One model, dual tasks: a novel distributionally adaptive learning framework for ECG classification and generation addressing intra- and inter-patient variability

Electrocardiogram (ECG) classification under the inter-patient paradigm remains a critical challenge due to significant intra- and inter-patient variability in cardiac signals. It causes domain shift and data imbalance issues that hinder model generalization across diverse patient cohorts, limiting the accuracy and robustness of existing classification approaches. We propose a novel distributionally adaptive learning framework that simultaneously performs ECG classification and label-supervised synthetic ECG generation using a generative adversarial network (GAN), addressing these challenges comprehensively. Our approach integrates a transformer-based generator with a convolutional discriminator to generate synthetic ECG signals, effectively alleviating data imbalance issues. To further mitigate domain discrepancy and improve generalization across unseen patient recordings, we introduce single-heartbeat test-time adaptation (TTA). We rigorously evaluate our model using the MIT-BIH Arrhythmia Database, MIT-BIH Supraventricular Database, and INCART Arrhythmia Database. Results outperform state-of-the-art models across intra- and inter-patient settings, highlighting the effectiveness of synthetic ECG augmentation and the robustness of our model in real-world clinical scenarios. It paves the way for personalized and adaptive ECG-based diagnostics, underscoring the potential of generative models for advancing digital health solutions in cardiovascular medicine.