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

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Not quantified with formal effect size metrics in the abstract; described qualitatively as increased domain specificity with greater layer freezing

• Primary Outcome: Domain-specific feature attribution in DTD explanations: morphology pre-training directed relevance toward P waves; rhythm pre-training concentrated explanations on R peaks. Domain specificity increased as more layers were frozen during fine-tuning.

• Nnt Or Sensitivity: Sensitivity and specificity not reported. Classification accuracy range: 85.70%–95.23% across fine-tuning configurations. No NNT calculable from available data.

• Confidence Interval: Not reported

Clinical Application

Currently a proof-of-concept methodology. Clinical deployment would require: (1) prospective validation on diverse, demographically representative ECG datasets; (2) formal sensitivity/specificity benchmarking against validated AF detection standards; (3) clinician usability evaluation of DTD explanations; (4) integration into existing ECG reporting workflows; (5) regulatory approval as a Software as a Medical Device (SaMD). Computational requirements for DTD at inference scale are not addressed. AF affects approximately 2–3% of the Australian population and is a leading cause of stroke. The RACGP supports opportunistic AF screening in primary care using single-lead ECG devices. The TGA regulates AI-based diagnostic tools as SaMD under the Therapeutic Goods Act; any clinical deployment of xECGArch would require TGA conformity assessment. There is no current PBS listing for AI-assisted ECG interpretation. The Australian Digital Health Agency's national digital health strategy supports AI integration in diagnostics, but requires demonstrated clinical safety and explainability — making this study's focus on interpretability directly relevant to the Australian regulatory and clinical governance context. External validation on Australian ECG datasets (e.g., from the Baker Heart and Diabetes Institute or Australian e-Health Research Centre repositories) would be necessary before local adoption. Patients undergoing automated ECG analysis for AF detection, particularly in settings where clinical explainability of DNN outputs is required for physician oversight — e.g., remote monitoring, wearable ECG devices, emergency department triage, and primary care AF screening programs

Abstract

Clinical decision-making increasingly relies on deep neural networks (DNNs), yet their deployment in practice requires transparent and interpretable predictions. Explainable artificial intelligence (xAI) methods can identify input regions relevant to a model's decision, but their clinical interpretability remains limited. In this study, we investigated whether inductive transfer learning (TL) can reinforce domain-specific feature separation in xECGArch, a two-branch convolutional neural network for atrial fibrillation (AF) detection from electrocardiograms (ECGs). Each branch was pre-trained on a task aligned with its designated feature domain, P wave detection for morphology and RR interval variability prediction for rhythm, then fine-tuned on binary AF classification using an iterative layer freezing schedule. Deep Taylor decomposition (DTD) was applied to analyze explanations across all configurations. Fine-tuning accuracy ranged from 85.70% to 95.23%, remaining comparable to the original xECGArch architecture and previous TL-based approaches. However, DTD analysis demonstrated that morphology pre-training directed relevance toward P waves, whereas rhythm pre-training concentrated explanations on R peaks, with domain specificity increasing as more layers were frozen. These findings suggest that inductive TL can encourage domain-specific feature attribution in DNNs, improving the alignment of post-hoc explanations with clinically meaningful ECG regions.

References

1. 1.Zannini, M., Hammer, A., Malberg, H., & Schmidt, M. (2026). Cross-domain transfer learning strategy enhances interpretability of deep learning model explanations. Scientific Reports. https://doi.org/10.1016/j.bspc.2018.03.016 [Note: DOI metadata inconsistency identified — the DOI string corresponds to an Elsevier Biomedical Signal Processing and Control publication; readers should verify the correct DOI via PubMed ID 42342760 before citing.]