Synapse
⌘+K
Synapse
PulseExploreJournal ClubResearchersJournals
Instagram
HomeJournal ClubExplore
June 25, 2026Scientific ReportsOpen Access

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

View Full Paper
Ask AI
Bookmark
Share

Why the study?

Deployment of deep neural networks in clinical practice requires transparent predictions, but the clinical interpretability of explainable artificial intelligence methods remains limited.

Does inductive transfer learning improve domain-specific feature attribution and interpretability of deep neural networks for atrial fibrillation detection from ECGs?

Population

ECGs for AF detection

Comparison

Inductive transfer learning with domain-specific pre-training vs original xECGArch architecture and previous TL-based approaches

Design

Model development and validation study

Authors

MZMatteo ZanniniAHAlexander HammerHMHagen Malberg

Discussion

Loading...

Member takes

Overview

Transfer learning sustains xECGArch accuracy; leaves open prospective clinical validation before deployment.

Key Points

  • The study aims to determine if inductive transfer learning can enhance the interpretability of deep learning models in atrial fibrillation detection.
  • Used a two-branch convolutional neural network (xECGArch) for AF detection with pre-training on P wave detection and RR interval variability.
  • Employed fine-tuning on binary AF classification with an iterative layer freezing schedule.
  • Applied deep taylor decomposition for analyzing model explanations across configurations.
  • Fine-tuning accuracy ranged from 85.70% to 95.23%, comparable to previous approaches.
  • DTD analysis showed morphology pre-training emphasized P wave relevance, while rhythm pre-training highlighted R peak explanations.
  • Domain specificity of feature attribution increased with more layers frozen.

Structured PICO

Does inductive transfer learning improve domain-specific feature attribution and interpretability of deep neural networks for atrial fibrillation detection from ECGs?

P
Population
19,047 10-second Einthoven lead II ECGs (rhythm dataset), 2,716 ECGs (morphology dataset), and 9,854 ECGs (xECGArch dataset for AF/non-AF classification) from publicly available databases.
I
Intervention
Inductive transfer learning strategy with cross-domain pre-training (P wave detection for morphology branch, SDRR prediction for rhythm branch) and iterative layer freezing during fine-tuning.
C
Comparator
Original xECGArch architecture and configurations with different numbers of frozen layers.
O
Outcome
Binary AF classification performance (Accuracy, F1 score) and domain-specific feature attribution (interpretability) assessed via Deep Taylor decomposition.surrogate

Inductive transfer learning with domain-specific pre-training enhances the interpretability of deep learning models for AF detection by aligning post-hoc explanations with clinically meaningful ECG regions.

Cite This Study

Zannini et al. (2026) studied this question.

synapsesocial.com/papers/6a3d91bf408ebb922448b168https://doi.org/10.1038/s41598-026-59076-8
View Full Paper
Ask AI
Bookmark
Share