Machine learning classifiers for predicting catheter ablation responses using non-contact electrograms during persistent atrial fibrillation

Background:

Identification of atrial sites which are effective ablation targets remains challenging in atrial fibrillation (AF) therapy.

Purpose: We thought to test machine learning algorithm in predicting the responses of ablating electrograms (EGMs) and their effect on terminating persistent atrial fibrillation (persAF).

Methods:

A total of 3,206 non-contact electrograms (EGMs) of 51 ablation lesion sites from ten persAF patients undergoing left atrial (LA) catheter ablation (2048-channel Ensite Array) were used as training dataset. AF cycle length (AFCL) changes before and after ablating each LA site were recorded in Labsystem Pro recording system. The EGMs were labelled in four classes by AFCL changes (with 10 ms threshold) after ablation: AF termination; AFCL increase; AFCL unchanged and AFCL decrease. Dominant frequency (DF) and organisation index (OI) were calculated from all EGMs (264 seconds) and used as input features. A group of machine classifiers were trained using the training dataset to predict the ablation response and deployed. 5-fold cross-validation was considered (80% of the data for training; 20% for validation).

Results:

The accuracy of the investigated classifiers was 69.66 % ± 12.60 %. The best performing Fine k-nearest neighbour (KNN) classifier achieved 85.3% of accuracy in the classification of the four classes. For AF termination classification (Area under the curve (AUC) = 0.98) from all four classes, a sensitivity of 87% and a specificity of 98% were achieved, whilst classifying AFCL increase group (AUC = 0.96) resulted in a sensitivity of 84% and a specificity of 92%.

Conclusion:

This work presents a machine learning framework to identify EGMs that are responsible for maintenance of persAF and potential targets for catheter ablation using panoramic non-contact mapping. Supervised learning algorithms on frequency features of long EGMs showed the ability to predict ablation responses measured by AFCL changes and AF termination. Targeting atrial regions with appropriate frequency characteristics might improve ablation outcome in persAF.