AI ECG Screening: Detecting Heart Failure Early From a Routine ECG

AI ECG screening is one of the clearest examples of clinical AI doing something humans cannot. A deep learning model reading a standard 12-lead ECG can flag a weak left ventricle, low ejection fraction, that a cardiologist cannot reliably see by eye. In 2025 the FDA cleared two more AI-ECG tools for this purpose, and a large UK trial published in 2026 tested AI stethoscopes in routine primary care.

This matters because left ventricular systolic dysfunction is often silent until a patient presents with breathlessness or swelling. Echocardiography confirms it, but you cannot echo everyone. An ECG is cheap, fast and already done in every OPD, emergency room and health check.

This guide explains how AI-ECG works, what the evidence shows, which tools are cleared, and what an Indian hospital needs to put it into practice.

How AI ECG works

A conventional ECG reading looks for patterns cardiologists have defined: ST elevation, bundle branch block, atrial fibrillation. AI-ECG models are trained differently. Researchers pair hundreds of thousands of ECGs with an echocardiogram done around the same time, then train a neural network to predict the echo result from the ECG alone.

The model learns subtle combinations of waveform features that correlate with a weak heart, features no human rule describes. The output is usually a probability or a positive/negative flag: "ECG suggests possible low ejection fraction; consider echocardiography".

Key points:

• It is a screening and triage tool, not a diagnosis. Echo confirms.
• It needs the raw digital waveform, not a photo of a paper printout.
• Performance depends on the population it was trained and validated on.

The evidence: from EAGLE to TRICORDER

The landmark trial is EAGLE, published in Nature Medicine in 2021. It cluster-randomised 120 primary care teams from 45 clinics and hospitals in the US to have access to AI-ECG results or usual care. Across 22,641 adults without known heart failure:

• New diagnoses of low ejection fraction rose from 1.6% in the control arm to 2.1% with AI (odds ratio 1.32).
• That is about five additional diagnoses per 1,000 patients screened.
• When the AI was positive, clinicians with access ordered echocardiograms more often (49.6% vs 38.1%).

The lesson is twofold. The AI found real disease, and its impact depended on clinicians acting on it. Half of AI-positive patients still did not get an echo.

The AI stethoscope is a related approach. Eko's stethoscope records a single-lead ECG and heart sounds during a routine exam, and received FDA clearance for low ejection fraction detection in April 2024 using an algorithm developed with Mayo Clinic. In January 2026, the TRICORDER trial02156-7/fulltext) from Imperial College London was published in The Lancet. It was a cluster-randomised trial across 205 NHS general practices, and its primary result was negative: simply offering practices the AI stethoscope did not significantly increase heart failure detection at 12 months, largely because clinicians used it at their discretion and not very often. Among patients who were actually examined with it, detection was higher; Eko's summary reports 2.3 times more heart failure, 3.5 times more atrial fibrillation and 1.9 times more valvular disease. Coverage of the trial also noted that about two-thirds of patients flagged for suspected heart failure did not have it on confirmatory testing. The lesson is the same as for any AI screening tool: the algorithm can work and the programme still fail if it is not built into routine workflow.

Cleared AI-ECG tools

Tempus's clearance is typical of the labelling: for ECGs recorded in a healthcare facility, not a stand-alone diagnostic, not for patient monitoring, and not for paced rhythms. Anumana's algorithm, originally developed at Mayo Clinic, was cleared under a product code for reduced ejection fraction machine learning-based notification software. For its original October 2023 clearance, Anumana reported a sensitivity of 84.5% and specificity of 83.6% in a retrospective validation of about 16,000 patients, for detecting an ejection fraction of 40% or less. The July 2025 clearance of its 12-lead version also included a Predetermined Change Control Plan, which lets the company update the model within pre-agreed limits.

Cardiology is one of the busiest areas of AI clearances: one regulatory analysis counted roughly 90 AI/ML cardiology devices cleared in 2025 alone, spanning ECG, echo and imaging tools.

Beyond low ejection fraction

AI-ECG research has gone well beyond heart failure. Published models target atrial fibrillation risk from a sinus rhythm ECG, hypertrophic cardiomyopathy, cardiac amyloidosis, valvular disease, and even estimates of age and sex from the ECG. Most of these are research tools or have limited clearances. Treat them as promising rather than ready.

The most practical near-term targets for a hospital remain:

• Low ejection fraction screening in at-risk adults.
• Faster triage of ECGs that need urgent cardiology review.
• Rhythm interpretation support for non-cardiologists.

The Indian context

India has a large and young burden of cardiovascular disease, uneven access to cardiologists outside cities, and enormous ECG volumes in OPDs, emergency departments, health checks and pre-operative assessments. That combination suits AI-assisted ECG reading.

AI-assisted ECG interpretation is already in use in India. Bengaluru-based Tricog, for example, offers ECG interpretation combining AI with remote cardiologists, and says it serves around 2,000 diagnosis points across Indian clinics, hospitals and diagnostic centres. Most such services focus on rhythm and acute findings (such as STEMI) for quick triage, rather than low ejection fraction prediction.

Two practical obstacles stand out:

• Paper ECGs. Many machines print and the paper goes in the file. AI needs digital waveforms, so the machine must export data (often a vendor XML, PDF with embedded data, or DICOM waveform format).
• Validation in Indian patients. The cleared low-EF models were mostly developed on US populations. Indian patients differ in age profile, body habitus and disease mix. Ask vendors for Indian validation data, and CDSCO's medical device software guidance now asks AI developers to disclose performance across Indian sub-populations.

What it takes to deploy AI-ECG

The pathway is the part most often skipped. EAGLE showed that even in a trial, many AI-positive patients did not get an echo. In an Indian hospital, the simplest fix is to generate an echo booking suggestion automatically in the HIS when the AI is positive, with the doctor confirming. Our guide to OPD management systems covers how orders and follow-ups flow in a busy OPD.

Who should be screened

AI-ECG works best where the chance of hidden disease is meaningful and an ECG is already being done. Sensible starting groups:

• Adults over 40 with diabetes, hypertension or chronic kidney disease attending OPD.
• Patients with unexplained breathlessness, fatigue or ankle swelling before echo is available.
• Pre-operative assessments in older adults.
• Patients with prior heart attack or known coronary disease who have not had a recent echo.

Screening low-risk young adults adds false positives and echo load for little gain. Tempus's clearance, for example, is limited to adults 40 and over at risk of heart failure.

Integration architecture

A practical setup has four parts:

1. Acquisition. ECG carts export to a network folder or send data directly to an ECG management system.
2. Normalisation. A service converts vendor formats into a standard representation and links each ECG to the patient's UHID and encounter.
3. AI scoring. The ECG is sent to the AI service (cloud or on-premise) and the result is stored with the ECG.
4. Results and workflow. The result appears in the EMR or HIS alongside the ECG, triggers suggested orders, and is logged for audit.

Standards matter here. HL7 v2 and FHIR can carry results and orders, and DICOM has a waveform format some machines support. Our HL7 FHIR integration guide explains these. At Algon this kind of device-to-record plumbing is routine healthcare integration work; the cleared AI model comes from a specialist vendor.

Costs

Indicative cost drivers, which vary widely by vendor and scale:

• Per-ECG or subscription fees for the AI service.
• Hardware upgrades where older carts cannot export digital data.
• Integration effort to connect carts, the AI service and the HIS.
• Echo capacity. More positive screens mean more echocardiograms. Plan cardiology capacity before you switch it on.
• Clinical time to review and act on results.

The last two are frequently forgotten. A screening programme that finds patients but cannot echo them within weeks creates risk and frustration.

Questions to ask an AI-ECG vendor

1. What exactly is cleared or approved, in which countries, and for which ECG machines?
2. What population was it validated on, and is there any Indian data?
3. What are sensitivity, specificity and positive predictive value at the threshold you recommend?
4. What input format does it require, and can our machines provide it?
5. How are results delivered: API, HL7, FHIR, PDF?
6. What happens to ECG data after scoring, and where is it stored?
7. How are model updates communicated?

Mistakes to avoid

• Screening everyone, including low-risk young adults, which raises false positives.
• No echo pathway for positives.
• Using photos of paper ECGs when the tool needs digital waveforms.
• Treating a negative AI result as ruling out heart failure in a symptomatic patient.
• Not measuring how many positives were confirmed.

Conclusion

AI ECG screening for low ejection fraction is backed by a randomised trial, has several FDA-cleared tools, and fits naturally into the ECG volumes Indian hospitals already handle. The hard parts are digital ECG capture, local validation and a dependable pathway from positive result to echocardiogram.

Start with one at-risk group, such as diabetic or hypertensive OPD patients over 40, connect your ECG machines, and measure confirmed diagnoses. For a related screening programme with strong Indian evidence, see our guide to AI diabetic retinopathy screening in India, and for broader data work see healthcare data analytics for hospitals.