Machine learning model for predicting left atrial thrombus or spontaneous echo contrast in non-valvular atrial fibrillation patients based on multimodal echocardiographic parameters

BACKGROUNDAlthough clinical prediction models have been proposed to predict thrombosis risk in non-valvular atrial fibrillation (NVAF), machine learning (ML) based models to predict thrombotic risk were limited. This study aimed to develop a robust ML-based predictive model that integrates multimodal echocardiographic data and clinical risk factors to evaluate the risk of thrombosis in patients with NVAF. METHODS AND RESULTSA total of 402 NVAF patients scheduled for AF radiofrequency ablation and/or left atrial appendage closure at the First Affiliated Hospital of Guangxi Medical University from January 2020 to December 2023 were prospectively collected. Among them, there were 289 males (71.9%) and 113 females (28.1%), with a mean age of 59.7 years. There were 142 patients (35.3%) with left atrial thrombus/spontaneous echocardiographic contrast (LAT/SEC) and 260 patients (64.7%) without LAT/SEC. Clinical data, biochemical markers, and multimodal echocardiographic parameters were collected to construct the model. After screening the influencing factors with Least Absolute Shrinkage and Selection Operator (LASSO) regression, we explored seven ML models - Logistic Regression (LR), Decision Tree (DT), K-Nearest Neighbors (KNN), Light Gradient Boosting Machine (LightGBM), Support Vector Machine (SVM), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost), to forecast the risk of thrombosis in NVAF patients. A variety of metrics, such as accuracy, precision, recall, F1 score and area under the curve (AUC) were used to evaluate the performance of the models. The DeLong test was applied to compare area under receiver operating characteristic (AUROC) curves across different ML models. Decision curve analysis (DCA) was used to gauge the clinical utility of each ML model by comparing their clinical net benefit. To gain insight into the risk prediction model, we used Shapley additive explanations (SHAP) analysis and investigated the contributions of the different variables. The incorporation of multimodal echocardiographic parameters and clinical risk factors using advanced ML algorithms markedly enhanced the accuracy of predicting thrombosis risk in individuals with NVAF. Specifically, the XGBoost model (AUC 0.959, 95% CI 0.925-0.993) slightly outperformed the traditional LR model (AUC 0.949, 95% CI: 0.911-0.987) in predicting thrombosis risk in NVAF patients, and showed superior predictive ability compared to other ML algorithms. Additionally, XGBoost offered greater clinical net benefit within a threshold probability range of 0.1 to 1.0. SHAP analysis revealed that left atrial structure (left atrial volume index, three-dimensional sphericity index), hemodynamic parameters (left atrial acceleration factor and S/D ratio), and functional parameters (peak atrial longitudinal strain and left ventricular ejection fraction) were important features in predicting the risk of thrombus formation in NVAF patients, with reduced peak atrial longitudinal strain being the most important risk factor for predicting thrombus. CONCLUSIONSDeveloping a predictive model utilizing ML techniques that incorporate multimodal echocardiographic parameters in conjunction with clinical risk factors has the potential to enhance the predictive accuracy of the thrombosis risk in individuals with NVAF. The XGBoost model shows that decreased PALS, hemodynamic abnormalities and left atrium spherical remodeling are significant factors correlated with increased risk of thrombus in NVAF.