Key result
Deep learning model accurately maps patient-specific PVI regions, achieving a ~0.93 Dice score.
Why the study?
Pulmonary vein isolation is standard ablation for atrial fibrillation, but recurrence remains a major clinical challenge.
Does a deep learning model accurately generate patient-specific PVI lines from pre-ablation 3D voltage maps in patients with successful AF ablation?
Observational (n=171)
Yes
Does a deep learning model accurately generate patient-specific PVI lines from pre-ablation 3D voltage maps in patients with successful AF ablation?
A deep learning model can accurately reproduce patient-specific PVI patterns associated with successful AF ablation outcomes, offering a potential tool for data-driven precision ablation planning.
May aid personalized PVI planning in AF ablation; leaves open clinical efficacy pending prospective validation.
Pulmonary vein isolation (PVI) is an established standard ablation for atrial fibrillation (AF), however, AF recurrence remains a major clinical challenge. We developed a deep learning model to generate patient-specific PVI lines on pre-ablation 3D voltage maps, using lesion sets from successful AF ablation cases with documented freedom from recurrence for more than one year as ground truth. Using a U-Net-based architecture trained and evaluated on 513 maps from 171 such cases, the model reproduced the anatomical and electrophysiological features of these PVI lesion sets. On the held-out test set, the model achieved a mean Intersection over Union of 0.87 and a Dice score of 0.93. As a proof of concept, these findings suggest that the model can reproduce patient-specific PVI patterns associated with successful outcomes; whether this translates into reduced recurrence requires prospective clinical validation.
No takes yet. Share an insight, caveat, or question.
Sakamoto et al. (2026) conducted an observational in Atrial Fibrillation (n=171). Deep learning model for PVI line generation vs. Ground truth (lesion sets from successful AF ablation cases) was evaluated on Intersection over Union (IoU) for PVI region segmentation (AI Model 1) on the held-out test set (95% CI 0.86-0.88). A deep learning model trained on successful atrial fibrillation ablation cases accurately generated patient-specific pulmonary vein isolation regions, achieving a mean Intersection over Union of 0.87 and a Dice score of 0.93 on a held-out test set.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: