CSPANet, a novel deep learning architecture, achieved 92.6% accuracy, 79.0% sensitivity, and 95.0% specificity in differentiating concealed accessory pathways and AVNRT from sinus rhythm ECGs.
Does CSPANet improve the prediction of concealed accessory pathways and AVNRT from sinus rhythm ECGs compared to classical CNNs?
A novel deep learning architecture, CSPANet, demonstrated high accuracy in differentiating concealed accessory pathways and AVNRT from sinus rhythm ECGs, potentially aiding in ablation planning.
ABSTRACT Background Concealed accessory pathways (CAP) and atrioventricular nodal reentry tachycardia (AVNRT) represent diagnostically challenging forms of paroxysmal supraventricular tachycardia, with conventional sinus rhythm ECGs often failing to reveal characteristic abnormalities. Methods We developed CSPANet, a novel deep learning architecture that integrates a Channel and Spatial Parallel Attention (CSPA) module for enhanced ECG feature extraction. The architecture features parallel processing through two specialized attention mechanisms: a channel attention submodule that adaptively weights clinically significant ECG leads using complementary feature pathways, working in concert with a spatial attention submodule that captures essential morphological patterns through synergistic multi‐scale pooling and convolutional feature extraction. Results In a comparative study of nine classical CNNs, ResNet50 demonstrated superior performance, achieving the highest sensitivity and specificity and validating the efficacy of residual learning for this task. The proposed CSPANet, integrating our novel channel and spatial parallel attention (CSPA) mechanism, achieved a test set accuracy of 92.6%, sensitivity of 79.0%, specificity of 95.0%, and precision of 79.7%, surpassing all other representative attention mechanisms. Ablation studies confirmed the individual and synergistic contributions of the CSPA and Stem modules, with their combined integration yielding the most significant performance gains, including an 11.7% increase in sensitivity and an 8.8% increase in precision over the baseline ResNet20 model. Conclusion CSPANet's ability to differentiate CAP and AVNRT from sinus rhythm ECGs offers a transformative clinical tool, facilitating optimized ablation planning and enhancing procedural safety. By addressing a key diagnostic gap, this approach underscores the potential of deep learning to refine arrhythmia management.
Wang et al. (Mon,) conducted a other in Concealed accessory pathways (CAP) and atrioventricular nodal reentry tachycardia (AVNRT). CSPANet (Channel and Spatial Parallel Attention network) vs. Classical CNNs (including ResNet50 and ResNet20) was evaluated on Test set accuracy for differentiating CAP and AVNRT from sinus rhythm ECGs. CSPANet, a novel deep learning architecture, achieved 92.6% accuracy, 79.0% sensitivity, and 95.0% specificity in differentiating concealed accessory pathways and AVNRT from sinus rhythm ECGs.