Title: EVL-ECG: Streamlining ECG Analysis via Multi-Dimensional Heterogeneous Knowledge Distillation

Abstract:

While high-precision ECG interpretation is becoming increasingly dependent on large-scale foundation models, their practical application in clinical edge settings is currently obstructed by prohibitive computational requirements. Although knowledge distillation (KD) offers a viable pathway to address this bottleneck, conventional techniques struggle to translate the intricate spatio-temporal dependencies of ECG signals when moving between disparate architectures. To overcome these limitations, this study introduces EVL-ECG, a specialized framework engineered for cross-architecture distillation of cardiac diagnostic logic. The proposed method incorporates three distinct ECG-specific advancements: first, Multi-Head Cross-Attention Alignment, which resolves architectural inconsistencies to retain detailed morphological characteristics; second, Optimal Transport-based Visual Feature Matching, which employs optimal transport theory to sustain global structural integrity across ECG leads, even when token representations do not align; and third, Geometric Intra-Architecture Relation Matching, designed to extract and transfer the teacher model’s underlying diagnostic reasoning. Benchmark results indicate that EVL-ECG surpasses current baselines by achieving gains of up to 2.4% in AUC and 1.1% in clinical accuracy. Furthermore, the framework facilitates the creation of an efficient 2B-parameter ECG foundation model, making it well-suited for deployment in clinical environments with limited resources.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC