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

Geometric Deep Learning is a modern approach to deep learning that focuses on the assessment of the structure and inherent symmetry of the data. How to best represent Electrocardiograms (ECG) data is unclear. The single-lead ECG can be interpreted as a shift-invariant one-dimensional grid as common for time series. However, the leads' spatial relationships in multi-lead ECGs are non-Euclidean due to the spread of electrical activation along the heart tissue. In the domain of Electroencephalograms (EEG), Graph Neural Networkss are being adopted to represent channel interactions. This method is transferred to ECGs as well, calling for a systematic literature review to compare approaches and identify leading directions and research gaps. Consequently, we conducted this review about Geometric Deep Learning-based approaches to ECG data. However, since all but two approaches used Graph Neural Networkss, we focus more heavily on them. Our results suggest missing diversity in applications for 12-lead ECG. It is mainly used for the recognition of arrhythmias, where single- or two-lead approaches already showed superb performance. It is unclear whether multi-lead spatial modeling would significantly improve classification for this use case. While a variety of approaches from EEG analysis have swept over to ECG, there is no consensus on topology, feature set or network architecture. In consequence, we propose two directions of future work. A more application-driven branch to compare existing approaches and expand on their explainability, and a more fundamental branch to focus on theoretical insights and experiment with simulation data to integrate physiological and anatomical information.