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

Figures

• 

  Figure 1. Flow chart of patient selection and allocation. LAA indicates left atrial appendage; NVAF, nonvalvular atrial fibrillation; and SMOTE, synthetic minority oversampling technique.

• 

  Figure 2. The process of radiomics analysis. The radiomics analysis pipeline comprised 6 sequential steps: (A) image preprocessing and segmentation: Digital Imaging and Communications in Medicine‐formatted CCTA images were isotropically resampled to ensure uniform voxel dimensions. The LAA was segmented as the ROI using 3D‐Slicer, combining automated thresholding with manual refinement; (B) feature extraction: a total of 1231 radiomics features were extracted from each ROI, encompassing first‐order features, shape features, and textural features derived from GLCM, GLRLM, GLSZM, NGTDM, and GLDM, in addition to wavelet features; (C, D) feature selection using random forest: all extracted features were normalized (Z score) before selection. A random forest algorithm was employed for feature selection, using an ensemble of 800 decision trees in pretraining to stabilize error estimation, followed by training with 300 trees to identify the most predictive feature subset; (E) feature importance ranking: the importance of each selected feature was quantified based on the Mean Decrease Gini index, which represents the average reduction in node impurity contributed by a feature across all trees in…

• 

  Figure 3. ROC curves and efficiency parameters for the LAAT detection models in the internal validation set. A, B, ROC curves and overall AUC of LAAT detection models; C, D, Radar chart and line chart of efficiency parameters. Model 1 (clinical parameters): CHA2DS2‐VASc score; Model 2 (clinical enhanced): Model 1 + atrial fibrillation type + D‐dimer; Model 3 (imaging integrated): Model 2 + LAA‐diameter + LAA‐length + anteroposterior diameter of the left atrium + LAAmin/AAmin; Model 4 (Radiomics‐LAAT): Model 3 + Radiomics score. AUC indicates area under the curve; LAA, left atrial appendage; LAAT, left atrial appendage thrombus; and ROC, receiver operating characteristic. *Difference is considered statistically significant.

• 

  Figure 4. Calibration curves and decision curve analysis for LAAT detection models in the training and internal validation sets. A, Calibration curves of training set; B, Calibration curves of internal validation set; C, Decision curve analysis of training set; D, Decision curve analysis of internal validation set. Model 1 (clinical parameters): CHA2DS2‐VASc score; Model 2 (clinical enhanced): Model 1 + atrial fibrillation type + D‐dimer; Model 3 (imaging integrated): Model 2 + LAA‐diameter + LAA‐length + anteroposterior diameter of the left atrium+ LAAmin/AAmin; Model 4 (Radiomics‐LAAT): Model 3 + Radiomics score. Predicted probability is the anticipated likelihood of an event occurring, observed probability is the observed likelihood, bias‐corrected line depicts the model's performance following multiple independent samplings, and apparent line depicts the actual performance of the model.LAA indicates left atrial appendage; and LAAT, left atrial appendage thrombus.

• 

  Figure 5. External validation for LAAT detection models. A, Receiver operating characteristic curve of LAAT detection models; B, Calibration curves of LAAT detection models. Model 1 (clinical parameters): CHA2DS2‐VASc score; Model 2 (clinical enhanced): Model 1 + atrial fibrillation type + D‐dimer; Model 3 (imaging integrated): Model 2 + LAA‐diameter + LAA‐length + anteroposterior diameter of the left atrium + LAAmin/AAmin; Model 4 (Radiomics‐LAAT): Model 3 + Radiomics score. LAA indicates left atrial appendage; LAAT, left atrial appendage thrombus; NPV, negative predictive value; and PPV, positive predictive value.