AI-Powered Research Discovery© 2026 ResearchPod · Berlin, Germany

About Research Blog NotebookLM Speechify Contact Terms Privacy iOS App

Characterising cardiac tissue properties with graph neural networks | ResearchPod

arxiv

Characterising cardiac tissue properties with graph neural networks

Ching-En Chiu, Yoo Ri Kim, Magdi Saba, Danilo Mandic et al./Aug 16, 2026/Original/PDF

Generate podcast

ResearchPod Summary

Research Question

Methodology

Key Findings

Paper Details

arXiv: 2608.15843

Topics

Graph Neural Networks (GNNs)Cardiac ElectrophysiologyPremature Ventricular Complexes (PVCs)Extracellular Potentials (EGMs)

Read next

EMPOWERING GNNS FOR DOMAIN ADAPTATION VIA DENOISING TARGET GRAPHHaiyang Yu et al.6 min
FastPFRec: A Fast Personalized Federated Recommendation with Secure SharingZhenxing Yan et al.9 min
Incremental Residual Reinforcement Learning Toward Real-World Learning for Social NavigationHaruto Nagahisa et al.4 min

Can a deep learning framework combine temporal electrogram signal analysis with spatial graph representations to accurately characterise underlying cardiac tissue properties from spatially sparse, noisy intracardiac measurements?

To evaluate tissue characterisation without relying on intraprocedural arrhythmia episodes, the researchers generated synthetic electrogram (EGM) signals by solving the isotropic Aliev-Panfilov monodomain model on 2D flat squares and curved ellipsoid surfaces. They introduced three types of focal pathologies: fibrosis (via reduced diffusion coefficients), rapid depolarisation (via altered kinetics), and high excitability (via a negative excitation threshold).

The proposed high-resolution framework couples a temporal encoder module (a 1D convolutional neural network that compresses individual EGM traces into latent embeddings) with a spatial module (a graph neural network using graph attention layers). The graph represents sparse electrode nodes and dense tissue nodes, allowing information to be propagated across spatial neighborhoods. The model was trained on flat surfaces and tested on its ability to localize pathologies under conditions of electrode sparsity and noise, as well as its zero-shot and few-shot generalisation to curved geometries.

On 2D flat surfaces, the graph neural network framework substantially outperformed a baseline low-resolution CNN-only approach, achieving average precisions of 0.96 for single-patch fibrosis, 0.97 for rapid depolarisation, and 0.95 for high excitability. The high-resolution model also demonstrated strong robustness against electrode sparsity, maintaining reliable performance even when half of the electrodes were masked.

When transferred to curved surfaces in a zero-shot setting, the model successfully identified fibrosis and rapid depolarisation but struggled with high excitability. However, few-shot fine-tuning using just 10 curved surface simulations restored exceptional performance across all pathology types, yielding average precisions above 0.97 and highlighting the framework's strong generalisation potential.

Significance

By localizing pathological tissue substrates from sparse measurements during normal sinus rhythm, this approach could help clinicians plan ablation targets for premature ventricular complexes even when arrhythmias are infrequent or non-inducible during procedures, significantly reducing procedural time and improving success rates.

Key Terms

Premature ventricular complex (PVC): An early ventricular depolarisation originating outside the normal conduction pathway that can cause symptoms or cardiomyopathy in high burdens.
Extracellular potential (EGM): Electrical signals measured by catheters placed on the endocardium that reflect underlying cardiac action potentials.
Graph neural network (GNN): A class of deep learning models designed to perform message-passing operations across data structured as interconnected nodes and edges.
Aliev-Panfilov model: A mathematical monodomain model used here to simulate cardiac action potential propagation and transmembrane potentials on 2D surfaces.
Zero-shot generalisation: The ability of a machine learning model to perform successfully on unseen target domains or geometries without any additional training or fine-tuning.

AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.