Machine-learning model for predicting left atrial thrombus in patients with paroxysmal atrial fibrillation

Figures

• 

  Fig. 1. Flowchart of participants selection. AF, atrial fibrillation; LAT, left atrial thrombus; TEE, transesophageal echocardiography

• 

  Fig. 2. Receiver operating characteristic curves for comparative analysis of machine learning algorithms and CHA2DS2-VASc scores in LAT. AUC, area under the curve; CI, confidence interval; LAT, left atrial thrombus

• 

  Fig. 3. Variable importance and performance analysis of LAT predictive modeling using random forest algorithm. AF, atrial fibrillation; BMI, body mass index; BNP, brain natriuretic peptide; BP, blood pressure; CrCl, creatinine clearance; DCM, dilated cardiomyopathy; HCM, hypertrophic cardiomyopathy; INR, international standardized ratio; LAD, left atrial diameter; LVEF, left ventricular ejection fraction; LVEDD, left ventricular end-diastolic diameter; NT-proBNP, N-terminal pro-B-type natriuretic peptide; TIA, transient ischemic attack. ROC, receiver operating characteristic scurves; AUC, area under the curve; CI, confidence interval

• 

  Fig. 4. Verification of model performance using testing cohort and post-calibration analysis. Model performance was evaluated using testing cohort data before (left) and after calibration (right). Blue bars represent predicted values and orange bars indicate actual observed values. Predictive impact was evaluated by stratifying detection rates into quintile. There was a statistically significant disparity between predicted and actual values before calibration (p < 0.001), and the disparity was no longer statistically significanct after calibration (p = 0.284)

• 

  Fig. 5. Predicted distribution of LAT detection rates in the final model