Application of Graph Neural Networks on ECG Data: A Systematic Literature Review.

Area of Science:

• Cardiology
• Artificial Intelligence
• Signal Processing

Background:

• Geometric Deep Learning (GDL) offers novel methods for analyzing structured data.
• Representing Electrocardiogram (ECG) data, especially multi-lead, poses challenges due to non-Euclidean spatial relationships.
• Graph Neural Networks (GNNs), successful in Electroencephalogram (EEG) analysis, are increasingly applied to ECGs.

Purpose of the Study:

• To systematically review Geometric Deep Learning approaches for ECG data analysis.
• To compare existing methods and identify research gaps and future directions.
• To focus on GNN-based methods due to their prevalence in the literature.

Main Methods:

• Systematic literature review of Geometric Deep Learning applications in ECG.
• Analysis focused on Graph Neural Network methodologies for ECG signal processing.
• Categorization of approaches based on ECG lead usage, application, and network architecture.

Main Results:

• Graph Neural Networks are the predominant GDL approach for ECG analysis.
• Limited diversity in applications for 12-lead ECG, primarily focused on arrhythmia recognition.
• Existing single- or two-lead approaches show high performance, questioning the added value of complex multi-lead spatial modeling for this specific use case.
• Lack of consensus on optimal topology, feature sets, and network architectures for GDL in ECG.

Conclusions:

• Future research should explore a wider range of applications for 12-lead ECG GDL models.
• Further investigation is needed to determine the benefits of multi-lead spatial modeling for arrhythmia detection.
• Proposed future work includes application-driven research for explainability and fundamental research for theoretical insights and physiological integration.