Anumana

The models are the product. Anumana sells software that reads a standard 12-lead ECG and returns a finding a clinician cannot see on the trace: low ejection fraction, pulmonary hypertension or cardiac amyloidosis, each a neural network cleared as machine learning notification software. There is no device, laboratory or service layer underneath.

Its newer WatchMate software for left atrial appendage closure planning, cleared in June 2026, is also a deep learning pipeline, from frame classification and segmentation to a 3D mesh and landing zone estimate. Remove the models and nothing is left to sell.

What each model may not do is published; the instructions behind it are held back. The indications pages state that the low ejection fraction model is not a stand-alone diagnostic, is not for monitoring and must not be used on paced rhythms; that a positive result suggests further evaluation; and that in a high risk patient a negative result should not rule out further non-invasive testing. The pulmonary hypertension and amyloidosis indications carry the same structure. Results arrive in the record as a binary finding, and the clearance summaries give the predictive values behind it.

The instructions for use are released only through a contact form, and nothing published says how the software handles a poor-quality trace. The validation paper's suggestion that a negative result could defer echocardiography sits uneasily with the label's warning against ruling out on a negative result in high risk patients.

The models, their training data and the rules for changing them are published. The clearance summaries describe each as a neural network reading 10 seconds of 12-lead voltage data: the low ejection fraction model was developed from 93,722 Mayo Clinic patients with an ECG and echocardiogram within two weeks, and the amyloidosis model from 4,251 patients with the disease and 45,730 matched controls. Each carries an FDA-cleared change control plan: a retrained version must be validated in a multi-center study that adds new data (at least 20 percent for the ejection fraction and amyloidosis models, 10 percent for pulmonary hypertension), must match or beat the prior version's sensitivity and specificity, and users are told through revised labeling or release notes.

The about page describes the platform as modeled on nearly 2.9 million ECG-echocardiogram pairs from more than 676,000 patients, a larger figure than the clearance summary's training set, without saying how the two relate.

A short chain: the models are Anumana's own, cleared, and run in the customer's environment. The ECG-AI page places processing on a health system virtual machine or inside its ECG management system, and each model's change control plan governs how it may be replaced, so no outside model sits between the trace and the result. The training data source is named: research-authorized Mayo Clinic records, through the company's founding partnership with Mayo Clinic and nference.

No subprocessor list is published, and nothing says whether support, telemetry or the new validation data the change control plans require involve customer data or third parties.

A randomized trial of the approach is published, with independent-site validation behind the cleared product. In the EAGLE trial (Nature Medicine, 2021), 120 Mayo Clinic primary care teams were cluster-randomized; teams given the AI-ECG result diagnosed new low ejection fraction (50 percent or below in the trial) in 2.1 percent of patients against 1.6 percent without it, across 22,641 adults (odds ratio 1.32). That trial tested Mayo Clinic's research version of the model, before Anumana's clearance. The cleared software was then validated at four US health systems outside Mayo (JACC, 2026; 13,960 patients, area under the curve 0.92), and the pulmonary hypertension and amyloidosis models in retrospective multi-site studies summarized in their clearances.

Most authors have Mayo Clinic or company ties, and Mayo holds a financial interest in Anumana. A cost-effectiveness analysis of EAGLE and a randomized trial in Nigeria that paired the model with a digital stethoscope are listed on the publications page.

One architectural statement, and nothing on retention or training use. The ECG-AI page says the algorithm is installed in the health system's own environment, on a virtual machine or inside a compatible ECG management system, where processing occurs entirely within that system; on that description the traces need not leave the customer. Nothing states whether any data returns to Anumana, whether customer traces feed the new data the change control plans call for, or how long anything is kept. The only privacy policy covers the website (effective April 2021).

The models were trained on Mayo Clinic records described as research-authorized, and Mayo co-founded the company; what patients were told about that use is not published. Safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled, goes undescribed as well.

No status is stated, and the only privacy document covers the website. The Anumana privacy policy (effective 6 April 2021) says it applies only to information collected through the site. No business associate position, agreement or statement that the product does not handle protected health information is published.

The ECG-AI page's description of processing inside the customer's own environment suggests the company may not receive patient data in normal operation, but it takes no position on that, and support access and model updates are not addressed.

An outside check of the software, and no attestation of the company. The pulmonary hypertension and amyloidosis clearance summaries record cybersecurity testing as part of FDA premarket review, and the company publishes a vulnerability disclosure policy that invites good faith security research. No SOC 2 report, ISO/IEC 27001 certificate, HITRUST certification or trust portal is published. Because the software runs inside the customer's environment, the customer's own controls carry much of the exposure.

Each model is cleared and identifiable, with its indication and a change control plan. K232699 (September 2023) cleared the low ejection fraction model for adults at risk of heart failure; K250652 (July 2025) extended it to more Philips carts with a predetermined change control plan. K252360 (28 March 2026) cleared the pulmonary hypertension model for adults presenting with shortness of breath, and K253801 (7 April 2026) the cardiac amyloidosis model for adults being evaluated for signs of it, both under 21 CFR 870.2380 with new product codes and their own change control plans. K260300 (15 June 2026) cleared WatchMate software for left atrial appendage closure planning.

The platform and ECG-AI pages still list the pulmonary hypertension and amyloidosis models as under development, while the press releases and indications pages describe them as cleared. The company describes each as the first of its kind on a standard 12-lead ECG; that is the company's claim.

Subgroup performance is published for each cleared model, including where it is weaker. The pulmonary hypertension summary (K252360) tabulates sensitivity and specificity by age, race and body mass index across 21,066 patients: sensitivity was 76.7 percent below a BMI of 30 and 64.2 percent at 35 or above. The amyloidosis summary (K253801) reports significant variation by race, with the lowest performance in Black patients (sensitivity 76 percent, specificity 88.0 percent) against White patients (80.6 and 91.6 percent), by age, by BMI (sensitivity 66.2 percent at 35 or above) and by ECG manufacturer. The low ejection fraction summary (K232699) found no significant difference by site, sex or race, and variation by age, BMI and prior heart failure or infarction.

Each model also has an FDA-reviewed predetermined change control plan that sets how it may be retrained and validated. The FAQ's statement that results were similar across racial groups applies to the low ejection fraction model, not to the amyloidosis model.

Error rates are published with their denominators, and nothing stands behind them. The multisite validation of the low ejection fraction model (JACC, 2026) reports 84.5 percent sensitivity and 83.6 percent specificity in 13,960 patients from four US sites, with a positive predictive value of 30.5 percent and a negative predictive value of 98.4 percent; the pulmonary hypertension clearance gives 73.0 percent sensitivity and 74.4 percent specificity in 21,066 records, with an estimated positive predictive value of 11.5 percent among US adults with shortness of breath; the amyloidosis clearance gives 78.9 and 91.2 percent. A buyer can hold the models to those figures.

The about page and FAQ quote 90.2 percent sensitivity and 85.1 percent specificity for the low ejection fraction model from data on file, which the published validation does not show. No warranty, indemnity or remediation commitment is published; the website terms disclaim warranties for the site itself.

Named inputs and a standards-based path into the record. The clearances name the ECG carts the software reads from: GE Dash 3000, GE MAC 5500, GE MAC VU360 and Philips PageWriter TC70, with K250652 adding the PageWriter TC50 and TC30. The ECG-AI page describes the integration: an ECG-AI orderable and results view built in the EHR, HL7 data points for exchange, and XML or DICOM waveform files processed inside the ECG management system, with the result returned to the EHR as a binary finding.

The company names no EHR vendor, ECG management system product or customer deployment.

Two deployment options are stated, both inside the customer's environment. The ECG-AI page describes installing the ECG-AI framework on a virtual machine configured by the health system, or running it within a compatible ECG management system, where the page says algorithmic processing occurs entirely within that system; waveforms are read as XML or DICOM files and results written back over HL7. On that description the ECG stays with the health system.

What is left open is the network access the virtual machine option requires, whether anything is sent back to Anumana for support or model updates, and which ECG management systems qualify. The software is sold in the United States only.

No price is published, and how the product is paid for is discoverable. The FAQ says the low ejection fraction model is eligible for Medicare reimbursement in outpatient settings, and the ECG-AI page lists setting up CPT Category III codes for charge capture as part of installation, alongside a virtual machine or ECG management system install and HL7 configuration. No license fee, per-study price or implementation charge is published, and access starts with a demonstration request. Reimbursement eligibility is payer context, not the price a health system pays; pharmaceutical collaborations are contracted separately.

Coverage is named by population for each model, with validation behind it, and not by setting. The low ejection fraction model is for adults at risk of heart failure, including postpartum women and people on cardiotoxic drugs; the pulmonary hypertension model for adults with shortness of breath; the amyloidosis model for adults being evaluated for it. Each clearance summary reports a multi-site US validation in that population, and the randomized trial ran in primary care.

The company does not state which care settings each model has been validated in beyond that, and the software is sold in the United States only.