¹Institut für Medizinische Informatik, Universitätsmedizin GöttingenCampus-Institut für Data Science (CIDAS), Georg-August-Universität Göttingen, Göttingen, Germany
²University Medical Center Göttingen, Department of Medical Informatics, Göttingen, Germany
³Clinic of Cardiology and Pneumology, Göttingen, Göttingen, Germany
⁴University Medical Center Göttingen, Göttingen, Germany
⁵Technical University of Denmark, Kongens Lygnby, Denmark
⁶Institute for Predictive Deep Learning in Medicine and Healthcare, Justus-Liebig Universität, Gießen, Germany

Text

Introduction: The application of Artificial Intelligence (AI) in clinical research is progressing rapidly. In the context of electrocardiogram (ECG) analysis, conventional approaches treat signals as waveforms or tabular data. Graph Neural Networks (GNNs) are particularly well-suited for modelling data with complex, non-Euclidean relationships, such as biosignals, and their intricate dynamics. However, their application in the clinical setting is limited, as regulatory frameworks like the GDPR or EU AI Act demand transparency and accountability for AI. Therefore, it is essential to integrate explainable AI (XAI) algorithms that are designed to provide insights into AI-driven decision-making processes.

Methods: In this study, we present a novel GNN classification pipeline tailored for ECG and showcase its applicability on the open-source PTB-XL dataset (~21k ECGs). To model the physiological structure of cardiac signals more accurately, each of the 12-lead signals is encoded into a series of sequentially and spatially connected nodes and edges representing the physiological and temporal dependencies. These graphs are then used to train a Graph Convolutional Network (GCN) to classify Complete Right Bundle Branch Block (CRBBB) and Complete Left Bundle Branch Block (CLBBB) against control.

To understand the intrinsic mechanisms of the model's decision-making process, the post-hoc explainability method GNNExplainer is applied to the trained GCN model. GNNExplainer identifies a compact subgraph and a subset of node features that are most relevant to a specific prediction. We specifically focus on correctly classified unseen patient samples and identify the most influential nodes and edges that contribute to the model's decision. Finally, the most important leads are visualized and compared to the clinical domain experts' knowledge.

Results: The GCN model demonstrates a high classification performance of CRBBB, CLBBB, and control ECGs, with an accuracy of 0.96, an area under the curve (AUC) of 0.989, and a F1 score of 0.962. Notably, the model not only achieves strong overall performance but also shows high accuracy across all individual classes, with particularly reliable differentiation of pathological from control signals.

GNNExplainer identifies specific signal segments, particularly the QRS complex in ECG leads V1 to V6, which are clinically essential for the diagnosis of CRBBB and CLBBB. This supports the interpretability and reliability of the model in a diagnostic context.

Conclusion: Our findings demonstrate that ECG-GNN models can achieve high diagnostic performance while aligning with clinically relevant features. Future work will aim to extend the approach to a wider range of cardiac conditions, patching clinically relevant features, and investigate other datasets and interpretability methods.

These results highlight the potential of GNNs in combination with XAI to foster clinician confidence in AI-driven decision-support systems, paving the way for trustworthy AI models in clinical cardiology.

The authors declare that they have no competing interests.

The authors declare that an ethics committee vote is not required.

Literatur

[1] Ying R, Bourgeois D, You J, Zitnik M, Leskovec J. GNNExplainer: Generating Explanations for Graph Neural Networks [Preprint]. arXiv. 2019. Version Number: 4. DOI: 10.48550/arXiv.1903.03894
[2] Scarselli F, Gori M, Tsoi AC, Hagenbuchner M, Monfardini G. The graph neural network model. IEEE Trans Neural Netw. 2009 Jan;20(1):61-80. DOI: 10.1109/TNN.2008.2005605
[3] Kipf TN, Welling M. Semi-Supervised Classification with Graph Convolutional Networks. Version Number: 4. 2016.
[4] Wagner P, Strodthoff N, Bousseljot RD, Kreiseler D, Lunze FI, Samek W, Schaeffter T. PTB-XL, a large publicly available electrocardiography dataset. Sci Data. 2020 May 25;7(1):154. DOI: 10.1038/s41597-020-0495-6
[5] Zhang H, Liu W, Chang S, Wang H, He J, Huang Q. ST-ReGE: A Novel Spatial-Temporal Residual Graph Convolutional Network for CVD. IEEE J Biomed Health Inform. 2023 Oct 23. DOI: 10.1109/JBHI.2023.3327025