Physiology-Aware Graph Attention for Early Ventricular Reentry Prediction

Ma{ë}lis Morier, Jairo Rodr{í}guez Padilla, Patrick Gallinari, Maxime SermesantarXiv 2026 · arXiv

TL;DR

This paper introduces a physiology-aware graph attention classifier that predicts whether a cardiac simulation will develop sustained ventricular reentry using only the first 75ms of data. By incorporating physiological descriptors like the diastolic interval into a Graph Neural Network (GNN), the authors achieve over 94% accuracy while providing a massive $5 \times 10^4$ speedup over traditional biophysical simulations.

What Problem Does This Paper Address?

Sudden cardiac death is frequently caused by ventricular tachycardia (VT), where electrical signals become trapped in "reentrant" loops around post-infarct scar tissue. Predicting which patients are at risk is difficult because reentry depends on a complex interplay between scar geometry, tissue recovery times, and the specific location of the electrical stimulus (pacing).

While biophysical "in silico" simulations can model these mechanisms, they are computationally expensive, often taking minutes or hours to simulate a few seconds of heart activity. This makes it impractical to screen thousands of potential pacing sites or scar configurations. The researchers sought a way to predict the long-term outcome of these simulations (reentry vs. no reentry) almost instantly, using only the very beginning of the simulation data.

Methods

The authors represent the cardiac mesh as a graph where nodes are mesh vertices and edges represent connectivity. They utilize a Residual GATv2 (Graph Attention Network) architecture to process this data.

  • Input Features:
    • Nodes: The first five voltage samples (75 ms) of the action potential, capturing the initial paced activation.
  • Edges: Eight attributes including edge length, tissue labels (healthy, scar, or border-zone), stimulation frequency, and—most importantly—the Diastolic Interval (DI) and amplitude calculated from the first two paced beats.
  • Architecture: The model consists of six GATv2 layers. The final layers use residual connections to preserve information flow. Node embeddings are aggregated via global mean pooling and passed to a Multi-Layer Perceptron (MLP) to output a reentry probability.
Overview of the proposed graph classifier pipeline

Datasets

The study relies on synthetic electrophysiology simulations generated using a modified Mitchell–Schaeffer (MS) ionic model.

  • Scale: A total of 8,004 simulations were used (5,000 for training, 2,004 for validation, and 1,000 for a balanced test set).
  • Generalization: The data was split by "anatomy-disjoint" sets. The training and validation sets used 16 heart meshes, while the test set used 2 completely held-out meshes to ensure the model generalizes to new heart shapes rather than just memorizing specific geometries.

Key Insights and Ideas

  • The Importance of Physiology: Ablation studies revealed that the Diastolic Interval (DI) is the most critical feature. DI summarizes local recovery timing; if tissue hasn't recovered enough before the next wave arrives, reentry is more likely. Removing DI caused the largest drop in model performance (F1-score fell to 0.6468).
  • Beyond "Scar Recognition": A major challenge in this field is ensuring a model doesn't just learn to identify "bad-looking" scars. The authors performed a "matched-substrate" analysis, testing the same scar with different pacing locations. The model correctly predicted different outcomes for the same anatomy, proving it understands the dynamic interaction between the stimulus and the substrate.
  • Attention as a Qualitative Tool: By inspecting the GAT attention maps, the authors found that different "heads" of the network focus on distinct regions like the border-zone, the stimulation site, and the early propagation path, offering a degree of interpretability.
Figure 2: Matched-substrate analysis. Lines connect mean reentry probabilities for non-reentry and reentry stimuli with the same substrate.

Key Results

  • High Predictive Power: On the independent test set, the model achieved 94.6% accuracy, 95.4% sensitivity, and a 0.987 ROC-AUC.
  • Extreme Speedup: Forward inference takes roughly 11.35 ms per graph. When compared to the full biophysical simulation pipeline (which takes ~1076 seconds), the GNN approach is approximately 50,000 times faster.
  • Robustness: The model maintained high sensitivity (100% on a specific imbalanced subset) even when tested on data with a lower prevalence of reentry, mimicking clinical conditions.
Figure 3: Selected attention maps and one encoder-output channel for a matched reentry and non- reentry pair.

Limitations and Future Work

  • Synthetic Data Bias: The model was trained on simplified Mitchell–Schaeffer simulations. Future work needs to validate these findings on more complex ionic models and, eventually, clinical patient data.
  • Fixed Protocols: The current dataset uses a fixed pacing cycle length. Real-world clinical "S1-S2" pacing protocols are more varied and complex.
  • Lack of Localization: While the model is excellent at predicting if reentry will occur, it does not yet localize the "critical isthmus" (the specific path of the reentry), which is required for guiding ablation procedures. The authors suggest combining this classifier with localization tasks in future iterations.

Summary updated

Abstract

Sudden cardiac death mostly originates from post-infarct ventricular tachycardia. This reentry is sustained by propagation around scar tissue, but its risk remains difficult to predict from infarct imaging alone. While in silico induction studies can reveal these mechanisms, exploring many substrates and pacing locations remains computationally expensive. We propose a graph attention classifier for early prediction of sustained ventricular reentry from simulated cardiac electrophysiology. Each simulation is represented as a graph with early nodal voltage samples and edge attributes encoding geometry, stimulation frequency, tissue labels, amplitude, and diastolic interval. These descriptors are computed only from the first two paced activations, before sustained reentry is observed. On a balanced independent test set of 1000 simulations, the classifier achieves 94.6% accuracy, 95.4% sensitivity, 93.8% specificity, 94.6% F1-score, 0.987 ROC-AUC, and 0.985 PR-AUC. Ablation experiments show that diastolic interval is the most critical descriptor. Qualitative attention inspection suggests that different heads focus on border-zone, stimulation location, and early propagation path. These results support graph attention as a fast early-prediction approach for large-scale in silico ventricular reentry studies.
BibTeX
@misc{morier2026physiology,
  author = {Ma\{ë\}lis Morier and Jairo Rodr\{í\}guez Padilla and Patrick Gallinari and Maxime Sermesant},
  title = {{Physiology-Aware Graph Attention for Early Ventricular Reentry Prediction}},
  year = {2026},
  eprint = {2610.26430},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.26430},
}
Figure