Objective To evaluate the values of CHADS2 score, CHA2DS2-VASc score, and self-modified CHA2DS2-VASc-LA2 score in predicting left atrial thrombus (LAT) in patients with non-valvular atrial fibrillation (NVAF). Methods A total of 203 NVAF patients who underwent transesophageal echocardiography (TEE) examination and cardiac multislice CT scans before AF radiofrequency ablation in our department from June 2007 to June 2012 were included in this study. LAT was detected by TEE in 39 patients (thrombosis group) and not detected in 164 patients (control group). The patients' general condition, medical history, admission examination, CHADS2 score and CHA2DS2-VASc score were subjected to univariate analysis and multivariate logistic regression analysis. Then a new CHA2DS2-VASc-LA2 scoring system was formed by combining left atrial volume index (LAVI), recording LAVI ≥32 mL/m² as 2 points and CHA2DS2-VASc score, which was based on the results of logistic regression analysis. The receiver operating characteristic curve (ROC) was used to compare the values of CHADS2, CHA2DS2-VASc and CHA2DS2-VASc-LA2 scores for predicting LAT formation in NVAF patients. Results ROC curve analysis showed that CHADS2 score had a low predictive value (AUC [area under the curve]=0.661, P=0.002); CHA2DS2-VASc score (AUC=0.731, P<0.001) and CHA2DS2-VASc-LA2 score (AUC=0.771, P<0.001) had middle predictive values. The CHA2DS2-VASc-LA2 scores of the three patients with CHADS2 score being 0 in thrombosis group was increased, and there was no patient in the thrombosis group with CHA2DS2-VASc-LA2 score being 0. Conclusion Compared with CHADS2 and CHA2DS2-VASc systems, CHA2DS2-VASc-LA2 score has a better performance in predicting LAT in low-risk patients with NVAF. It is also suggested that pre-ablation TEE may be unnecessary if the CHA2DS2-VASc-LA2 score is 0.