ECGFounder

ECGFounder is a foundation model for electrocardiogram (ECG) analysis designed to serve as a general-purpose backbone for cardiovascular diagnosis. While deep learning has produced strong task-specific ECG classifiers, most are trained on small, narrowly labeled datasets and generalize poorly across recording domains, devices, and lead configurations. ECGFounder addresses this by pretraining a single model on a very large, expert-annotated corpus and then transferring it to downstream tasks, including the increasingly common single-lead signals captured by wearable devices.

The model was built on the Harvard-Emory ECG Database (HEEDB), using real-world annotations from cardiology experts spanning 150 diagnostic categories. It was developed by a collaboration led by Shenda Hong's group at Peking University together with clinical researchers at Massachusetts General Hospital / Harvard Medical School and Emory University, and first released as a preprint in October 2024. Pretrained checkpoints for both 12-lead and single-lead variants are publicly available.

By coupling the scale of HEEDB with broad diagnostic label coverage, ECGFounder aims to be a reusable starting point for ECG research, lowering the data and compute burden for groups that cannot assemble million-scale labeled datasets of their own.

#Key Features

• Large-scale expert supervision: Pretrained on over 10 million ECGs with 150 expert-annotated diagnostic categories, capturing a far broader label space than typical ECG datasets.
• 12-lead and single-lead variants: A standard 12-lead model plus a single-lead model trained with lead augmentation that simulates axis inversion, enabling deployment on wearable and Holter-style devices.
• Transfer-ready backbone: Released checkpoints support fine-tuning on downstream tasks such as arrhythmia detection and demographic or clinical-event prediction, demonstrated on MIMIC-ECG and PTB-XL.
• Strong cross-domain generalization: Validated externally on CODE-test, PTB-XL, and PhysioNet, maintaining high AUROC across datasets it was not trained on.

#Technical Details

ECGFounder uses a RegNet-based 1D convolutional neural network with stage-wise scaling and bottleneck blocks that combine group convolutions and channel-wise attention. The primary variant has 76.3 million parameters; ablations across 11.7M, 25.6M, and 110M parameter models found 76.3M optimal. Training used 7,519,035 ECGs from 1,319,128 patients, with a held-out set of 834,926 ECGs from 146,570 patients, drawn predominantly from 10-second, 12-lead clinical recordings. On a committee-reviewed internal test set the model reached an average AUROC of 0.968 (95% CI 0.955-0.982), exceeding 0.95 for 80 individual diagnoses. External evaluation produced average AUROCs of 0.981 on CODE-test and 0.924 on PTB-XL, and the single-lead variant scored 0.975 for normal sinus rhythm and 0.957 for atrial fibrillation on PhysioNet data.

#Applications

ECGFounder targets both research and clinical-adjacent workflows. Cardiology and signal-processing groups can fine-tune it for specific tasks—arrhythmia classification, cardiac event detection, demographic inference, or estimation of clinical variables—without training a large model from scratch. The single-lead variant is aimed at consumer wearables and ambulatory monitors, where only one lead is available, while the 12-lead model suits hospital and clinic settings. Downstream studies have already adapted it for laboratory-value estimation and ICD-code-based disease profiling, illustrating its use as a transferable starting point.

#Impact

ECGFounder is one of the first ECG foundation models built at the scale of millions of expert-labeled recordings with broad, externally validated diagnostic coverage, and its public 12-lead and single-lead checkpoints have made it a practical base for subsequent work (for example AnyECG and distillation studies). Its main limitations stem from supervised pretraining on a single institutional database, which may carry annotation and population biases, and from a licensing discrepancy worth noting: the released code and weights on GitHub and Hugging Face are under the MIT license, while the preprint states CC BY 4.0. Even so, by demonstrating that large-scale expert supervision yields a robust, transferable ECG backbone, it helps push biosignal analysis toward the reusable foundation-model paradigm already common in protein and language modeling.