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

Abstract

Objective: Left atrial thrombus (LAT) poses a significant risk for stroke and other thromboembolic complication in patients with atrial fibrillation (AF). This study aimed to evaluate the incidence and predictors of LAT in patients with paroxysmal AF, utilizing machine learning techniques based on data from the Chinese Atrial Fibrillation study.

Methods: A large-scale multi-center retrospective study was conducted involving patients diagnosed with non-valvular paroxysmal AF. LAT incidence was assessed, and potential risk factors were analyzed. Machine learning algorithms, including decision tree, random forest, AdaBoost, k-Nearest Neighbor, and logistic regression, were employed to develop a predictive model for LAT.

Results: Of the 49,515 patients with paroxysmal AF, 1,058 patients (2.1%, 95% CI 2.0%-2.3%) were identified with LAT. Sixty-one variables were initially included to train machine learning models, with the random forest algorithm demonstrating the best predictive performance (AUC 0.833, 95%CI 0.730-0.924). The final model, refined to include nine essential features, achieved an AUC of 0.787 (95%CI 0.670-0.883). Calibration analysis indicated no significant difference between predicted and observed values (p = 0.181). The median predicted probabilities of LAT across quintiles were 2.3%, 7.0%, 11.8%, 16.6%, and 21.5%.

Conclusion: This simplified prediction model effectively identifies the risk of LAT in patients with paroxysmal AF, providing a valuable tool for clinical decision-making. Further studies are needed to explore AF management and risk stratification in other AF subtypes.