Interpreting Deep Neural Networks for Single-Lead ECG Arrhythmia Classification.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Cardiac arrhythmia is a major cause of death, requiring early diagnosis.
• Traditional ECG analysis by cardiologists is limited by expert availability.
• Deep learning (DL) offers scalable arrhythmia diagnosis but lacks clinical interpretability.

Purpose of the Study:

• To develop interpretable DL models for cardiac arrhythmia classification.
• To correlate DL model outputs with specific ECG segments for clinical validation.

Main Methods:

• Applied Gradient-weighted Class Activation Mapping (Grad-CAM) for Convolutional Neural Network (CNN) interpretability.
• Utilized input deletion masks for Long Short-Term Memory (LSTM) model saliency analysis.
• Validated model competence against established baselines.

Main Results:

• Saliency visualizations provided insights into DL model predictions.
• Interpreted model outputs aligned with established medical literature on arrhythmia classification.
• Demonstrated the clinical relevance of DL interpretability in cardiology.

Conclusions:

• Developed and validated novel interpretability methods for DL-based arrhythmia detection.
• Enhanced clinical trust and utility of AI in diagnosing cardiac arrhythmias.
• Facilitated better correlation between AI findings and medical expertise for patient care.