GraphEP: Cardiac Electrophysiology as Spatiotemporal signals flowing on Graphs

Cardiac electrophysiology (EP) simulations generate high-dimensional spatiotemporal signals over patient-specific anatomical domains, but the scarcity of publicly available datasets on realistic cardiac geometries limits the development and systematic evaluation of machine-learning surrogates. We introduce GraphEP v1, a reproducible data-generation pipeline and graph-based benchmark for learning the propagation of cardiac electrical activity on realistic full-heart and left-ventricular meshes. GraphEP v1 builds upon openCARP and a standardised electrophysiological calibration workflow that automatically adapts tissue conductivities to mesh resolution to reproduce prescribed conduction velocities.GraphEP v1 comprises 732 transmembrane-potential simulations across 20 different real left ventricles, covering healthy and pathological tissue substrates and multiple pacing configurations, with approximately 40,000 spatial nodes per simulation. Each simulation is distributed as a PyTorch Geometric graph containing anatomical coordinates, fibre-aligned geometric and electrophysiological features, tissue properties, stimulus information, graph connectivity, and node-wise transmembrane-potential trajectories. The dataset is organised to support controlled evaluation of generalisation across anatomical geometries, pacing sites, and tissue substrates, and is released together with the complete simulation and preprocessing pipeline to facilitate reproducibility and extension to additional geometries and electrophysiological conditions.Alongside the dataset, we provide a minimal benchmark for graph-based spatiotemporal regression, in which models learn the complete transmembrane-potential trajectory over the cardiac domain from anatomical, tissue, and stimulation information. The benchmark defines reproducible data splits, training and evaluation procedures, and baseline models that developers can use as a starting point for investigating more advanced graph neural networks, neural operators, and spatiotemporal architectures. In this setting, GraphEP v1 exposes several challenges that are relevant beyond cardiac modelling, including sharp temporal transitions, spatially coupled dynamics, long-range propagation, and geometric domain shift.By combining realistic electrophysiological simulations, graph-native data representations, reproducible generation tools, and an extensible learning benchmark, GraphEP v1 aims to provide a common testbed for developing machine-learning surrogates of cardiac electrophysiology and, more broadly, methods for learning complex spatiotemporal dynamics on irregular anatomical domains. ! Important !Download the files and follow the content of README.md, it contains instructions to prepare the dataset, use the data generation pipeline and reproduce the training of a geometry informed neural operator (GINO).