Automated quantification of left atrial function using AI: a novel predictor of thrombus formation in AF patients undergoing cardioversion or catheter ablation

Abstract Background Transthoracic echocardiography (TTE) is routinely used for cardiac imaging in patients undergoing catheter ablation or cardioversion for atrial fibrillation (AF). While left ventricular ejection fraction (LV EF) is an established parameter for assessing cardiac function and has been associated with left atrial appendage thrombus (LAT), less is known about the significance of left atrial ejection fraction (LA EF) as a predictor of LAT. Objectives This study aimed to evaluate the utility of LA EF measured by an AI model using TTE views in assessing atrial function and its association with the risk of LAT formation, in comparison to LV EF, in patients undergoing catheter ablation or cardioversion. Methods We analysed a retrospective cohort of consecutive patients undergoing transoesophageal echocardiography (TEE) to rule out LAT and TTE before catheter ablation or cardioversion between July 2020 and December 2022. A dedicated workstation was used to manually measure LV EF and LA EF. Additionally, a state-of-the-art machine learning model was used to automatically measure LA EF (AI-derived) and left atrial volume index (LAVI, AI-derived) using the bi-plane Simpson method. All LA EF and LAVI measurements were fully automated, without any user input. Receiver operating characteristic (ROC) curve analysis and multivariable logistic regression were used to assess and compare the predictive value of LA EF and LV EF for LAT. Results Among the 157 patients (median age 66, 64% male), 13 (8.2%) were found to have LAT. ROC curve analysis demonstrated that both AI-derived and manually measured LA EF were comparable to LV EF in predicting LAT, with AUCs of 0.90 (95% CI: 0.84–0.97), 0.92 (95% CI: 0.85–0.98), and 0.84 (95% CI: 0.71–0.97), respectively. In multivariable logistic regression analysis, after adjusting for age, sex, and LAVI (AI-derived), both AI-derived LA EF and manually measured LV EF were found to be significant, independent predictors of LAT presence, with OR = 0.82 (95% CI: 0.69–0.94) and OR = 0.92 (95% CI: 0.86–0.97), respectively. Age also showed a significant association with LAT (p = 0.02). No LATs were observed in patients with LA EF values above 25% (AI-derived) and 21% (manually measured), whereas the highest LV EF value in a patient with LAT was 62%, suggesting that normal LA EF values may significantly reduce thrombus risk. Conclusion LA EF measured by AI using TTE appears to be a promising tool for assessing atrial function and stratifying LAT risk in patients with AF undergoing ablation or cardioversion. The AI model showed good agreement with values manually defined by expert echocardiographers and required no user input, thereby minimizing the risk of user error. While manual measurement of LA EF is time-consuming, the AI-derived method is fast, capable of averaging multiple beats and does not require ECG gating. Further refinement may enhance its clinical utility as a rapid, reproducible tool.Forrest plot multivariable regression Violin plot for clot presence