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

Figures

• 

  Figure 1. Overview of data flow and CLIP-style pre-training. Patient records from the YNHHS that contained paired TTE reports and ECGs acquired ≤30 days apart were split into 2015 2018 data (A) for contrastive pre-training and 2019–2023 data (B) for downstream fine-tuning and evaluation. Panel A shows pre-training; a CNN encodes 1-lead ECG signals while a RoBERTa-based encoder processes the corresponding TTE text. The two embeddings (ZECG, ZTTE) are pulled together for matching pairs (green arrows) and pushed apart for non-matching pairs (red arrows), aligning both modalities in a shared latent space. The weights learned in this stage serve as initialisation for subsequent fine-tuning on the temporally distinct cohort. Panel B shows fine-tuning and testing; pre-trained weights are transferred to a distinct 2019–2023 cohort, split 80/10/10% (train/validation/test) at the patient level (≤5 ECGs per subject in train; 1 in validation/test). Separate classifier heads are trained for three echo-derived labels (LVEF ≤40%, LV diastolic dysfunction, and composite SHD) using cross-entropy with class weighting and early stopping (patience = 6). To test label-efficiency, the training set is randomly…

• 

  Figure 2. Signal encoder architecture optimized for embedding used in a contrastive framework. The seven-layer 1D CNN model encoded the 10 s 1-lead ECGs. Each convolutional block consists of (Conv, BatchNorm, ReLU, MaxPool), with kernel sizes tapering from 7 to 3 and filters doubling every other layer (16, 32, 64) to capture both broad and fine-grained temporal features. Max-pooling strides of 2 and 4 progressively down-sample the 10 s input while expanding the receptive field. The final convolutional output is flattened and passed through two FC layers (64 and 32 units) with 0.5 dropout, producing a 512-dimensional ECG embedding that feeds the contrastive loss. Abbreviations: CNN, convolutional neural network; ECG, electrocardiogram. Conv, convolution. BatchNorm, Batch normalisation. ReLU, Rectified linear unit activation function. Max pool, maximum pooling layer. FC, fully connected (dense) layer; L2 normalisation, Euclidean (L2) vector norm.

• 

  Figure 3. Model performance across different fractions of data for LVEF ≤40, LVDD, and the composite outcome. Test-set AUROC (mean ± 95% CI) for detecting LVEF ≤40%, LV diastolic dysfunction, and composite SHD as a function of the fraction of labelled training data used for fine-tuning (0.1% to 100%). The pre-trained Wearable-Echo-FM is shown in blue, and the randomly initialised CNN is shown in orange. Abbreviations: LVEF, left ventricular ejection fraction. LVDD, left ventricular diastolic dysfunction. Composite SHD, LVEF ≤40%, moderate or severe aortic stenosis, aortic regurgitation, mitral regurgitation and/or severe left ventricular hypertrophy. AUROC, area under the receiver operating characteristic curve. Wearable Echo FM, pretrained model with weights initialized from the pretraining step. Standard Model, with random weights initialized.