Automated identification of post-infarction ventricular tachycardia ablation targets with spatio-temporal graph convolutional networks on 3-D electroanatomical manifolds

Q: What are the key findings of this study?

Spatio-temporal graph convolutional networks achieved an AUC of 0.942 for detecting abnormal ventricular potentials, failing to outperform the =0.06).

Q: What is the clinical evidence from this study?

Study design: Observational. Population: Post-infarction ventricular tachycardia (n=9). Intervention: Spatio-temporal graph convolutional network (ST-GCN) vs. Conventional bipolar-voltage rule (<1.5 mV) and gradient boosting (XGBoost) on 38 expert descriptors. Primary outcome: Detection of abnormal ventricular potentials (AVPs) measured by leave-one-patient-out cross-validation AUC-ROC (95% CI 0.888-0.977, p=1.000).

Q: What does this research mean for the field?

Spatio-temporal graph convolutional networks do not outperform standard bipolar-voltage thresholding or gradient boosting for identifying abnormal ventricular potentials because expert annotations are largely determined by bipolar amplitude. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.NEUTRAL.

Amine BOUGARNE

doi:10.21203/rs.3.rs-11035410/v1

September 29, 2026Open Access

Automated identification of post-infarction ventricular tachycardia ablation targets with spatio-temporal graph convolutional networks on 3-D electroanatomical manifolds

View Full Paper

Ask AI

Bookmark

Share

Key result

Graph neural networks fail to outperform standard voltage mapping or XGBoost for detecting abnormal ventricular potentials.

95% CI 0.888-0.977
P=1.000
n=9

Why the study?

Manual annotation of abnormal ventricular potentials during substrate mapping for post-infarction ventricular tachycardia ablation is slow and operator dependent, motivating evaluation of whether a spatio-temporal graph convolutional network detects these potentials better than established alternatives.

Does a spatio-temporal graph convolutional network (ST-GCN) defined on the reconstructed left-ventricular surface detect abnormal ventricular potentials (AVPs) better than established alternatives (such as the conventional bipolar-voltage rule or gradient boosting on expert descriptors) in patients with post-infarction ventricular tachycardia?

Key Points

To determine whether a spatio-temporal graph convolutional network defined on 3-D endocardial meshes improves detection of abnormal ventricular potentials compared to conventional voltage rules and machine learning baselines.
Evaluated the ARGO dataset comprising 1962 annotated bipolar electrograms from 9 post-infarction ventricular tachycardia patients mapped using CARTO 3.
Constructed patient-specific graphs with mapping sites as nodes, geodesic distances along triangulated endocardial meshes as edge weights, and trained a spatio-temporal graph convolutional network.
Benchmarked the network against six comparators, including a clinical bipolar-voltage threshold (<1.5 mV) and gradient boosting on 38 expert descriptors, using leave-one-patient-out cross-validation.
Gradient boosting achieved the highest area under the ROC curve at 0.973 +/- 0.033, followed by the conventional bipolar-voltage rule at 0.952 +/- 0.053 and the spatio-temporal graph network at 0.942 +/- 0.052.
No pairwise performance comparison significantly favoured the spatio-temporal graph convolutional network (Holm-adjusted p >= 0.06).
Voltage stratification demonstrated that all electrograms below 0.5 mV were abnormal potentials, while 95.6% above 1.5 mV were physiological, showing that expert labels are largely driven by signal amplitude.

Study Design

Type

Observational (n=9)

Multicenter

No

Structured PICO

Does a spatio-temporal graph convolutional network (ST-GCN) defined on the reconstructed left-ventricular surface detect abnormal ventricular potentials (AVPs) better than established alternatives (such as the conventional bipolar-voltage rule or gradient boosting on expert descriptors) in patients with post-infarction ventricular tachycardia?

P

Population

9 patients (median age 68 years, 22.2% female) with post-infarction ventricular tachycardia contributing 1,962 endocardial bipolar electrograms mapped with CARTO 3.

E

Exposure

Spatio-temporal graph convolutional network (ST-GCN) defined on the reconstructed 3-D left-ventricular endocardial surface, combining a dilated temporal encoder, Chebyshev graph convolutions, multi-head graph attention, and an auxiliary head regressing the distance to delivered radiofrequency lesions.

C

Comparator

Six comparators trained under identical leave-one-patient-out cross-validation: (1) conventional bipolar-voltage rule (<1.5 mV), (2) gradient boosting (XGBoost) on 38 expert descriptors, (3) 1-D CNN (temporal encoder alone), (4) two-layer bidirectional LSTM (BiLSTM), (5) two-layer Transformer encoder, and (6) ST-GCN without voltage map/substrate attributes.

O

Outcome

Automated detection of abnormal ventricular potentials (AVPs) vs non-AVPs (annotated by consensus of three independent expert electrophysiologists), evaluated by Area Under the Receiver Operating Characteristic curve (AUC-ROC) and Precision-Recall AUC (PR-AUC) under leave-one-patient-out cross-validation.surrogate

Main Result

Absolute Event Rate: 0.942% vs 0.952%

p-value: p=1.000

Complex spatio-temporal graph neural networks do not outperform simple clinical voltage rules or gradient boosting on expert features for identifying post-infarction VT ablation targets in the ARGO benchmark, largely because expert annotations in this dataset are heavily driven by bipolar signal amplitude.

Limitations

Small sample size of nine patients from a single centre and a single mapping platform limits statistical power and generalisability
Reference standard is expert visual consensus (largely driven by bipolar amplitude) rather than clinical outcome
Delivered radiofrequency lesions reflect operator strategy rather than an absolute spatial ground truth
Only endocardial left-ventricular data are included, with no representation of epicardial or intramural abnormal ventricular potentials
Small sample size of 9 patients constrains statistical power, where the smallest attainable one-sided Wilcoxon p-value is 0.00195.
In the ARGO dataset, expert AVP annotation is nearly a deterministic function of bipolar signal amplitude (100% AVP below 0.5 mV, 95.6% physiological above 1.5 mV), which limits how well the dataset can separate competing modelling strategies.
Mapping points sit on mesh vertices with short neighbour distances, making geodesic and Euclidean distances nearly interchangeable (median ratio 1.005) and preventing surface curvature from meaningfully contributing.
Moderate inter-annotator agreement among human experts (Fleiss' kappa 0.641 for binary AVP task) sets a ceiling on automated model evaluation.

Implication

Graph networks add no value over voltage or XGBoost for AVP detection; hypothesis-generating and should not yet change mapping practice.

Abstract

Abnormal ventricular potentials (AVPs) recorded during substrate mapping guide catheter ablation in post-infarction ventricular tachycardia, and annotating them by hand is slow and operator dependent. We asked whether a spatio-temporal graph convolutional network (ST-GCN) defined on the reconstructed left-ventricular surface detects AVPs better than established alternatives. Using the ARGO dataset (nine patients, 1962 annotated bipolar electrograms, three independent electrophysiologists plus a consensus), we built one graph per patient with nodes at mapping sites and edges weighted by geodesic distance along the triangulated endocardial mesh. The network combined a dilated temporal encoder, Chebyshev graph convolutions, multi-head graph attention, and an auxiliary head regressing the distance to delivered radiofrequency lesions. Six comparators were trained under identical leave-one-patient-out cross-validation, among them the conventional bipolar-voltage rule (below 1.5 mV) and gradient boosting on 38 expert descriptors. Decision thresholds were estimated on pooled inner-validation patients only. Gradient boosting reached the highest AUC (0.973 +/- 0.033), ahead of the voltage rule (0.952 +/- 0.053) and the ST-GCN (0.942 +/- 0.052); no pairwise comparison favoured the ST-GCN (Holm-adjusted p > = 0.06). A 13-variant ablation showed that removing spatial message passing (0.949) or substituting Euclidean for geodesic edge weights (0.954) left performance unchanged, whereas removing the wideband amplitude channel lowered it to 0.921. Stratifying by voltage explains the pattern: below 0.5 mV every annotated electrogram was an AVP, above 1.5 mV 95.6% were physiological. In ARGO, expert AVP annotation is largely determined by bipolar amplitude, which limits how well the dataset can separate competing modelling strategies.

Discussion

No takes yet. Share an insight, caveat, or question.

Cite This Study

Amine BOUGARNE (2026) conducted an observational in Post-infarction ventricular tachycardia (n=9). Spatio-temporal graph convolutional network (ST-GCN) vs. Conventional bipolar-voltage rule (<1.5 mV) and gradient boosting (XGBoost) on 38 expert descriptors was evaluated on Detection of abnormal ventricular potentials (AVPs) measured by leave-one-patient-out cross-validation AUC-ROC (95% CI 0.888-0.977, p=1.000). Spatio-temporal graph convolutional networks achieved an AUC of 0.942 for detecting abnormal ventricular potentials, failing to outperform the <1.5 mV voltage rule (0.952) or XGBoost (0.973; p>=0.06).

synapsesocial.com/papers/6ae28f53cbadf70daf6c44fa https://doi.org/10.21203/rs.3.rs-11035410/v1

View Full Paper

Ask AI

Bookmark

Share

Automated identification of post-infarction ventricular tachycardia ablation targets with spatio-temporal graph convolutional networks on 3-D electroanatomical manifolds

Key Points

Study Design

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

Automated identification of post-infarction ventricular tachycardia ablation targets with spatio-temporal graph convolutional networks on 3-D electroanatomical manifolds

Key Points

Study Design

Structured PICO

Main Result

Limitations

Implication

Abstract

Discussion

Cite This Study

Keep reading

Keep reading