The left atrial (LA) morphological profile, anatomically contiguous with the left atrial appendage (LAA), exhibits hemodynamic properties associated with thrombogenic predisposition in nonvalvular atrial fibrillation (NVAF). Integrating these structural biomarkers with clinical parameters enables noninvasive prediction of thrombosis risk.

This single-center retrospective study analyzed 253 NVAF patients undergoing pre-ablation dual-phase delayed LA computed tomography angiography (CTA). A machine learning (ML) model incorporating clinical and radiomics features was developed to predict LAA thrombosis/blood stasis. Multi-framework interpretation revealed robust predictive performance: global accuracy 92%, thrombosis subgroup F1-score of 0.97 (95%CI: 0.89–1.00) with area under the curve of 1.00 (AUC: 95%CI: 0.99–1.00), blood stasis subgroup F1-score of 0.90 (95%CI: 0.81–0.97) with AUC of 0.97. Model reliability was confirmed by Cohen's κ = 0.88 and 5-fold cross-validation (CV) score (mean score 0.91, range 0.88–0.94). Contribution visualization analysis identified clinical parameters as the main predictors, lipid-related indicators showed high discriminative value, while the radiomics parameters LA sphericity and radiomics texture features provided incremental calibration.

The multimodal model integrating clinical profiles with CTA-derived radiomics effectively stratifies LAA thrombosis and blood stasis risks, demonstrating an exceptional discriminatory accuracy for thrombus detection.

‡ These authors share first authorship.

Received: 22 May 2025

Accepted after revision: 06 March 2026

Accepted Manuscript online:
19 March 2026

Article published online:
31 March 2026

© 2026. Thieme. All rights reserved.

Georg Thieme Verlag KG
Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany