ECGFM: A foundation model for ECG analysis trained on a multi-center million-ECG dataset

Abstract

The electrocardiogram (ECG) is widely used for diagnosing heart conditions due to its cost-effectiveness, non-invasiveness, and accessibility. Between 2014 and 2017, the First Affiliated Hospital of Zhengzhou University collected over a million clinical ECGs from diverse primary hospitals, each accompanied by initial diagnostic results. Effectively utilizing this vast dataset with potential label inconsistencies is a key challenge. In this study, we introduce ECGFM, a foundation model pre-trained on over a million clinical ECGs to achieve deep ECG comprehension. ECGFM comprises a convolutional encoder, a transformer decoder, and task-specific heads, pre-trained through three complementary sub-tasks: (i) contrastive predictive learning for unsupervised representation learning, (ii) normal/abnormal classification, and (iii) diagnostic text generation. Given potential label unreliability, active learning is integrated with the classification task to select key data for re-annotation, enhancing supervision quality. To enable ECGFM’s adaptability to downstream tasks with any-lead inputs, a transferred convolutional encoder is trained to align feature distributions. ECGFM’s effectiveness is evaluated using diverse public datasets, including PTB-XL, Georgia, CPSC, CinC 2020, MITDB, and the “Hefei High-tech Cup” dataset. Fine-tuning ECGFM (training only a task-specific head) delivers strong performance across datasets, nearing fully supervised methods. Additionally, in a one-month online test with 7951 recordings, ECGFM achieved a recall of 0.9335, a precision of 0.9571, and an F1-score of 0.9451, underscoring its robustness and potential for real-world applications.