Key result
Transformer-based ECG model predicts high-risk left-sided Kent ERP with AUC 0.77 and maintains strong localization.
Why the study?
Determining whether invasive electrophysiology study is needed for risk stratification in asymptomatic pre-excitation patients remains controversial.
Does a transformer-based deep learning model using 12-lead ECGs accurately localize accessory pathways and predict high-risk ERP in patients with WPW syndrome?
Does a transformer-based deep learning model using 12-lead ECGs accurately localize accessory pathways and predict high-risk ERP in patients with WPW syndrome?
A transformer-based deep learning model applied to standard 12-lead ECGs can accurately localize accessory pathways and predict high-risk conduction properties, potentially serving as a non-invasive tool to support ablation planning.
Answers are grounded in this paper and cite their sources.
May support non-invasive WPW risk stratification; extends localization data but leaves open prospective validation.
Background In the ESC guideline, catheter ablation is recommended in asymptomatic patients with pre-excitation in whom electrophysiology (EP) testing identifies high-risk properties, such as a short antegrade ERP in accessory pathway (AP) ≤250 ms. However, determining whether the invasive EP study is needed in asymptomatic pre-excitation patients remains controversial. Methods This retrospective study included patients with WPW syndrome who underwent EPS and successful ablation from 2005 to 2024. APs were categorized by EPS into right vs. left, septal vs. lateral, and left septal vs. left lateral. ERP was labeled high-risk (≤250 ms) or low-risk (≥270 ms), excluding intermediate ERP (251–269 ms). Two transformer-based models were trained: one for AP localization and one for high-risk ERP classification, using segmented lead-level ECGs with patient-level split (70% train, 10% validation, 20% test). External validation used datasets from three other tertiary hospitals. Results Among 584 patients, left lateral pathways were the most frequent (34.2%), followed by septal pathways accounting for 27%. 110 patients (24.8%) exhibited a Kent ERP ≤250 ms, indicating high-risk conduction properties based on EPS. The transformer-based model achieved strong classification performance: AUCs were 0.97 for right vs. left APs, 0.94 for septal vs. lateral, and 0.87 for left septal vs. left lateral localization. (Figure 1) For high-risk ERP (≤250 ms) prediction, the model achieved an AUC of 0.77 when applied to left-sided pathways only. (Figure 2) Prediction accuracy declined when including right-sided APs or intermediate ERP values, supporting their exclusion from the final training cohort. Model performance remained consistent across external validation datasets from three independent tertiary centers, with AUCs ranging from 0.87 to 0.94 for localization tasks and 0.75 for high-risk ERP prediction. Conclusion This study demonstrates that a transformer-based deep learning model trained on standard 12-lead ECGs can accurately localize accessory pathways and predict high-risk Kent ERP criteria in patients with pre-excitation. This model, developed with EPS-validated data and externally validated across multiple centers, may serve as a reliable non-invasive tool to support ablation planning and early risk stratification in clinical practice.
No takes yet. Share an insight, caveat, or question.
Choi et al. (2026) studied this question. A transformer-based ECG model accurately predicted high-risk Kent ERP (≤250 ms) with an AUC of 0.77 for left-sided pathways and maintained strong localization performance.
Synapse has enriched 10 related papers. Consider them for comparative context: