Abstract
OBJECTIVE: To establish a machine-learning model of left atrial appendage thrombogenic milieu (LAATM) in patients with nonvalvular atrial fibrillation (NVAF), and analyze the corresponding risk factors to guide clinical decision-making. METHODS: Patients with NVAF were selected and divided into LAATM group and non-LAATM group, according to the results of transesophageal echocardiography (TEE). The LAATM group included LAA thrombus formation, sludge and spontaneous echo contrast. The patient data was collected and preprocessed. The machine learning algorithms of random forest (RF), support vector machine (SVM) and extreme gradient boosting (XGBoost) were used to establish a predictive model for LAATM in patients with NVAF. Shapley additive explanation (SHAP) was used to sort the feature importance of clinical factors. RESULTS: A total of 1217 patients were selected in this study, including 112 patients in LAATM group and 1105 patients in non-LAATM group. In terms of predictive performance, AUC value of RF model was 0.97, F1 score was 0.93, accuracy was 0.98, precision was 0.99, and recall was 0.89; AUC value of SVM model is 0.96, F1 score is 0.89, accuracy is 0.97, precision is 0.95, recall is 0.84; The AUC value of the XGBoost model is 0.96, F1 score is 0.88, accuracy is 0.97, precision is 0.98, and recall is 0.82. The prediction efficiency of RF model is the best. The prediction results of RF model were visualized by SHAP diagram, indicating that Homocysteine (HCY), NT-proBNP, C-reactive protein (CRP), glycosylated hemoglobin (HbA1c) and ABC stroke score were the top five risk factors affecting the formation of LAATM in patients with NVAF. CONCLUSION: The RF model achieved the best predictive performance between the three prediction model. HCY, NT-proBNP, CRP, HbA1c and ABC stroke score were the top five risk factors affecting the formation of LAATM in patients with NVAF.