Explaining deep learning for ECG using time-localized clusters.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Deep learning (DL) models, particularly convolutional neural networks (CNNs), have shown promise in analyzing electrocardiogram (ECG) data.
• However, the interpretability of these complex models remains a significant barrier to their widespread clinical adoption and understanding of underlying electrophysiological mechanisms.

Purpose of the Study:

• To introduce a novel post-hoc explainability method for CNNs applied to ECG analysis.
• To enhance the interpretability of DL models in cardiology, facilitating clinical trust and providing deeper electrophysiological insights.

Main Methods:

• The proposed method clusters feature activations from the final residual blocks of a 1D-ResNet.
• Each ECG is segmented into a sequence of clusters, with assignment entropy quantifying per-timestamp uncertainty.
• Evaluation was performed on PTB-XL super-class classification and CODE-15% age regression using 10-fold cross-validation.

Main Results:

• Cluster proportions demonstrated a correlation with predicted labels and aligned with key ECG landmarks (P/QRS/T/TP).
• A random forest model trained on cluster proportions achieved 94.9% accuracy in reproducing CNN predictions.
• The method revealed class-dependent representational stability through encoder uncertainty, outperforming Grad-CAM in certain aspects.

Conclusions:

• Time-localized clusters successfully extracted physiologically relevant structures from CNN activations.
• The method faithfully summarizes discriminative information captured by the encoder.
• This provides an architecture-light, post-hoc tool for auditing CNN-based ECG models and identifying potential label quality issues.