Machine learning for enhancing the ablation outcomes using recurrence quantification analysis in human persistent atrial fibrillation.

Background: Several ablation strategies, including dominant frequency (DF), complex fractionated atrial electrograms (CFAEs), and rotors, have been used to target the drivers of atrial fibrillation (AF). The success rate of these strategies remains suboptimal. Objective: This work aims to develop a model using recurrence quantification analysis (RQA) variables and machine learning (ML) algorithms to predict the responses of ablating intracardiac electrograms (EGMs) and their impact on terminating AF and cycle length changes. Methods: 3206 non-contact EGMs were collected from 10 persistent AF patients. Two classes were considered as labels based on EGM's responses to ablation (positive and negative responses). Nine RQA-based variables were extracted from the EGMs. Ten ML classifiers were trained and tested using leave one patient out a 10-fold cross-validation (LOPOCV)). Results: The decision tree outperformed other ML models, achieving a balanced accuracy of 73.46%, F1_score of 74.05, and AUROC of 0.74. The high performance was achieved using the three most important features (determinism, longest diagonal line, and recurrence rate) with the importance of 30%, 23%, and 13%, respectively. Statistical analysis showed high values of all median RQA variables for the EGM's negative responses over the positive ones. Conclusions: Our results show that RQA variables can effectively highlight the electrophysiological differences between EGM positive and negative responses to catheter ablation. The model might aid cardiologists in predicting ablation outcomes and reducing the number of unsuccessful ablations. A comparison between the proposed approach and previous works shows the superiority of this model.