Cross-domain transfer learning strategy enhances interpretability of deep learning model explanations

Area of Science:

• Artificial Intelligence
• Biomedical Engineering
• Cardiology

Background:

• Deep neural networks (DNNs) are crucial for clinical decision-making, but their lack of transparency hinders adoption.
• Explainable AI (xAI) methods offer insights into DNN predictions but often lack clinical interpretability.
• Atrial fibrillation (AF) detection from electrocardiograms (ECGs) requires models that are both accurate and interpretable.

Purpose of the Study:

• To investigate if inductive transfer learning (TL) can enhance domain-specific feature separation in a deep learning model for AF detection.
• To improve the clinical interpretability of xAI explanations for ECG-based AF classification.

Main Methods:

• Utilized a two-branch convolutional neural network (xECGArch) for ECG analysis.
• Applied inductive TL by pre-training each branch on domain-specific tasks (P wave detection for morphology, RR interval variability for rhythm).
• Fine-tuned the model for AF classification using iterative layer freezing and analyzed explanations with Deep Taylor Decomposition (DTD).

Main Results:

• Fine-tuning achieved high accuracy (85.70%–95.23%), comparable to existing methods.
• DTD analysis revealed that morphology pre-training focused explanations on P waves, while rhythm pre-training focused on R peaks.
• Domain specificity of explanations increased with more frozen layers during fine-tuning.

Conclusions:

• Inductive TL effectively promotes domain-specific feature attribution in DNNs for ECG analysis.
• This approach enhances the alignment of xAI-generated explanations with clinically relevant ECG features.
• Improved interpretability of AI models can facilitate their clinical integration for conditions like AF.