Conditional Adversarial Multi-Task Deep Learning for Robust Cross-Site Generalization in 12-Lead ECG Classification

While deep learning techniques have greatly advanced automated electrocardiography (ECG) interpretation, models trained on homogeneous, single-site datasets often fail to generalize due to confounding batch effects arising from differences in equipment, patient demographics, and acquisition protocols. To address this, we propose a Bayesian-inspired conditional adversarial multi-task framework that simultaneously optimizes ECG classification and removes confounding variables, with the adversarial site-ID serving as a proxy variable to ensure marginalization over the learned conditional distribution of site-specific variability. Compared with the baseline, our foundation model achieved better generalization across all key metrics on both the internal and public datasets. t-SNE visualizations of the learned latent space further confirmed successful de-confounding, revealing well-mixed, site-invariant clusters. This straightforward approach thus offers a robust, interpretable solution for scalable ECG classification across heterogeneous clinical environments.