ProtoECGNet: Case-Based Interpretable Deep Learning for Multi-Label ECG Classification with Contrastive Learning.

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Deep learning models for electrocardiogram (ECG) classification achieve high performance.
• Clinical adoption is hindered by a lack of transparent and faithful model explanations.
• Existing post hoc explanation methods, like saliency maps, may not accurately reflect model decision-making processes.

Purpose of the Study:

• To introduce ProtoECGNet, a novel prototype-based deep learning model for interpretable, multi-label ECG classification.
• To enable faithful, case-based explanations for ECG diagnoses by grounding decisions in learned prototypes.
• To develop a transparent alternative to black-box deep learning models in clinical settings.

Main Methods:

• Developed ProtoECGNet, a multi-branch deep learning architecture integrating 1D and 2D CNNs with global and time-localized prototypes.
• Implemented a structured prototype loss for multi-label learning, incorporating clustering, separation, diversity, and a novel contrastive loss.
• Trained and evaluated the model on the PTB-XL dataset, covering all 71 diagnostic labels.

Main Results:

• ProtoECGNet achieved competitive performance compared to state-of-the-art black-box models on multi-label ECG classification.
• The model provided structured, case-based explanations, enhancing interpretability.
• Clinician review confirmed that the learned prototypes were representative and clear, validating their quality.

Conclusions:

• Prototype learning can be effectively scaled for complex, multi-label time-series classification tasks like ECG analysis.
• ProtoECGNet offers a practical pathway toward transparent and trustworthy deep learning models for clinical decision support.
• The model's interpretability and performance suggest significant potential for improving AI adoption in healthcare.