Abstract
Artificial intelligence (AI) has demonstrated significant potential in ECGanalysis and cardiovascular disease assessment. Recently, foundation modelshave played a remarkable role in advancing medical AI. The development of anECG foundation model holds the promise of elevating AI-ECG research to newheights. However, building such a model faces several challenges, includinginsufficient database sample sizes and inadequate generalization acrossmultiple domains. Additionally, there is a notable performance gap betweensingle-lead and multi-lead ECG analyses. We introduced an ECG Foundation Model(ECGFounder), a general-purpose model that leverages real-world ECG annotationsfrom cardiology experts to broaden the diagnostic capabilities of ECG analysis.ECGFounder was trained on over 10 million ECGs with 150 label categories fromthe Harvard-Emory ECG Database, enabling comprehensive cardiovascular diseasediagnosis through ECG analysis. The model is designed to be both an effectiveout-of-the-box solution, and a to be fine-tunable for downstream tasks,maximizing usability. Importantly, we extended its application to lower rankECGs, and arbitrary single-lead ECGs in particular. ECGFounder is applicable tosupporting various downstream tasks in mobile monitoring scenarios.Experimental results demonstrate that ECGFounder achieves expert-levelperformance on internal validation sets, with AUROC exceeding 0.95 for eightydiagnoses. It also shows strong classification performance and generalizationacross various diagnoses on external validation sets. When fine-tuned,ECGFounder outperforms baseline models in demographic analysis, clinical eventdetection, and cross-modality cardiac rhythm diagnosis. The trained model anddata will be publicly released upon publication through the bdsp.io. Our codeis available at https://github.com/PKUDigitalHealth/ECGFounder