A Comprehensive Explanation Framework for Biomedical Time Series Classification.

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Cardiology

Background:

• Deep learning models are increasingly used for biomedical time-series classification.
• Interpreting these complex models, especially for critical applications like atrial fibrillation (AF) detection, remains a challenge.

Purpose of the Study:

• To develop and evaluate a post-hoc explainability framework for deep learning models in quasi-periodic biomedical time-series classification.
• To apply this framework to AF detection using electrocardiography (ECG) signals.

Main Methods:

• Utilized a state-of-the-art pretrained deep learning model for AF detection.
• Implemented global explanation to analyze model behavior across data classes and identify influential regions in repetitive patterns.
• Applied local explanation to examine specific signal classifications and model outcomes.

Main Results:

• Global explanation confirmed that crucial AF detection features (e.g., R-R interval regularity, P-wave absence) heavily influence the model's decisions, aligning with clinical expert knowledge.
• Local explanation provided insights into network behavior for both correct and incorrect classifications, highlighting informative regions and potential misclassification causes.

Conclusions:

• The proposed explainability framework enhances the understanding of deep learning models in biomedical time-series analysis.
• The findings support the clinical relevance of the identified features for AF detection and offer a method to diagnose model behavior.