Wearable Single-Lead ECG Detects Fine-Grained Structural Heart Disease Through Echo-Report Supervision

The human heart is a complex pump, and when its structure begins to change—when a valve leaks, a chamber swells, or the muscle weakens—the consequences can be severe. These conditions, known as structural heart diseases, often develop silently over years, only revealing themselves when the damage is already advanced. The gold standard for spotting these issues is an ultrasound of the heart, called an echocardiogram. It provides a clear, detailed picture of the heart's anatomy, but it requires expensive machines and highly trained technicians. This creates a bottleneck: millions of people live with undetected heart problems because they cannot easily access this specialized test. In contrast, the electrocardiogram, or ECG, is a simple, cheap, and ubiquitous tool that records the heart's electrical activity. While wearable devices have made single-lead ECGs available to the public, they have historically been limited to detecting rhythm problems, not the deeper structural issues that require an ultrasound to see.

A team of researchers has now bridged this gap, demonstrating that a simple, single-lead ECG can be used to detect a wide range of structural heart diseases with high accuracy. They developed a new artificial intelligence system that learns to read the electrical signals of the heart by studying the connection between those signals and the written reports of ultrasound exams. Instead of relying on thousands of manually labeled examples for every specific disease, the system was trained to align the raw ECG waveforms with the natural language descriptions found in thousands of physician-written ultrasound reports. By doing so, the AI learned to recognize the subtle electrical footprints left by structural changes, effectively turning a basic electrical recording into a powerful screening tool.

The researchers tested their system, which they named AnyECG-Echo, on a massive dataset involving over 25,000 patients from two different hospitals in China. The goal was to see if the model could identify thirteen distinct types of structural heart problems, ranging from a weakened pumping muscle to enlarged heart chambers and leaking valves. The results were striking. The system successfully identified these conditions with a high degree of reliability, even when tested on a completely new group of patients from a different city, proving that the knowledge it learned was not just memorized from one specific hospital's data. For critical conditions like reduced pumping function and global heart enlargement, the model performed with an accuracy comparable to the best existing tools, despite using only a single electrical lead instead of the twelve leads typically used in hospital settings.

What makes this discovery particularly significant is the breadth of what the system can detect. Previous artificial intelligence models for heart disease often focused on a narrow set of conditions or required complex, multi-lead inputs that are not available on consumer wearables. This new framework, however, covers a diverse spectrum of issues, including specific valve diseases and the enlargement of individual heart chambers. The researchers found that the system was particularly adept at spotting a weakened left ventricle, a condition where the heart muscle fails to pump blood effectively, as well as the enlargement of the heart's upper and lower chambers. While the system showed slightly lower accuracy for a few specific conditions, such as certain valve leaks or the widening of the main artery, it maintained a strong ability to flag the vast majority of structural issues that matter most for patient health.

To ensure the system was not just guessing or finding statistical tricks, the team looked closely at how the AI made its decisions. They examined which parts of the electrical signal the model focused on when it identified a disease. They found that the AI consistently paid attention to the specific segments of the heartbeat that medical experts know are relevant to those conditions. For instance, when detecting a leaky valve, the model highlighted the parts of the signal associated with the heart's relaxation phase, just as a cardiologist would expect. Furthermore, the researchers discovered that the probability scores generated by the AI acted like a digital biomarker. When the system predicted a higher risk of a specific problem, it corresponded directly with worse measurements from actual ultrasound exams, such as a lower pumping percentage or thicker heart walls. This suggests the system is not merely categorizing patients into "sick" or "healthy" but is actually tracking the severity of the disease in a continuous, meaningful way.

The study also revealed that this approach is remarkably efficient. The system achieved its high performance using a dataset that was significantly smaller than those required by other leading models, which often rely on millions of records. This efficiency suggests that the method of learning from written reports rather than manual labels is a powerful way to teach machines about heart health. The researchers noted that the system's performance remained consistent across different age groups and between men and women, though it did show a slight decline in accuracy for the oldest patients, likely due to the accumulation of other age-related electrical changes in the heart.

Ultimately, this work points toward a future where the first line of defense against heart disease is as simple as a few seconds of recording on a wearable device. By validating the system across different hospitals and time periods, the researchers have shown that it is robust enough to be used outside the controlled environment of a clinical trial. The findings suggest that in the near future, a person could use a smartwatch or a handheld device to screen for serious structural heart problems, receiving an immediate alert if a follow-up ultrasound is needed. This could transform the landscape of heart care, allowing for early detection in communities and homes where ultrasound machines are unavailable, potentially saving lives by catching silent heart disease before it causes irreversible damage.