What question did this study set out to answer?

The aim is to develop an AI model that effectively detects structural heart diseases using 1 lead ECGs.

March 28, 2026Open Access

Wearable Echo FM: An ECG echo foundation model for 1 lead electrocardiography

Key Result

Wearable-Echo-FM detected structural heart diseases from 1-lead ECGs using only 0.5% of training data, markedly outperforming a baseline model (AUROC 0.819-0.863 vs 0.496-0.582).

Key Points

The aim is to develop an AI model that effectively detects structural heart diseases using 1 lead ECGs.
Developed Wearable-Echo-FM foundation model using 194,551 ECG-echo pairs.
Contrastive pre-training was applied using convolutional neural network and RoBERTa encoders.
ECG encoder was fine-tuned with progressively larger ECG datasets to classify cardiac disorders.
Evaluated performance against a randomly initialized CNN on a held-out test set.
Achieved AUROC of 0.894 for LVSD and 0.849 for diastolic dysfunction on full training set.
With only 0.5% of data, model outperformed baseline CNN significantly in all tested conditions.
Results suggest improved efficiency in developing SHD screening models using limited labeled data.

Structured PICO

Does Wearable-Echo-FM improve the detection of structural heart diseases from 1 lead ECGs compared to a randomly initialized CNN?

Population

77,378 adults with 194,551 1 lead ECG and echocardiogram pairs

Intervention

Wearable-Echo-FM (foundation model contrastively pre-trained on 1 lead ECGs and echocardiographic text reports)

Comparator

Randomly initialized convolutional neural network (CNN)

Outcome

Detection of left-ventricular systolic dysfunction (LVSD), diastolic dysfunction, and a composite structural heart disease (SHD)surrogate

Contrastive pre-training of 1 lead ECGs with echocardiographic text reports significantly reduces the amount of labeled data required to develop accurate AI screening models for structural heart disease.

Main Result

Absolute Event Rate: 0% vs 0%

Abstract

Abstract Background Artificial intelligence (AI) models can now detect patterns of structural heart diseases (SHDs) from electrocardiograms (ECGs), though scaling them requires the broader use of 1 lead ECGs that are now ubiquitous in wearable and portable devices. However, model development for these devices is limited by a lack of diagnostic labels for SHDs for wearable ECGs. Methods Here, we present Wearable-Echo-FM, a foundation model that encodes 1 lead ECGs with information from echocardiographic text reports. Using 194,551 1 lead ECG echo pairs from 77,378 adults (2015-2018), we contrastively pre-trained convolutional neural network (CNN) and RoBERTa encoders. The ECG encoder was fine-tuned on a distinct progressively larger ECG set (250 to 250,260 ECGs) to detect different cardiac disorders (i) left-ventricular systolic dysfunction (LVSD), (ii) diastolic dysfunction, and (iii) a composite SHD. This was compared with a randomly initialized CNN, with both approaches evaluated in an independent held out test set. Results With the full training set, Wearable-Echo-FM matched the baseline CNN (AUROC 0.894 vs 0.884 for LVSD; 0.849 vs 0.843 diastolic dysfunction; 0.887 vs 0.869 composite). With only 0.5% (∼1000 ECGs) of data, it markedly outperformed baseline (0.855 vs 0.548; 0.819 vs 0.582; 0.863 vs 0.496, respectively). Conclusions Contrastive pre-training of 1 lead ECGs on echocardiographic text reduces label requirements for label efficient development of SHD screening models on 1 lead ECGs, providing a foundation for future validation on wearable and portable devices.