What question did this study set out to answer?

This research aims to determine if an ECG-based risk model can better predict thromboembolic and major adverse cardiac events in atrial fibrillation than traditional methods.

February 8, 2026

ECG-based artificial enhancement for risk stratification of thromboembolic and major adverse cardiac events in atrial fibrillation compared with CHA2DS2-VA scoring

Q: What does this research mean for the field?

An AI-driven ECG system provides superior predictive performance for thromboembolic and major adverse cardiac events in atrial fibrillation patients compared to the CHA2DS2-VA scoring system, achieving an AUROC of 0.923. Novelty: ClaimNovelty.NOVEL_FINDING. Consensus alignment: ConsensusAlignment.CHALLENGES_CONSENSUS.

Key Result

AI-driven ECG risk stratification achieved AUROC 0.923, outperforming CHA2DS2-VA score in predicting thromboembolic and major adverse cardiac events in AF patients.

Key Points

This research aims to determine if an ECG-based risk model can better predict thromboembolic and major adverse cardiac events in atrial fibrillation than traditional methods.
Analyzed data from 10,181 atrial fibrillation patients
Developed an AI-driven ECG system using transformer deep neural networks
Processed 60,413 ECGs alongside clinical records and CHA2DS2-VA scores
Validated clinical analysis with a team of seven cardiology experts
The AI-driven ECG system showed significant differentiation between low and high thromboembolic risks (35.22 vs. 80.30, p < 0.001)
Achieved a high area under the receiver operating characteristic curve (AUROC) of 0.923
Notable performance variations across CHA2DS2-VA scores, with significant precision differences
Kaplan-Meier curves indicated marked survival differences between risk groups in terms of major adverse cardiac events.

Structured PICO

Does an AI-driven ECG system improve predictive performance for thromboembolic and major adverse cardiac events in atrial fibrillation patients compared to the CHA2DS2-VA score?

Population

10,181 patients with atrial fibrillation (AF) with 60,413 ECGs recorded between 2006 and 2021 at a university hospital

Intervention

AI-driven ECG system employing transformer deep neural networks for risk stratification

Comparator

Traditional CHA2DS2-VA scoring system

Outcome

Predictive performance for thromboembolic events and major adverse cardiac events (MACE)composite

An AI-driven ECG system demonstrated high predictive accuracy (AUROC 0.923) for thromboembolic events in atrial fibrillation, offering a potentially superior alternative to the traditional CHA2DS2-VA score.

Main Result

Absolute Event Rate: 0% vs 0%

Abstract

Abstract Background The CHA2DS2-VA score is commonly utilized to evaluate thromboembolic risk in atrial fibrillation (AF) patients, guiding anticoagulation therapy decisions. However, its predictive accuracy has limitations, necessitating more refined risk assessment methods. Recent advancements in computational techniques allow electrocardiogram (ECG)-based modeling, which may enhance the precision of predicting thromboembolic events and major adverse cardiac events (MACE). This study examines an artificially enhanced ECG-based method and compares its effectiveness to the traditional CHA2DS2-VA scoring system. Objective To assess whether an artificially enhanced ECG-based risk stratification model provides superior predictive performance for thromboembolic and major adverse cardiac events in atrial fibrillation patients compared to the CHA2DS2-VA scoring system. Methods Using data from 10,181 AF patients, we developed and validated an AI-driven ECG system employing transformer deep neural networks. This system processed 60,413 ECGs from a university hospital, recorded between 2006 and 2021, alongside clinical records including CHA2DS2-VA scores and longitudinal follow-up data. Clinical analysis was meticulously validated by a team of seven experts, including cardiologists, to ensure robustness beyond traditional data warehouses. Results The AI-driven ECG system effectively differentiated between low and high thromboembolic risks, demonstrating a significant difference in AI scores (35.22 ± 10.17 vs. 80.30 ± 8.77, p 0.001) and achieving high predictive accuracy with an AUROC of 0.923 (95% CI 0.918 – 0.928). Performance metrics varied significantly across CHA2DS2-VA scores, with notable precision and recall differences between scores 0 and 2-8. Kaplan-Meier curves illustrated a marked survival difference between the risk groups in terms of MACE and related hospitalizations. Conclusion The application of computationally enhanced ECG analysis provides a more precise risk stratification for thromboembolic and major adverse cardiac events in atrial fibrillation patients. This method offers a promising alternative to conventional scoring systems, facilitating more informed clinical decision-making regarding anticoagulation therapy.

Bookmark