Intracardiac thrombus formations despite continuous oral anticoagulation in atrial fibrillation patients undergoing catheter ablation procedures: pilot development of a machine learning prediction model
Background:
Atrial fibrillation (AF) increases the risk of stroke, primarily due to thrombus formations (TF) in the left atrium (LA), especially in advanced atrial disease. Undetected TF despite continuous oral anticoagulation remain a safety concern, particularly in patients undergoing AF catheter ablations (CA) procedures. This retrospective study aimed to develop a machine learning (ML) model to predict TF in those patients.
Methods:
A retrospective cohort of 1335 AF patients who underwent CA and transesophageal echocardiography were screened for TF. Initially, n = 93 TF+ patients were identified and propensity-score-matched (age, sex) to 93 TF− controls. Clinical, echocardiographic data and p-wave morphology in sinus rhythm (SR) were assessed. After excluding patients without SR ECGs 65 TF+ and 84 TF− patients were used for the ML models. External validation was performed.
Results:
We analyzed 149 patients undergoing LA CA procedures on DOACs 96.0%/VKA 4.0%. The TF+ patients (20% solid thrombi, 80% prethrombotic formations) had a mean CHA₂DS₂-VASc of 3.89 and were in clinical arrhythmias in 96.9% at the time of TF, seen in a TOE performed before planned CA. The ML models showed a good performance in recognizing TF in the test-dataset, achieving a ROC AUC of 0.886 and 0.826 which outperformed CHA₂DS₂-VASc in our cohort (AUC 0.641). Some P-wave-characteristics in SR are ranked among the top ML features.
Conclusion:
ML algorithms, including echo parameters and p-Wave parameters reflecting atrial substrate, offer a promising approach for accurate prediction of TF in anticoagulated AF patients undergoing CA. Further validation could enable implementation of a clinically applicable risk stratification tool to enhance preprocedural decision-making.
