Source-linked AI summary
Characterising cardiac tissue properties with graph neural networks
Ching-En Chiu, Yoo Ri Kim, Magdi Saba, Danilo Mandic, Marta Varela
TL;DR
Identifying cardiac tissue properties from sparse intracardiac measurements remains difficult because PVCs can be infrequent or absent during procedures. This paper uses graph neural networks trained on synthetic electrograms to detect abnormal tissue and transfers the model to curved surfaces with few-shot fine-tuning, achieving average precisions above 0.97 across three pathologies and two curved-surface types.
Problem
Clinical identification of PVC ablation targets from sparse electrograms remains largely empirical, while PVCs may be infrequent or absent during procedures.
Method
The study trains graph neural networks on synthetic electrograms generated with a 2D Aliev–Panfilov model to characterise cardiac tissue properties.
Results
Average precision exceeded 0.97 across three pathology types and two curved-surface geometries after fine-tuning on just over 10 graphs.
Takeaways & Limitations
The framework can transfer from flat to curved geometries with minimal fine-tuning and process sparse, irregular electrode layouts relevant to PVC ablation.
Takeaways & Limitations
The results demonstrate transfer across synthetic geometries within one simulator and do not yet establish generalisation to other models, anatomies, or clinical noise.
Abstract
from arXiv · showhide
Characterising electrophysiological properties of cardiac tissue efficiently and accurately from spatially sparse intracardiac measurements is clinically important for localising ablation targets and improving arrhythmia treatment. We developed a graph neural network-based framework trained on synthetic electrogram signals on 2D flat surfaces to identify areas of interest in the context of cardiac ablation for premature ventricular complexes (PVCs). Our method achieved an average precision of 0.96, 0.97, and 0.95 for the detection of single-patch fibrosis, rapid depolarisation and high excitability, respectively. The trained model can then be applied to 2D curved surfaces with few-shot fine-tuning, demonstrating its generalisation capability. Future work will develop this framework further for clinical use in PVC ablation.
1 Introduction
The study aims to characterise myocardial electrophysiological properties from spatially sparse intracardiac electrograms using CNN- and GNN-based methods, identifying high-resolution regions relevant to PVC ablation. This addresses limitations of empirical, operator-dependent target identification and methods focused primarily on ectopic-origin localisation.
- Clinical context: PVCs are common arrhythmias, detected in 40–75% of apparently healthy individuals and affecting approximately 1–4% of the general population.They are early ventricular depolarisations originating outside the normal conduction pathway.
- Clinical challenge: Current EGM-based ablation-target identification remains largely empirical, relies heavily on operator expertise, and is limited by infrequent intracardiac measurements.Activation mapping commonly identifies the earliest local activation site for PVCs.
- Study aim: The framework aims to characterise myocardial electrophysiological properties from spatially sparse EGM signals using convolutional and graph neural networks.GNNs represent cardiac geometries as graphs and update node states through iterative neighbour message passing.
- Study aim: The method identifies, at high spatial resolution, regions of fibrosis, abnormal repolarisation, and enhanced automaticity.These regions represent underlying myocardial properties that may inform future PVC ablation procedures.
- Related approaches: Existing automated PVC-ablation approaches improve activation-mapping reproducibility and efficiency but depend on PVC presence and primarily localise ectopic origins rather than characterising electrophysiological properties.They operate either coarsely from surface ECGs or at higher resolution from intraprocedural EGMs.
2 Methods
The study uses synthetic cardiac simulations to generate sparse electrogram measurements over flat and curved geometries with localized pathological heterogeneities. A CNN encodes temporal electrogram features, while a GNN propagates them across tissue graphs to produce high-resolution pathology classifications and evaluate robustness and generalization.
- Synthetic data generation: 100 locations measure electrograms on the square, whereas 192 locations measure electrograms on curved tissue surfaces with a 0.01 mm outward-normal offset.Signals are paced three times at random nonpathological locations, recorded for 1935 ms, and sampled at 775 Hz.
- Synthetic data generation: Pathological heterogeneities model fibrosis, fast depolarisation, and high excitability through altered diffusion coefficient D, depolarisation parameter k, and excitation threshold a.Fibrosis uses D sampled from [0.02, 0.04] mm2 TU−1 versus baseline D = 0.1 mm2 TU−1; fast depolarisation uses k sampled from [9.6, 16.0] versus baseline k = 8.0; high excitability uses a = −0.025.
- Deep learning framework: The model combines a CNN temporal encoder producing 64-dimensional electrode embeddings with a GNN spatial module that propagates sparse measurements to high-resolution tissue nodes.The graph connects each electrode to its 20 nearest tissue nodes, while tissue nodes connect to other tissue nodes within a 1.0 mm radius.
- Experimental setup: The experimental setup trains on 80% of simulations and tests on 20%, compares the full CNN+GNN high-resolution model with a CNN-only low-resolution baseline, and evaluates noise, sparsity, and curved-surface transfer.Zero-shot transfer uses the flat-surface model directly on curved simulations; few-shot adaptation fine-tunes it on 10 additional lower-curvature simulations before testing on held-out graphs.
3 Results · 2D Square
In 2D square surfaces, the GNN-based high-resolution model localised abnormal tissue properties beyond electrode resolution and outperformed the low-resolution CNN model across key metrics. The HR model remained robust to substantial electrode sparsity, while noise affected pathology detection unevenly, particularly for fibrosis.
- 2D Square: The evaluated performance includes average precision, macro F1, abnormal-class precision and recall at threshold 0.5, and ROC AUC for LR and HR models.These results are summarised in Table 1, with example datasets shown in Fig. 2.
- 2D Square: The GNN-based HR model localises abnormal tissue properties in detail beyond electrode resolution and achieves higher F1 scores than the CNN-only LR model.The LR model detects pathological EGMs with good recall but has consistently poor precision because of limited electrode resolution.
- 2D Square: The HR model also outperforms the LR model in threshold-independent average precision and ROC AUC across all pathology types.Average precision is defined as precision averaged over ranked predictions as the threshold varies, weighted by changes in recall.
- 2D Square: Both models perform slightly worse in ROC AUC when detecting multiple fibrosis patches than single-patch cases.The reduced performance reflects the greater difficulty of ranking several dispersed regions.
- 2D Square: The HR model maintains average precision above 70% across all three heterogeneity types up to half of the electrodes being masked.Performance under increasing electrode sparsity is shown in Fig. 3, alongside visual examples in Fig. 2.
- 2D Square: Electrode sparsity reduces the HR model’s ability to identify gaps between closely placed fibrotic patches in the multi-patch case.Fig. 3 reports performance for both random and regular electrode downsampling, with error bars over three random-sampling rounds.
- 2D Square: Noise affects pathology detection unevenly, with fibrosis particularly sensitive for both LR and HR models, while examples include noisy and noise-free signals.The passage also notes varying noisy-signal performance for high excitability and fast depolarisation detection.
2D Curved Surfaces
On curved surfaces, the model generalised unevenly without adaptation, detecting fibrosis and fast depolarisation above 0.850 average precision but not heightened excitability effectively. Fine-tuning on just over 10 graphs raised average precision above 0.97 across all three pathologies and both curvature settings, with better performance at lower curvature.
- Zero-shot performance: Above 0.850 average precision was achieved for fibrosis (low D) and fast depolarisation (high k) in zero-shot curved-surface detection.The model identified approximate locations, although pathology shapes were inaccurate.
- Zero-shot performance: Heightened excitability (negative a) did not generalise well when the flat-surface-trained model was applied directly to curved surfaces.
- Few-shot fine-tuning: Above 0.97 average precision was achieved across all three pathology types and both curved-surface settings after fine-tuning on just over 10 graphs.Few-shot results were visualised on two held-out graphs.
- Few-shot fine-tuning: Higher performance was obtained in the lower-curvature setting than in the higher-curvature setting after fine-tuning.
4 Discussion & Conclusions
The CNN/GNN framework characterises cardiac tissue properties from sparse electrode measurements and transfers across electrode layouts and flat-to-curved geometries with limited fine-tuning. Its current evidence is restricted to synthetic transfer within one simulator, motivating evaluation across broader electrophysiological models and realistic patient conditions.
- Contributions: The CNN/GNN method detects fibrosis, abnormal repolarisation, and heightened excitability from electrode measurements, supporting PVC ablation applications.It performs on simulated data and adapts to sparse, irregular electrode distributions and curved geometries with minimal fine-tuning.
- Generalisation: The model trained on flat surfaces operates on unseen curved surfaces directly or with improved performance after minimal fine-tuning for curvature artefacts.Transfer is non-trivial because wave propagation speed and tissue properties depend on curvature.
- Methodological interpretation: Attention-based message passing weights neighbouring nodes by inverse distance, allowing the graph model to incorporate irregular layouts and transfer features across geometries.The authors hypothesise that this mechanism supports generalisability to patient-specific geometries using limited data.
- Limitations: The demonstrated transfer is limited to planar and curved synthetic datasets generated with the same isotropic Aliev–Panfilov model and similar pathology-generation procedure.Thus, the current results establish transfer across synthetic geometries within one simulator.
- Future work: Further work will test other electrophysiological models, heterogeneities, anisotropic conduction, non-zero-thickness geometries, realistic patient anatomies, and clinical EGM noise and artefacts.The long-term goal is to identify PVC ablation targets clinically.