Researchers say this is the first prospective study to demonstrate that an AI algorithm can detect multiple structural heart conditions using data from the single-lead ECG sensor embedded in the back and digital crown of a smartwatch.
“In our study, we explored whether the same smartwatches people wear every day could also help find these hidden structural heart diseases earlier, before they progress to serious complications or cardiac events,” said Arya Aminorroaya, who is also a research affiliate at the Cardiovascular Data Science (CarDS) Lab at Yale School of Medicine in New Haven, Connecticut.
The researchers trained the AI algorithm using over 266,000 12-lead ECG recordings from more than 110,000 adults. From this large dataset, they developed a model designed to detect structural heart disease using just a single-lead ECG - similar to what a smartwatch can capture. To simulate real-world smartwatch conditions, they isolated one lead from the 12-lead ECG and introduced artificial “noise” to mimic the random interference often seen in wearable device recordings.
The model was then externally validated using ECG data from patients at community hospitals and participants in a large population-based study in Brazil. Finally, to test its real-world performance, the researchers prospectively enrolled 600 participants who completed 30-second single-lead ECGs using a smartwatch, allowing the team to evaluate the algorithm’s diagnostic accuracy in everyday settings.
The analysis found:
- When tested on single-lead ECGs collected from hospital-grade equipment, the AI model performed exceptionally well, achieving 92 % accuracy on a standard performance scale (with 100 % representing perfect classification).
- In the real-world study using smartwatch-recorded ECGs from 600 participants, the algorithm maintained strong performance, correctly detecting structural heart disease with 88 % accuracy.
- It also demonstrated solid clinical utility, identifying most individuals with heart disease (86 % sensitivity) and showing high reliability in ruling out disease, with a 99% negative predictive value.
On its own, a single-lead ECG is limited; it can’t replace a 12-lead ECG test available in health care settings. However, with AI, it becomes powerful enough to screen for important heart conditions. This could make early screening for structural heart disease possible on a large scale, using devices many people already own.
Rohan Khera, M.D., M.S., Study Senior Author and Director, CarDS Lab
Study Background, Design, and Methodology
To develop the AI-ECG algorithm, researchers used a dataset of 266,054 12-lead ECGs collected from 110,006 patients who received care at Yale New Haven Hospital between 2015 and 2023. Each ECG was paired with an echocardiogram to determine the presence or absence of structural heart disease, allowing the AI to learn from confirmed diagnoses.
The model was then externally validated using two independent datasets:
- 44,591 adults who received care at four community hospitals, and
- 3014 participants from the Brazilian Longitudinal Study of Adult Health (ELSA-Brasil) - a large-scale population study focused on chronic conditions, especially cardiovascular disease and diabetes.
To better prepare the algorithm for real-world smartwatch use, researchers added artificial signal interference - comparable to static or fuzz - during training. This made the AI more robust in interpreting lower-quality or noisy ECG signals, which are common with wearable devices.
In the prospective study phase, 600 participants were recruited to test the AI’s performance in a real-world setting. Each wore a smartwatch equipped with a single-lead ECG sensor and recorded a 30-second ECG on the same day they received a standard echocardiogram.
The median age of participants was 62. Roughly half were women, with the racial and ethnic breakdown as follows: 44 % non-Hispanic white, 15 % non-Hispanic Black, 7 % Hispanic, 1 % Asian, and 33 % other or mixed backgrounds. About 5 % of participants were found to have structural heart disease based on the ultrasound results.
Study limitations included the relatively low number of confirmed cases of structural heart disease in the prospective sample and a noted rate of false positives, which could affect generalizability and clinical application.
We plan to evaluate the AI tool in broader settings and explore how it could be integrated into community-based heart disease screening programs to assess its potential impact on improving preventive care.
Arya Aminorroaya, M.D., M.P.H., Study Author, Internal Medicine Resident, Yale New Haven Hospital