Key result
XAI-ECGNet edges past standard deep learning baselines with a macro-AUROC of ~0.93 for ECG classification.
Why the study?
Although the 12-lead ECG is the primary non-invasive diagnostic tool for cardiovascular disorders, interpretable deep-learning ECG classifiers remain limited, which restricts clinical adoption.
Does the XAI-ECGNet hybrid CNN-SE-BiLSTM framework improve multi-label cardiac diagnosis classification accuracy and provide faithful, clinically consistent explanations from 12-lead ECG signals compared to baseline deep learning models?
Does the XAI-ECGNet hybrid CNN-SE-BiLSTM framework improve multi-label cardiac diagnosis classification accuracy and provide faithful, clinically consistent explanations from 12-lead ECG signals compared to baseline deep learning models?
XAI-ECGNet achieves high multi-label diagnostic accuracy (macro-AUROC 0.9337) on 12-lead ECGs using a compact 4.02M-parameter hybrid architecture while demonstrating quantitatively faithful, clinically consistent explanations across Grad-CAM, SHAP, and Integrated Gradients.
Incremental gains support hybrid ECG model refinement; leaves open prospective clinical validation before any workflow integration.
Cardiovascular disorders cause approximately 17.9 million deaths annually, and the 12-lead electrocardiogram remains the primary non-invasive diagnostic tool, yet interpretable deep-learning ECG classifiers remain limited, restricting clinical adoption. We present XAI-ECGNet, which is based on 1D residual CNNs, squeeze-and-excitation (SE) channel attention on learned, lead-derived feature channels, and bidirectional LSTM (BiLSTM) temporal modelling. On PTB-XL (21,799 ECGs, 18,885 patients), XAI-ECGNet outperforms all of the top performing baselines in terms of macro-AUROC, which is 0.9337, within the range of the baselines’ bootstrap confidence intervals, and four of five diagnostic superclasses have AUROC above 0.90. Ablation results indicate that the full model is within five-seed variability (σ = 0.0023) of the CNN-only backbone (0.9333); per-class analysis reveals that SE attention is crucial for BiLSTM to offer benefit, as it recovers and surpasses CNN-only performance on four out of five classes after BiLSTM alone fails to do so. Three post-hoc methods (Grad-CAM, SHAP, Integrated Gradients) show the model learns clinically consistent representations without explicit clinical priors. We also report quantitative faithfulness evaluation, SHAP background-sample-size justification, and zero-shot cross-dataset validation on CPSC 2018, demonstrating genuine generalisation for NORM and CD, and interpretable label-taxonomy mismatch for STTC. XAI-ECGNet combines accurate multi-label ECG classification with a rigorously and quantitatively assessed explainability framework, but prospective cardiologist validation is required before any claim to clinical use can be made.
No takes yet. Share an insight, caveat, or question.
Metre et al. (2026) studied Multi-label cardiac disorders (Normal Sinus Rhythm, Myocardial Infarction, ST/T-wave Changes, Conduction Disturbance, Hypertrophy) (n=18,885). XAI-ECGNet (explainable hybrid 1D residual CNN, SE channel attention, and BiLSTM model) vs. Canonical deep learning baselines (ResNet1D, Inception1D, bidirectional GRU) and CNN-only backbone was evaluated on Macro-averaged Area Under the Receiver Operating Characteristic curve (macro-AUROC) on the 5-class PTB-XL diagnostic superclass task. XAI-ECGNet achieved a macro-AUROC of 0.9337 on the PTB-XL 5-superclass ECG benchmark, outperforming canonical ResNet1D (0.931) and CNN-only (0.9333) baselines.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: