Abstract 4145351: Wearable-Echo-FM: An ECG-Echo Foundation Model for Single Lead Electrocardiography

Background: Single-lead electrocardiograms (ECGs) are widely available on wearable and portable devices and can be leveraged as screening tools for structural heart disorders (SHDs) with novel deep learning ECG applications. However, scope of model development is limited by the need for large numbers of labeled examples. We present a novel approach to encode unstructured echocardiographic information on 1-lead ECGs to enable the design of wearable-adapted artificial intelligence (AI)-ECG models with efficient development for detecting SHDs. Methods: We linked 274057 unique ECGs from 77378 patients with reports of echocardiograms done within 30 days across a large, diverse health system (2015-2019). We extracted lead I ECG signals that are typically acquired on wearable devices. We defined 2 discrete encoders, a convolutional neural network (CNN) for ECGs and a RoBERTa-based text encoder for echocardiographic reports. We applied contrastive pre-training for 200 epochs to learn a shared signal-text embedding. In a transfer learning experiment involving an age- and sex-matched training pool of temporally-distinct ECGs all from 2019-2023, we trained sequentially smaller amounts to detect LVEF<40%, LVDD (left ventricular diastolic dysfunction), and a composite measure of LVEF<40%, moderate or severe aortic stenosis, aortic regurgitation, mitral regurgitation, and/or severe LVH (IVSd ≥15 mm) using our Wearable-Echo-FM pre-trained versus standard single-lead ECG CNN model. We examined the performance of these models on the same test set. Results: There were a total of 11923 patients (1 ECG per patient, age, 65.3±17.2 2016 (17%) women) in the test set. For models trained on 100% data, the Wearable-Echo-FM model achieved AUROCs comparable to a standard CNN for LVEF < 40% (0.894 vs 0.884), LVDD (0.849 vs 0.843) and a composite of SHDs (AUROC of 0.887 vs 0.869) in the test set. In few-shot learning experiments that only used 0.5% of the training data (~1000 ECGs), the pre-trained model achieved an AUROC of 0.855 vs. 0.548 for the standard model for LVEF < 40%, 0.819 vs. 0.582 for LVDD, and 0.863 vs. 0.496 for the composite SHD. Conclusions: We propose an ECG-Echo foundation model that learns shared embeddings between 1-lead ECGs and complex structural cardiac phenotypes recorded on unstructured echocardiographic reports. Our approach enhances the efficiency of developing AI algorithms for SHD screening on wearable and portable devices.