Background Atrial fibrillation and flutter (AF/AFl) increase the risk of thromboembolic events by promoting clot formation in the left atrium (LA), which can be visualised using transoesophageal echocardiography (TEE). Current guidelines recommend initiation of oral anticoagulation (OAC) in patients with AF/AFl based solely on CHA2DS2VASc score (congestive heart failure, hypertension, age ≥75 years, diabetes mellitus, history of stroke or thromboembolism, vascular disease, age 65 to 74 years, female sex). Yet, some patients develop left LA thrombus (LAT) and experience thromboembolic events despite OAC. Purpose We sought to develop and externally validate a machine learning model for prediction of presence of LAT based on clinical, laboratory and transthoracic echocardiography (TTE) features. Methods We analyzed data from the multicenter, prospective LATTEE registry (Left Atrial Thrombus on Transesophageal Echocardiography) that included patients from 13 sites who underwent TEE before cardioversion or ablation between November 2018 and March 2021. We used XgBoost model to predict presence of LAT in TEE based on 29 clinical features, 10 biomarkers and 5 TTE measurements. We trained and tested the model internally using 10-fold hold-out cross validation and data from 12 sites (N=2489). We then tested the final model externally using data from the 13th site (that had recruited most patients, N=400). We compared the predictive performance with that of CHA2DS2VASc score using areas under receiver operating curve (AUC) and DeLong test. Results In the training and internal testing cohort the median age was 67 (Inter Quartile Range [IQR] 59, 74), 63% were male, 85% received OAC and LAT was found in 8.4%. Ablation was the indication for TEE in 43%, cardioversion in 57%. In internal, 10-fold hold-out cross validation, the model achieved AUC of 0.755 (95% confidence interval [CI]: 0.722, 0.788) while CHA2DS2VASc performed significantly worse with AUC of 0.638 (95% CI: 0.604, 0.673), P<0.0001 (Figure 1). Left ventricular ejection fraction, rhythm at the time of study (AF/AFl or sinus rhythm) and age received the highest feature importance ranking (Figure 2). In the external testing cohort the median age was 67 (IQR 59, 74), 66% were male, 88% received OAC, ablation was the indication for TEE in 49% of cases and LAT was found in 6.8%. In this external cohort, our model achieved AUC of 0.815 (95% CI: 0.741, 0.889) while CHA2DS2VASc performed significantly worse with AUC of 0.684 (95% CI: 0.583, 0.785), P=0.028. Conclusion Machine learning based on readily available clinical data allows accurate prediction of the presence of LAT in patients with AF/AFl irrespective of OAC treatment. Such score could be used to identify patients who should undergo TEE before ablation or cardioversion. Subsequent studies to clinically evaluate such application of our model as well as how the model can predict future thromboembolic events are warranted. Funding Acknowledgement Type of funding sources: None.