Key result
Novel adaptive ECG framework improves multi-label classification accuracy by up to ~14 points.
Why the study?
Automatic multi-label classification of 12-lead ECGs is challenging due to heterogeneous waveform morphology, long-range temporal dependencies, diagnosis-dependent inter-lead relationships, and severe label imbalance.
Does a multi-scale temporal feature fusion and lead-graph attention framework improve multi-label classification performance on 12-lead ECGs compared to existing models?
Does a multi-scale temporal feature fusion and lead-graph attention framework improve multi-label classification performance on 12-lead ECGs compared to existing models?
A novel AI framework utilizing diagnosis-conditioned sparse lead graphs and temporal feature fusion improves the automated multi-label classification accuracy of 12-lead ECGs.
May advance automated multi-label ECG classification; leaves open prospective clinical validation and generalizability.
OBJECTIVE: Automatic multi-label classification of 12-lead electrocardiograms remains challenging because of heterogeneous waveform morphology, long-range temporal dependencies, diagnosis-dependent inter-lead relationships, and severe label imbalance. This study develops an adaptive framework that jointly models temporal features and diagnosis-specific lead interactions. Approach. We propose a multi-scale residual-temporal network incorporating a diagnosis-conditioned, record-specific sparse lead graph. A shared temporal encoder independently extracts dynamic node features from each lead waveform. Learnable diagnosis prototypes participate in edge-score computation before message passing, generating a distinct graph for every recording and diagnostic category. For each target lead, only the top-k relationships are retained, while established lead connections are incorporated as a physiological soft prior with learnable strength. Prototype-guided graph predictions are integrated with global semantic predictions through class-wise gated fusion. Weighted asymmetric loss and validation-based class-specific thresholds address label imbalance. Main results. The model was evaluated on CPSC2018, G12EC, and PTB-XL across five multi-label settings. It achieved Macro-F1 scores of 0.8327 on nine-class CPSC2018, 0.5857 ± 0.0142 on 30-class G12EC, and 0.7782, 0.6421, and 0.4910 on the 5-, 12-, and 23-class PTB-XL tasks, respectively. The corresponding absolute gains over the strongest listed Macro-F1 comparators were 0.0147, 0.1428, 0.0226, 0.1335, and 0.0102. Macro-AUPRC was highest in all five settings, although Hamming Loss on the 23-class PTB-XL task was 0.0025 higher than that of ECG-STAR. Ablation and paired-bootstrap analyses supported the contribution of diagnosis-conditioned graph modelling. Significance. Conditioning sparse lead relationships jointly on each recording and diagnostic category extends fixed or class-agnostic lead graphs and provides an adaptive representation of disease-dependent multi-lead evidence. Clinical generalisability requires validation on independent hospital cohorts.
No takes yet. Share an insight, caveat, or question.
Zhou et al. (2026) studied Multi-label classification of 12-lead electrocardiograms. Multi-scale residual-temporal network with diagnosis-conditioned sparse lead graph vs. Other established baseline models (e.g., ECG-STAR) was evaluated on Macro-F1 scores and Macro-AUPRC across CPSC2018, G12EC, and PTB-XL datasets. An adaptive framework using multi-scale temporal feature fusion and lead-graph attention outperformed baseline models on multi-label 12-lead ECG classification, with Macro-F1 absolute gains up to 0.1428.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: