Multi-label ECG diagnosis via adversarial view decoupling and hierarchical label constraints
Authors: Meng, W., Wang, Q., Gan, C., Yu, H., Cheng, Y.
Journal: Knowledge Based Systems
Publication Date: 28/02/2026
Volume: 335
ISSN: 0950-7051
DOI: 10.1016/j.knosys.2025.115174
Abstract:Objective: Multi-lead, multi-label electrocardiogram (ECG) diagnosis relies on both rich lead information and inter-label dependencies. However, existing methods often struggle to distinguish critical and redundant features during lead fusion and fail to leverage the hierarchical structure of disease labels. Methods: To address these issues, a novel method termed AVD-HLC is proposed, which is based on adversarial view decoupling and hierarchical label constraints. A multi-scale feature extraction network is first designed to capture morphological and temporal dependencies of ECG, with lead views partitioned according to cardiac anatomical structure. An adversarial view decoupling module is then introduced to decouple view features into consistent and specific representations, with orthogonality constraints imposed to enhance feature fusion. Furthermore, a hierarchical label constraint module is developed to explicitly embed label hierarchies into the classification process. A top-down semantic guidance mechanism promotes knowledge sharing between parent and child labels, thereby improving the recognition of child labels, including rare diseases. Results: Experiments on two public datasets, PTB-XL and SPH, show that AVD-HLC consistently outperforms competitive baselines, achieving up to 1.5% AUC gain and 3% F1 improvement. Significance: By incorporating medical priors and structured feature decoupling, AVD-HLC enables more effective representation learning and interpretable multi-label ECG diagnosis.
Source: Scopus