standard care, a comparison against usual practice rather than against a held out test set. The pulmonary hypertension algorithm has a peer reviewed publication in the European Respiratory Journal. Three separate clearances each required their own evidence package, and the products are stated to be eligible for reimbursement, which is a further independent assessment of clinical value. An outside review of Mayo Clinic's artificial intelligence portfolio characterises this as one of its most evidence backed applications rather than a pilot.

Graded on an honest basis. No specific published stewardship position on customer data handling, retention or model improvement was located in this pass.

The question this record raises is a different one and it belongs on the axis. The models were built on more than 22 million patient records held by academic medical centres, and one of those institutions co founded the company and holds equity in it. The patients whose traces trained these algorithms were receiving care, not participating in a commercial venture, and the resulting product is sold back into healthcare at a price. That arrangement is lawful, increasingly common and arguably the best use of data already collected, and it is also the single most consequential fact about how this product came to exist. Nothing published describes what those patients were told or what governance sat over the transfer.

Graded on an honest basis and flagged for re verification. No published compliance statement, business associate agreement posture or privacy documentation attributable to the company was located in this pass.

One related datapoint sits with the parent rather than the subsidiary: nference has publicly described its de identification approach for the underlying data platform as exceeding the statutory requirement. That is a claim about the research data pipeline, not about how the commercial product handles a customer's patients, and the two should not be conflated.

A dedicated search for a trust centre, SOC 2 report, HITRUST or ISO 27001 certification returned nothing attributable to this company.

The honest qualification is that a cleared device manufacturer carries an assurance obligation that a software vendor does not, since marketing a class II device requires a quality management system subject to inspection, and the regulatory record shows that obligation is being met three times over. That is real third party oversight, but it covers design controls and manufacturing quality rather than information security, and a hospital security review asks a different question. Graded C for the absence of any published security posture, not for an absence of external scrutiny.

The deepest regulatory position in the electrocardiogram artificial intelligence segment. Three clearances, two of them firsts of their kind: pulmonary hypertension in March 2026 and cardiac amyloidosis in April 2026, both the first cleared for those indications on a standard 12 lead trace, alongside the earlier low ejection fraction algorithm which is available in the United States and the European Union.

Two further signals matter. Multiple algorithms have carried Breakthrough Device Designation, which this index treats as a development pathway rather than an approval, but here the designations converted into actual clearances, which is the outcome that designation is supposed to produce and frequently does not. And the amyloidosis product was selected among the first 15 devices in the agency's Total Product Life Cycle Advisory Programme pilot, meaning the regulator chose this company for a closer and more continuous form of engagement than the ordinary review process.

The training cohort is described as spanning diverse populations, and the cohort size is published, which is more than most vendors offer. No actual subgroup performance breakdown was located, so the diversity claim cannot be checked and no reader can tell where the models degrade.

The concern is specific to what these algorithms do. Detecting structural heart disease from an electrocardiogram means learning the relationship between an electrical trace and cardiac anatomy, and that relationship is affected by body habitus, chest wall structure, lead placement and sex. Ejection fraction and amyloidosis both have documented differences in presentation and diagnosis rates across sex and race, and a model trained largely at a small number of academic centres inherits whatever the referral patterns of those centres encoded. Publishing the cohort size without the breakdown answers the easier half of the question.

Publishing the denominator of a training set is rare and it is what carries this grade. The company states that the platform draws on more than twenty two million patient records spanning two decades from academic medical centre partnerships, and that the flagship model was developed from roughly two point nine million paired cardiac studies across more than six hundred and seventy thousand patients.

Giving both the study count and the unique patient count is the useful form, because the ratio tells a reader how many studies each patient contributed and therefore how much independent information the corpus actually holds, which a single headline number conceals. That is exactly what lets an outside reader judge whether a performance claim is plausible rather than taking it on trust. Held at C on three points.

The architecture is described only in the most general terms, per algorithm performance characteristics are not collected in one accessible place so a buyer evaluating a specific algorithm has to assemble the picture themselves, and the description of the training population as diverse is an assertion rather than a breakdown.

Diversity matters here in a measurable way, because cardiac signal characteristics vary with sex, age and body habitus, so a stated composition would be a performance disclosure rather than a fairness gesture. No warranty, indemnity or remediation commitment attaches. Ask for the training population composition, and per algorithm sensitivity and specificity with their validation cohorts.

Independent review literature describes the commercial electrocardiogram artificial intelligence platforms, this one included, as integrating through standard health data messaging and interchange protocols, and adoption at academic health centres is consistent with real workflow integration rather than a separate portal.

The delivery problem here is genuinely harder than for most software, because the input is a signal file produced by electrocardiogram carts from several manufacturers and the output has to reach a clinician inside a cardiology or primary care workflow. A related venture was created by the same parent specifically to handle the plumbing of getting device data to algorithms and results back to the point of care, which is an admission of how much of the difficulty sits there. Named integrations and standards level detail were not located.

Not described in public material. No hosting model, region, on premise option or retention schedule was located.

The architectural question a buyer should ask is where inference happens, because an electrocardiogram is a small file and a health system may reasonably expect the algorithm to run locally rather than sending every trace off site. Availability in both the United States and the European Union implies the residency question has been answered somewhere for the European market, since the same posture would not satisfy both, and that answer is not public.

No price, rate card or pricing mechanism is published, and the route to information is a demonstration request.

One real disclosure lifts this above the floor: the company states that its cleared algorithms are eligible for reimbursement, which tells a buyer something material about how the product is paid for and is more than most software vendors offer. Reimbursement eligibility means the economic question for a health system is coverage and coding rather than a licence fee, and that is a genuinely different purchase. The specific codes, rates and coverage conditions are not published, so the buyer has the shape of the answer without the numbers.

The reach is wide because the input is ordinary. An electrocardiogram is performed in primary care, emergency departments, preoperative assessment and cardiology alike, so an algorithm operating on a standard 12 lead trace can screen opportunistically anywhere the test is already being done, without adding a step or a device.

Indication coverage spans low ejection fraction, pulmonary hypertension and cardiac amyloidosis, three conditions that are underdiagnosed for the same reason, namely non specific early symptoms. A second and different buyer exists in pharmaceutical companies, with named algorithm development collaborations, where the commercial logic is finding undiagnosed patients for a therapy rather than improving care in a health system. Held at B because everything sits within cardiovascular medicine and the trace itself bounds what is reachable.