xECGArch: a trustworthy deep learning architecture for interpretable ECG analysis considering short-term and...

Area of Science:

• Artificial Intelligence in Medicine
• Cardiovascular Disease Diagnostics
• Machine Learning for Healthcare

Background:

• Deep learning models excel at classifying cardiovascular diseases from ECGs but suffer from a lack of interpretability, limiting clinical adoption.
• Existing methods often function as black boxes, hindering trust and validation in critical healthcare applications.

Purpose of the Study:

• To develop a novel deep learning architecture, xECGArch, for interpretable electrocardiogram (ECG) analysis.
• To enhance the clinical applicability of deep learning in cardiovascular disease detection by addressing the black-box problem.

Main Methods:

• xECGArch utilizes two independent convolutional neural networks (CNNs) to analyze short- and long-term ECG features, combined into an ensemble.
• Explainable artificial intelligence (xAI) methods were integrated to provide transparency, with perturbation analysis used to compare 13 xAI techniques.
• The model was specifically parameterized for atrial fibrillation (AF) detection using four public ECG databases.

Main Results:

• xECGArch achieved a 95.43% F1 score in classifying atrial fibrillation (AF) versus non-AF on an unseen ECG test dataset.
• Deep Taylor decomposition was identified as the most trustworthy xAI method for explaining xECGArch predictions.
• The architecture effectively captures short-term (morphology) and long-term (rhythm) features relevant to clinical diagnosis.

Conclusions:

• xECGArch offers an interpretable deep learning approach for ECG analysis, enhancing trust and clinical utility.
• The model's ability to differentiate between short- and long-term features aligns with clinical understanding of cardiovascular conditions.
• Future research will explore the correlation between xECGArch features and clinical findings for improved diagnosis and therapy.