Beyond the Beat: How AI Is Turning the ECG Into a Screening Engine
By Constantine Tarabanis, MD, Cardiology Fellow, Mass General Brigham
A century-old diagnostic tool is being repurposed to detect what neither clinician nor calipers can see, and the regulatory and reimbursement infrastructure is finally catching up.
The 12-lead electrocardiogram (ECG) is among the most familiar and reproducible diagnostic instruments in medicine. For nearly a century, its value has been almost entirely diagnostic, with waveform features mapped to defined pathologies such as ischemia and arrhythmia. That role is now expanding with the rise of artificial intelligence.
Applying neural networks to ECG interpretation is not, in itself, new, as publications dating back to the 1990s describe early automated rhythm classification. What is new, and what has driven the field’s acceleration over the past seven years, is the use of deep learning architectures, not to replicate established diagnostic labels but to infer underlying conditions that conventional ECG interpretation cannot capture. The 2019 demonstration that a neural network could detect asymptomatic left ventricular systolic dysfunction from a normal-appearing sinus rhythm ECG established that the ECG signal carries clinically relevant information beyond what is accessible to human or rule-based interpretation.
Why the ECG Is the Ideal Substrate
Two features make the ECG uniquely positioned for AI-driven screening at scale. It is ubiquitously generated, producing training datasets larger than those of other modalities. And it is inexpensive and rapidly acquired, requiring neither specialized infrastructure nor advanced operator training. Any algorithm extracting a clinically actionable signal from the ECG inherits those distribution advantages.
The field has produced a growing body of validated AI-ECG models for conditions the ECG was not built to diagnose, including paroxysmal atrial fibrillation, arrhythmic risk, hypertrophic cardiomyopathy, aortic stenosis, and cardiac amyloidosis, among others.
The open questions for AI-ECG are no longer primarily about architecture or signal processing. They concern validation, deployment, reimbursement, and clinical value.
The Regulatory Reality and Its Limits
Any AI-ECG tool intended for U.S. clinical use is, by definition, Software as a Medical Device (SaMD), classified along standard risk tiers. Most products reach the market with Class II 510(k) clearances. The FDA’s AI/ML SaMD Action Plan and Predetermined Change Control Plan guidance attempt to build a framework for algorithms that learn over time without requiring new clearance for every retraining. As of late 2025, more than 1,400 AI-enabled devices have been FDA-authorized, with cardiology the second-largest application domain after radiology.
Clearance, however, does not have clinical value. Many cleared devices have established retrospective predictive performance without ever showing that they change physician behavior or improve outcomes. Compounding this limitation is the field’s persistent reliance on internal, single-institution validation. Neural networks overfit to dataset-specific artifacts, and a model with high reported accuracy on its home dataset routinely loses several points of discrimination when deployed elsewhere. Too few published models are externally validated at all.
The prospective evidence base is thinner still, but expanding. The 2021 EAGLE randomized trial showed that AI-guided screening for low ejection fraction increased new cardiomyopathy diagnoses in primary care without inflating overall echocardiogram utilization. More recently, a 2024 pragmatic randomized trial from Taiwan showed an AI-ECG mortality-alert system reduced 90-day all-cause mortality in hospitalized patients, a rare example of AI-ECG moving the needle on a hard clinical endpoint. International prospective work has continued to expand this evidence base, including a recent JAMA Cardiology study evaluating the AiTiALVSD model in a Kenyan health system, demonstrating that AI-ECG performance can be sustained in low- and middle-income settings where echocardiographic capacity is limited.
Reimbursement Catches Up, Adoption Lags
Reimbursement is finally moving. The AMA created Category III CPT codes 0764T and 0765T in 2023 for AI-based ECG analysis, and effective January 1, 2025, CMS assigned them a national Medicare payment rate of $128.90. Condition-specific codes have followed, signaling that payers are beginning to treat AI-ECG as a reimbursable clinical service rather than a research curiosity.
Yet the gap between reimbursement and routine adoption remains wide. Large health systems remain cautious about integrating tools whose incremental clinical benefit and financial return have not been clearly established at the institutional level, and clinicians remain wary of opaque algorithmic outputs. Sustainable uptake will likely require business cases in which institutional clinical and financial interests align with deployment, whether through demonstrated effects on patient outcomes, downstream procedural volume, value-based contracts, or quality metrics tied to early detection.
An alternative direction, less susceptible to these institutional friction forces, lies outside conventional clinical workflows. Pharmaceutical applications are particularly compelling. Building on published work demonstrating that AI-ECG models can recover the electrocardiographic signature of specific drugs, including dofetilide and sotalol, such models could verify medication adherence in clinical trials without serum drug levels, enrich populations for trials of cardioactive therapies, or serve as objective adherence endpoints in regulatory submissions. The economic case in industry settings is often clearer than in routine care, and deployment risk is contained within a study population.
The Wearable Frontier and Its Limits
The migration of AI-ECG from the 12-lead machine to the wrist may represent the field’s most consequential near-term opportunity, with the potential to enable genuinely population-scale screening. The Mayo Apple Watch study demonstrated that a single-lead AI-ECG could detect low ejection fraction, a result that would have been considered implausible only a few years earlier. Comparable work in the consumer space is being advanced by Medical AI, Inc., whose AiTiALVSD model has been evaluated in a large general-population health-screening cohort and integrated into the Samsung Galaxy Watch for left ventricular systolic dysfunction detection, an early example of structural disease screening migrating from hospital infrastructure to consumer hardware.
The principal limitation, however, is that predictive performance is generally attenuated when models trained on 12-lead inputs are deployed on single-lead acquisitions. The Apple Watch, for example, remains FDA-cleared for atrial fibrillation detection from its own single-lead recordings, not for the prediction of atrial fibrillation from sinus rhythm ECGs or for the structural and systolic function indications that AI-ECG has demonstrated on the 12-lead. Differences in electrode contact, sampling rate, filtering, and baseline noise may erase subtle waveform features essential to certain prediction tasks, and because deep learning models offer limited interpretability, it remains unclear which signal characteristics are being lost.
Where the Field Goes Next
The open questions for AI-ECG are no longer primarily about architecture or signal processing. They concern validation, deployment, reimbursement, and clinical value.[L1] The next decade will be defined less by who builds the most accurate model on a held-out test set, and more by who runs the prospective trials, secures the reimbursement pathways, designs the workflows, and demonstrates to clinicians, patients, payers, and regulators that AI-guided ECG interpretation produces healthier patients than the standard of care it seeks to augment. A century after Einthoven, the ECG is becoming a window into latent biology. The work ahead is to make sure that the window is reliable, equitable, and clinically meaningful.
About the Author
Constantine Tarabanis, MD is a cardiology fellow at Mass General Brigham and a researcher at the Broad Institute of MIT and Harvard, focused on the development, validation, and clinical deployment of AI-ECG models. He serves as a scientific advisor to Medical AI, Inc..