A simple wearable-style electrocardiograph (ECG) may produce signs of hidden structural heart disease – not pulse or heart-rate data, but the heart’s electrical trace. This could help to identify people who need ultrasound scanning before symptoms appear, although real-world smartwatch recordings still need to be tested.
The heart can quietly change long before a person feels ill. A valve may start leaking. The main pumping chamber may thicken or weaken. For years, the person carrying these changes may feel nothing.
Waiting until palpitations or shortness of breath appear can mean missing the window when treatment is simplest. By the time some people receive echocardiography – ultrasound scanning that maps the heart in moving images – they may already need valve replacement or another invasive treatment.
“Early detection really matters,” says lead author Elianna Knight, a doctor and PhD student researching deep learning in health at Yale University in the United States. “But you do not necessarily know you need an echo before it is too late.”
That is where the wrist comes in. What if something millions of people already wear could help reveal who needs echocardiography?
Researchers have now shown that artificial intelligence (AI) can detect signs of structural disease from 1-lead ECG signals similar to those produced by many wearable and portable devices. The signal is electrical, and not an image, but the model appears to pick up traces of the heart’s shape and function that a human reader would usually not see. The results, published in the European Heart Journal – Digital Health, follow earlier AI-ECG studies showing that electrical traces can contain information usually confirmed by scans.
When rhythm is only part of the story
Smartwatches can already flag rhythm problems such as irregular heartbeats, and earlier work has shown that AI can detect weakened pumping from smartwatch ECGs in real-world settings. The new study takes the next step: from one pumping problem to a broader group of structural diseases that normally require imaging – thickened walls, faulty valves or weakened pumping – before people notice symptoms.
“You could potentially get people into the clinic faster and prevent some of these valve replacements” and cases of heart failure, says Knight.
ECGs are a frontline test for many heart problems: electrodes on the chest and limbs record the tiny electrical signals that coordinate each heartbeat.
“It captures the electrical signal that goes through the heart to tell it to pump, initiated by the sinoatrial node, your internal pacemaker,” Knight explains. ECGs are quick, widely available and crucial for detecting heart attacks and rhythm disorders.
But many dangerous heart problems are structural rather than rhythmic: whether the muscle has thickened, how well the chambers fill, how strongly they pump or whether the valves close properly. Diagnosing them usually requires an echocardiogram, an ultrasound scan that shows the heart moving in real time.
The model learned to connect electricity and structure
The question was whether that trace still carried structural clues a deep learning model could learn to recognise. To train the foundation model, Knight and her colleagues used 194,551 paired ECG and echocardiogram records from 77,378 adults treated in the Yale New Haven Health System. In each pair, the ECG showed the electrical signal, and the echocardiogram report described the heart’s shape and function. The model was then tested on later patient records.
The idea came from AI systems that learn to connect images with written descriptions. “Around that time, CLIP just came out,” Knight says, referring to OpenAI’s Contrastive Language–Image Pretraining. Here, the model matched ECG traces with the same patient’s echocardiogram report: “a text where a physician rigorously describes the structural heart diseases present in the echocardiogram”.
Although a standard 12-lead ECG records the heart from several angles, the researchers only gave the model Lead I, the lead that most closely resembles the signal captured when many wrist-worn or handheld devices record between the two arms. “Lead I is supposed to be the most similar to wearables” such as smartwatches and fitness trackers, Knight says. But Lead I from a clinical ECG is still only a proxy: the study did not test consumer smartwatch recordings directly.
The study is therefore a step from hospital ECG data towards wearable screening – not proof that a smartwatch can already diagnose structural heart disease.
Clues doctors cannot see in an ECG
Knight and colleagues trained the model to identify three clinically important problems: reduced left ventricular ejection fraction, meaning that the main pumping chamber pushes out too little blood with each beat; moderate or severe diastolic dysfunction, meaning that the heart has trouble relaxing and filling between beats; and a broader category covering reduced pumping, moderate or severe valve disease and severe thickening of the left ventricle.
On patient data it had not seen during training, the model reached AUROC scores of 0.894 for reduced pumping function, 0.849 for diastolic dysfunction and 0.887 for the combined structural heart disease category. AUROC measures how well a model ranks people with a condition above people without it: 1.0 would be perfect; 0.5 is no better than chance. A standard neural network trained on the same tasks showed whether ECG–echocardiogram training added anything beyond simply feeding many ECGs into AI.
The model was not simply copying what doctors already see in an ECG. Standard ECG interpretation is excellent for rhythm and electrical problems, but structural disease is usually confirmed by imaging. The new result moves earlier AI-ECG work from rhythm and pumping signals towards heart structure more broadly.
“Most physicians cannot detect structural heart diseases from ECGs,” Knight says. “Maybe the most expert cardiologists, but this is not really something that a human eye can understand.”
A warning, not a diagnosis
Knight emphasises that AI models reading wearable-style ECGs would not replace a 12-lead ECG and certainly not an echocardiogram. A warning from such a model would not be a diagnosis. The goal is to identify people who should be sent for proper testing sooner – while also measuring how many would be sent for scanning unnecessarily.
The caution matters because the study focused on what doctors call “actionable” structural disease – disease advanced enough to matter clinically, such as reduced pumping function or moderate to severe valve disease.
Nor was it tested on the general smartwatch-wearing public. The data came from people in a health system who had both ECGs and echocardiograms. Long-term studies in seemingly healthy people are still needed to show whether ECG signals change before the disease becomes entrenched and treatment becomes invasive.
Even within these limits, Knight thinks the model can still improve. “We just were not encoding a lot of information that was potentially really rich from physicians,” she says. “There are many new types of deep learning architectures that we could potentially employ to improve this.”
The advantage was clearest when labelled data were scarce. Wearable-Echo-FM only modestly outperformed a standard neural network. But when both models received only about 0.5% of the training data – 617 to 1251 ECGs depending on the task – the foundation model performed far better. ECG–echocardiogram training helped most when examples were limited, suggesting that the model had learned a broader link between electrical signals and heart structure rather than merely memorising one large dataset.
This matters because common conditions can be studied using large hospital datasets, whereas rarer structural heart diseases may leave researchers with only dozens or hundreds of labelled examples. A model that needs fewer examples could make rare-disease screening more realistic – as a starting-point for several related clinical questions and not one detector for one disease. Adapting the model to these conditions “is the next step,” Knight says.
