Key result
Graph neural networks detect AF from ECGs with 98% accuracy, outperforming CNNs while maintaining clinical interpretability.
Why the study?
Does a GNN-based XAI framework improve AF detection performance and interpretability compared to CNN and CNN-LSTM models in ECG recordings?
Does a GNN-based XAI framework improve AF detection performance and interpretability compared to CNN and CNN-LSTM models in ECG recordings?
A graph neural network combined with explainable AI provides highly accurate (98.0%) and clinically interpretable atrial fibrillation detection from ECGs.
Supports GNN-XAI integration for AF ECG screening; extends cardiovascular AI beyond conventional CNN-LSTM models.
Atrial fibrillation (AF) is the most common cardiac arrhythmia, markedly elevating the risk of stroke and cardiovascular death. Despite the remarkable efficacy of deep learning models in electrocardiogram (ECG)-based AF identification, their “black-box” characteristics hinder practical use owing to insufficient interpretability. To create and authenticate an explainable artificial intelligence (XAI) framework for AF diagnosis that integrates high precision with clinical interpretability via graph neural networks and advanced explainability methodologies. We established an extensive pipeline utilizing the MIT-BIH AF database, including 25 long-term recordings, each lasting 10 h. Following meticulous ECG preprocessing and R-peak identification, we identified beat-specific characteristics and developed temporal graphs illustrating beat-to-beat connection. Three deep learning architectures were assessed: convolutional neural networks (CNN), CNN-long short-term memory (CNN-LSTM) hybrid, and graph neural networks (GNN). Model interpretability was attained by SHapley Additive exPlanations and gradient saliency analysis using Captum. The GNN model attained exceptional performance with an accuracy, precision, recall, and F1-score of 98.0%, surpassing the CNN-LSTM’s 97.82% accuracy and the CNN’s 96.77% accuracy. The temporal graph form accurately encapsulated beat-to-beat interactions essential for AF identification. XAI analysis indicated that irregular R-R intervals and morphological differences in P-wave patterns were the most distinguishing traits, offering clinically interpretable insights aligned with recognized AF pathogenesis. Our methodology illustrates that GNNs may attain superior AF detection performance while preserving clinical interpretability via XAI strategies. The use of XAI reconciles the disparity between high-performance deep learning models and therapeutic relevance, possibly expediting AI adoption in cardiovascular diagnostics.
No takes yet. Share an insight, caveat, or question.
Maggu et al. (2026) studied this question. A graph neural network model achieved 98.0% accuracy in detecting atrial fibrillation from ECGs, outperforming CNN-LSTM and CNN architectures while maintaining clinical interpretability.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: