Background: Electrocardiogram (ECG) analysis plays an important role in detecting heart rhythm disorders. Although deep learning models have achieved high accuracy in ECG classification, they often fail to generalize well when tested on data collected from different hospitals, devices, or patient groups. In addition, the unequal distribution of ECG classes makes it difficult for these models to accurately recognize rare arrhythmias. Methodology: In this thesis, an Adaptive Reweighting Ensemble Distillation (ARED) framework is proposed to improve cross-domain ECG arrhythmia classification. The proposed method combines adaptive class reweighting, ensemble learning with multiple teacher models, and knowledge distillation. Different teacher models learn complementary information by focusing on different ECG classes, and this knowledge is transferred to a lightweight student model. The framework is evaluated using a leave-one-domain-out strategy, where the model is trained on two datasets and tested on a previously unseen dataset. Results: Experimental results on the Chapman--Shaoxing, PTB-XL, and Georgia ECG datasets demonstrate that the proposed framework achieves high classification performance under cross-domain conditions, with accuracies of 97.6%, 91.4%, and 99.1%, respectively. The results also show strong Macro-F1 and AUPRC scores, confirming reliable recognition of different arrhythmia categories across multiple domains. Conclusion: Overall, the proposed ARED framework provides an effective and computationally efficient solution for domain-generalized ECG diagnosis by combining adaptive class weighting, diverse teacher models, and multi-teacher knowledge distillation.
