Predicting intracardiac thrombus formations despite continuous oral anticoagulation in atrial fibrillation patients undergoing catheter ablation procedures: a machine learning approach

Atrial fibrillation (AF) increases stroke risk due to thrombus formations (TF) in the left atrium (LA), especially with advanced atrial disease. Limited data exist on TF occurrence despite continuous oral anticoagulation (OAC), posing safety concerns for AF patients undergoing left atrial catheter ablation (CA). Identifying high-risk pts could optimize preprocedural imaging decisions, especially for transesophageal echocardiography (TEE).

This retrospective study aimed to develop a machine learning model to predict TF in AF pts under continuous OAC. By analyzing clinical, ECG, and echo characteristics, we sought to differentiate patients with intracardiac thrombi or prethrombotic formations/sludge from those without TF.

Among 1,095 pts who underwent a TEE (06/2019-08/2024), 80pts with TF (group 1) pre or post CA (PVI, PVI+, first-time and repeat ablations) at a tertiary hospital were identified. The control group (group 0) without TF was chosen by propensity score matching on age and sex.

Baseline characteristics, echo parameters and sinus rhythm ECGs (if available) were extracted from medical records, with a specific focus on the P-wave due to its link with electrical remodeling and Atrial Cardiomyopathy.

An XGBoost algorithm was employed to develop a predictive model for TF, using 53 clinical features. The data was divided into a 70/30 split for training and testing and a 5-fold cross-validation (CV) was applied to the training set.

All results were obtained using RStudio Version 2024.04.2+764 with the packages XGBoost, TableOne and MatchIt.

Table 1 presents the pts characteristics of both groups. In group 1 we found 31 pts with CHADS2VASc-Score ≤ 3 (38.8%). All pts in group 1 were on OAC (7,5% Vit. K Antagonist/ 92,5 % NOAC) when TF occurred. SR ECG’s revealed a significantly higher P-wave area and P-wave dispersion in group 1 (Table 1).

The model achieved an Area Under the Curve (AUC) of 0.95 in the testing dataset (Fig. 1) and a mean AUC of 0.86 during CV in the training phase, indicating a high level of accuracy in predicting TF. By selecting a cut-off value of 0.6, the model demonstrated a sensitivity of 0.96 and a specificity of 0.75 (positive predictive value (PPV)=0.79; negative predictive value (NPV)=0.95). The importance metric identified P-wave dispersion as the most influential factor in the model. Even after retraining the model without this parameter, it still achieved an AUC of 0.91 (PPV=0.74;NPV=0.94;Sensitivity=0.96;Specificity=0.67) in the testing data and a mean AUC of 0.85 in CV.

Machine learning algorithms offer a promising tool for predicting TF in AF pts with continuous OAC undergoing CA with high predictive accuracy using existing medical records. Further research could lead to a prediction tool which is easily applicable in the clinical setting. This approach may supplement CHA₂DS₂-VASc scoring enhancing risk assessment.