Development and Validation of a Coronary Computed Tomography Angiography-Based Radiomics-Integrated Model for Noninvasive Detection of Left Atrial Appendage Thrombus in Atrial Fibrillation.

内容摘要

Timely detection of left atrial appendage thrombus (LAAT) is critical for stroke prevention in atrial fibrillation. Current diagnostic approaches such as traditional models and physician visual analysis face challenges in comprehensive capturing image information. The study aimed to develop and validate a multimodal model integrating coronary computed tomography angiography-based radiomics features with clinical parameters for noninvasive LAAT detection in patients with atrial fibrillation. The diagnostic study retrospectively enrolled 670 patients with nonvalvular atrial fibrillation undergoing coronary computed tomography angiography and transesophageal echocardiography from May 2015 to May 2023 and stratified into training and internal validation sets. An independent prospective cohort (n=114) from May 2023 to May 2025 served for external validation. Semiautomated LAA segmentation extracted 1231 radiomics features, with 25 features selected by random forest. A multimodal LAAT detection model was developed and evaluated using receiver operating characteristic, calibration curve, and decision curve analysis and compared against traditional models and physician visual analysis. The Radiomics-LAAT model achieved significantly superior discrimination in both internal validation (area under the curve, 0.963 [95% CI, 0.945-0.980], accuracy: 0.929) and external validation (areas under the curve, 0.920 [95% CI, 0.886-0.953], accuracy: 0.807), outperforming traditional models and physician visual analysis, with optimal calibration (Brier score: 0.067) and clinical net benefit. The Radiomics-LAAT model achieved high sensitivity (0.953), specificity (0.905), negative predictive value (0.950), and positive predictive value (0.910). The Radiomics-LAAT model significantly enhances noninvasive LAAT detection performance compared with traditional models and physician visual analysis, demonstrating its potential for stroke risk stratification in patients with atrial fibrillation.