The Hidden Signal in Every ECG You Already Run

Key Takeaway: An artificial intelligence model trained on standard ECG images can reliably detect patients with transthyretin amyloid cardiomyopathy—a treatable but chronically underdiagnosed cause of heart failure. Validated across five international cohorts, this tool could transform the widely used, low-cost test into a powerful first-line screening system capable of catching a disease that has been overlooked for years.

A 12-Lead Clue Hiding in Plain Sight

Picture a 74-year-old man sitting on an examination table. He’s had shortness of breath for two years. An echocardiogram showed thickened heart walls, leading to a diagnosis of “heart failure with preserved ejection fraction”—a label that, while technically correct, tells only half the story. Somewhere in the intricate tracings of his electrocardiogram lies a pattern too subtle for the human eye to reliably detect: the electrical signature of a protein called transthyretin silently infiltrating his heart muscle. If someone, or something, could read that signal, his entire trajectory might change. That “something” now exists.

A Disease in the Shadows

Transthyretin amyloid cardiomyopathy (ATTR-CM) arises when the transthyretin protein, produced in the liver, misfolds and deposits as amyloid fibrils in the heart walls. The result is a progressively stiff, thickened ventricle that struggles to fill and ultimately fails^[2]. It comes in two types: a hereditary variant caused by genetic mutations in the TTR gene, and the wild-type form, once thought to be rare, which predominantly affects men over 65. We now know that wild-type ATTR-CM is far more common than previously believed. Autopsy studies have found transthyretin amyloid deposits in roughly 25% of hearts from individuals over age 80^[3]. The disease hides behind more familiar diagnoses like hypertensive heart disease, aortic stenosis, and hypertrophic cardiomyopathy, and the average time from symptom onset to diagnosis has historically exceeded three years.

This diagnostic delay now carries immense weight because, for the first time, we have a disease-modifying therapy. Tafamidis, a transthyretin stabilizer, was shown to reduce mortality and cardiovascular-related hospitalizations in patients with ATTR-CM in the ATTR-ACT trial^[4]. A treatable disease that is routinely missed: this is precisely the kind of problem that scalable screening technology is designed to solve.

The AI Model: Development and Performance

An international team developed an AI model trained not on raw ECG waveform data, but on the standard 12-lead ECG images of the kind printed on paper or stored as PDFs in hospital systems worldwide. This design choice was deliberate: Image-based models are hardware-agnostic, meaning they can work with ECGs from nearly any machine, from any clinic, anywhere.

In internal validation, the model achieved an area under the receiver operating characteristic curve (AUROC) of 0.84 (95% confidence interval, 0.79–0.89)—a strong performance metric indicating it accurately distinguished patients with ATTR-CM from those without^[1]. Critically, this performance held up even in clinical subgroups that often complicate diagnosis, such as patients with left ventricular hypertrophy and aortic stenosis—the very populations where ATTR-CM most often hides.

The model was then stress-tested in five different international external cohorts, delivering AUROCs ranging from 0.78 to 0.89. In dedicated screening cohorts of high-risk individuals—the target population where such a tool would be used in practice—the diagnostic utility remained strong, with AUROCs between 0.76 and 0.91.

Perhaps the most clinically elegant finding involved a sequential screening strategy. When the AI-powered ECG was used as an initial filter, followed by an AI-guided echocardiogram analysis, the positive predictive value—the probability that a flagged patient actually has the disease—jumped from 0.24 to 0.66. In other words, two-thirds of patients flagged by the combined AI pipeline would indeed have ATTR-CM, a yield high enough to justify the cost and logistics of confirmatory nuclear scintigraphy or endomyocardial biopsy.

Why the ECG Captures This Signal

At first glance, the idea that a surface ECG can detect a protein deposition disease might seem almost implausible. But the biology makes sense. Amyloid infiltration disrupts the normal electrical architecture of the myocardium. The fibrils accumulate between cardiomyocytes, increasing the distance electrical impulses must travel and altering conduction velocity^[5]. Classic ECG findings in cardiac amyloidosis include low QRS voltage relative to wall thickness, a pseudoinfarct pattern with pathologic Q waves in the absence of coronary artery disease, and conduction abnormalities like first-degree atrioventricular block. None of these features alone is sensitive or specific enough. An AI model, however, can simultaneously integrate dozens of subtle waveform features—amplitude ratios, morphological nuances in the ST segment and T wave, timing relationships across all twelve leads—and synthesize a probability score in a way no human could replicate at scale.

Important Caveats

No single study, however well-designed, rewrites clinical practice overnight. The model’s positive predictive value as a standalone tool was a modest 0.24, meaning that without a second confirmatory step, roughly three out of four flagged patients would not have ATTR-CM. This underscores the importance of the sequential strategy. Furthermore, the prevalence of ATTR-CM in the training and validation cohorts directly impacts predictive values; real-world application in lower-prevalence settings could change these figures. Prospective implementation studies that measure whether AI-assisted ECG screening actually shortens the time to diagnosis and improves patient outcomes are the necessary next step.

What This Means for Tomorrow’s Practice

The vision for practice is both compelling and attainable. Every hospital, every emergency department, every primary care office already performs ECGs. The infrastructure is in place. Embedding an AI algorithm into existing ECG interpretation software requires no new hardware, no extra patient effort, and minimal workflow disruption. When a patient gets an ECG for any reason—palpitations, a pre-operative evaluation, a chest pain workup—the system quietly runs an analysis in the background. If the probability of ATTR-CM exceeds a threshold, an alert appears in the electronic health record, prompting the clinician to consider further workup.

For that elderly man on the exam table, this changes everything. Instead of a years-long diagnostic odyssey in a heart failure clinic, a silent algorithm reads his ECG, raises its hand, and essentially says: look closer. And because Tafamidis is available, looking closer now means having the opportunity to slow a disease that, left unchecked, would steal years from his life. This is the quiet revolution embedded within twelve leads and a few million machine-learning parameters.

Scientific Sources

1. Croon PM, et al. Electrocardiogram-Based Deep Learning to Prioritize Testing for Transthyretin Amyloid Cardiomyopathy. JAMA. 2026. PubMed: https://pubmed.ncbi.nlm.nih.gov/42663420/
2. Ruberg FL, et al. Transthyretin Amyloid Cardiomyopathy: JACC State-of-the-Art Review. J Am Coll Cardiol. 2019. DOI: 10.1016/j.jacc.2019.04.003
3. Tanskanen M, et al. Senile systemic amyloidosis affects 25% of the very aged and associates with genetic variation in alpha2-macroglobulin and tau: a population-based autopsy study. Ann Med. 2008. DOI: 10.1080/07853890701842988
4. Maurer MS, et al. Tafamidis Treatment for Patients with Transthyretin Amyloid Cardiomyopathy. N Engl J Med. 2018. DOI: 10.1056/NEJMoa1805689
5. Falk RH. Diagnosis and management of the cardiac amyloidoses. Circulation. 2005. DOI: 10.1161/CIRCULATIONAHA.104.489187