The ECG AI Arms Race

The 12-lead electrocardiogram is one of medicine’s workhorses — a ten-second electrical recording of heart activity that has been part of clinical practice since Willem Einthoven developed the galvanometer version in 1903. In the more than 120 years since, the ECG has been interpreted essentially the same way: a cardiologist or computerized algorithm looks for specific waveform patterns (P waves, QRS complexes, T waves, intervals, and their deformities) that correspond to known arrhythmias, ischemia, conduction disease, and electrolyte abnormalities.

The ECG tells you about the heart’s electrical activity at the moment of recording. That is, or was, its limit.

In 2019, a team at the Mayo Clinic published a paper in Nature Medicine demonstrating that a deep learning model could detect low ejection fraction — reduced pumping function of the left ventricle, which is a hallmark of heart failure — from a standard 12-lead ECG with an AUC of 0.93. This matters because ejection fraction is normally measured by echocardiography, a more expensive and technically demanding test. Low ejection fraction affects roughly 2 percent of the population, is largely asymptomatic until advanced, and is eminently treatable with a well-established class of medications. A screening test for it on a widely available, cheap tool would be clinically significant.

That paper opened a door. What followed was a cascade.

What AI ECG Models Can Now Do

The list of conditions that AI models can detect or predict from a 12-lead ECG, as validated in published studies through 2025, is remarkable and still growing:

Related reading

Atrial fibrillation — including paroxysmal AF that is not present at the time of the ECG (the model detects structural signatures that correlate with AF burden). Ejection fraction below 40 percent. Hypertrophic cardiomyopathy. Pulmonary hypertension. Hyperkalemia (high blood potassium). Anemia (the cardiac signatures of reduced oxygen-carrying capacity). Estimated biological age (which differs from chronological age and correlates with cardiovascular risk). Estimated sex (with >95 percent accuracy, which matters because sex differences in ECG morphology affect interpretation). Short QT syndrome, a rare but potentially lethal arrhythmia. And, most provocatively, all-cause mortality risk — not just cardiovascular mortality but overall survival, from a 10-second recording.

The mortality finding, from a large study by the Mayo group published in Nature Medicine in 2023, generated appropriate controversy. The model was predicting something real — its predictions were validated in independent cohorts and calibrated well — but what exactly it was predicting is not fully understood. The ECG captures cardiac function, and cardiac function is a global indicator of biological health. The model may be detecting subtle cardiac phenotypes that reflect systemic disease burden, biological aging, or both. It is predicting mortality in the same way that grip strength predicts mortality: not because cardiac electrical activity causes all causes of death, but because it reflects something broader about biological reserve.

The Screening Opportunity

The practical implication of these AI capabilities is substantial. An ECG costs roughly $15-30 to perform in a clinical setting and is already done routinely for surgical clearance, preemployment health assessments, and standard cardiac risk evaluation. If the same recording can screen for a dozen additional conditions — including conditions that are treatable if caught early — the cost-benefit arithmetic for expanded AI-ECG screening is compelling.

The ECG-AI tools with FDA clearance for specific indications include AliveCor’s KardiaBand (AF detection from single-lead ECG), Apple Watch’s ECG feature (also AF detection, CE marked and FDA authorized), and Mayo Clinic’s AI-ECG platform for low ejection fraction detection, which has been licensed for clinical deployment and is in use at several large health systems. Several commercial platforms (Eko, Cardiologs, Viz.ai) have cleared AI ECG analysis tools for various arrhythmia detection applications.

The most ambitious deployment vision is ECG-as-universal-screening: running AI analysis on every 12-lead ECG performed (at any indication) to screen for the full range of AI-detectable conditions simultaneously. Population-level implementation of this approach, without careful management of downstream resource implications, would generate enormous numbers of positive screens — the majority of which, in low-prevalence populations, would be false positives requiring further workup at significant cost and patient anxiety.

The Cascade Harm Problem

This is where ECG-AI transitions from impressive technology to complicated clinical management. A positive screen for low ejection fraction triggers an echocardiogram. In a high-volume deployment, even a specificity of 95 percent means 5 percent false positives — in a hospital doing 10,000 ECGs per year, that’s 500 unnecessary echocardiograms per year. Scaled nationally, the resource implications are substantial.

The paroxysmal AF application is particularly fraught. The AI model predicts AF risk (the probability that a patient has or will develop AF), not the presence of AF. A patient flagged as high-risk for AF may need extended Holter monitoring to determine whether AF is actually present, and the decision of whether to start anticoagulation — which carries its own bleed risk — should not be made on the basis of an ECG-AI score alone.

Guidelines from the European Society of Cardiology and American Heart Association on ECG-AI applications, issued in 2025, emphasized that AI-ECG outputs should be used to risk-stratify and prioritize further testing, not as standalone diagnostic criteria. The distinction between screening tool and diagnostic test is critical, and several reported adverse events in ECG-AI deployments involved clinical teams treating high-confidence AI scores as diagnoses rather than risk indicators.

Related reading

The Single-Lead Smartphone Revolution

Perhaps more clinically impactful than the 12-lead ECG AI for population health is the democratization of single-lead ECG through smartphones and wearables. The Apple Watch’s ECG feature (FDA authorized since 2018) has been used by millions of patients to record atrial fibrillation episodes. AliveCor’s Kardia devices have enabled primary care physicians in offices without ECG machines to record cardiac rhythm.

Studies of this consumer cardiac monitoring have found that it identifies atrial fibrillation in patients who were undiagnosed — the KardiaBand Heart Health Study found incident AF detection in roughly 1 in 100 high-risk patients screened using consumer devices. That may sound modest, but AF affects roughly 6 million Americans and is a leading cause of stroke, and earlier detection enables earlier anticoagulation that reduces stroke risk. The downstream clinical benefit, if it reaches the patients most at risk, is meaningful.

The limitation is familiar from all AI diagnostic deployments: the patients using Apple Watches and KardiaBand devices are disproportionately affluent, tech-literate, and already engaged with their health. The patients at highest risk from undetected AF — older, less affluent, less connected to healthcare — are less likely to own or use these devices. The technology exists to screen for AF in community pharmacies, barbershops, and public spaces using cheap consumer-grade devices and AI analysis. The healthcare system infrastructure to follow up on those screens, counsel the patients, and connect them to care is where the bottleneck actually lives.

What 2026 Shows

The ECG has become, through AI, something it never was: a multi-disease screening tool with unprecedented information density. That transformation is real, validated, and clinically significant for specific indications.

What the transformation doesn’t yet have is a corresponding transformation in how clinical systems use the information. The data now available from a routine 10-second recording exceeds what any individual clinician can act on without structured decision support, clear clinical pathways, and resource allocation frameworks that account for the downstream burden of positive screens. Building those systems is harder than building the models. It is also, ultimately, where the patient benefit actually lives.

The International Race

ECG-AI has become a quiet geopolitical competition that mirrors the broader AI landscape. China’s healthcare AI ecosystem — led by companies like Ping An Health, Infervision, and a cluster of hospital-university partnerships — has generated large ECG databases from a patient population of 1.4 billion and trained models on them at scale. Several Chinese ECG-AI systems have been validated in published studies with performance metrics competitive with or exceeding Mayo Clinic’s platform for AF detection and certain arrhythmia classes.

The implication is that ECG-AI, like AI diagnostics broadly, is not a technology where U.S. and European academic medical centers have a durable monopoly on innovation. The countries with the largest healthcare datasets and the fewest regulatory barriers to deploying AI in clinical settings are rapidly accumulating the training data and deployment experience that builds competitive advantage. Whether the models trained on Chinese patient populations generalize well to non-Chinese populations is an empirical question with practical implications for global deployment.

European ECG-AI development has followed a different model: consortium-based, with academic medical centers pooling data through frameworks like the European Reference Networks and national health data initiatives. The ECLAIR consortium (European Cardiology AI Research), formed in 2023 with 15 academic center participants, has published comparative validation studies that are among the most rigorous in the field. The consortium model produces more generalizable models than single-institution validation, but progresses more slowly than either U.S. commercial players or Chinese health system scale.

Related reading

The Hardware Convergence

The future of ECG-AI is probably not hospital-based 12-lead machines. It’s wearable continuous monitoring integrated with ambient computing. The Apple Watch ECG, introduced in 2018, established proof of concept for consumer-grade single-lead ECG. Since then, multiple devices — the Samsung Galaxy Watch series, Withings ScanWatch, Garmin Venu 3 — have added ECG capabilities. The next generation of these devices will have six-lead or twelve-lead equivalents, sufficient for the full range of AI diagnostic applications currently requiring clinical equipment.

When ECG-AI capability moves from the hospital into the wrist, the population exposed to cardiac screening changes from “people with symptoms who see a physician” to “people who wear a smartwatch.” That’s a younger, wealthier, more health-conscious population — a different selection from the population at highest cardiovascular risk. The technology will be most widely distributed where it’s least urgently needed.

Unless the clinical systems built around these devices — remote monitoring platforms, AI alert routing, primary care integration — are specifically designed to reach higher-risk populations who don’t self-select into wearable health technology. That design challenge is organizational and incentive-based, not algorithmic. The ECG algorithm is, in some ways, the easy part.

Einthoven invented the modern ECG in 1903 and won the Nobel Prize in 1924. It took decades for the technology to move from academic demonstration to routine clinical practice, because the barriers were organizational and economic rather than technical. The first hospitals with electrocardiograph machines were not the hospitals that most needed them. The pattern has a certain durability. AI ECG tools are better than Einthoven’s machine by any measurable standard. The distribution problem is the same one that’s always existed in medical technology: the people who build it and the people who need it are rarely the same people, and the gap between building and deploying is filled by exactly the structural factors — infrastructure, economics, clinical workflow, reimbursement — that algorithms don’t address. Solving the algorithm is the beginning. The rest is medicine.