Identification of Ablation Sites in Persistent Atrial Fibrillation Based on Spatiotemporal Dispersion of Electrograms Using Machine Learning.

Amina Ghrissi1, Fabien Squara2, Vicente Zarzoso3, Johan Montagnat1
1Université Côte d’Azur, CNRS, I3S Laboratory, 2Université Côte d’Azur, CHU Pasteur, Cardiology Department, 3Université Côte d'Azur


Background: A recent patient-tailored ablation protocol to treat atrial fibrillation (AF) consists in identifying ablation sites based on their spatiotemporal dispersion (STD). STD stands for a delay of the cardiac activation observed in intracardiac electrograms (EGMs) through contiguous leads. Interventional cardiologists use PentaRay multipolar mapping catheters to identify STD EGMs visually. This work aims at automatically identifying ablation sites by classifying EGM data acquired by PentaRay into ‘ablated’ vs. ‘non-ablated’ groups using machine learning (ML) techniques.

Materials & Methods: 35563 multichannel recordings are acquired from 15 persistent AF patients at the Cardiology Department of Nice University Hospital. 10 bipolar EGMs composed of 2500 samples each (1000 Hz sampling rate) are recorded per atrial location with the PentaRay. A mapped site is labeled as 'ablated' if its Euclidean distance to an ablated site is inferior to a margin of 5 mm specified by a partner cardiologist. We classify ‘ablated’ vs. ‘non-ablated’ mapped sites using: (1) multivariate logistic regression (MLR) (2) CNN9 a 9 layered convolutional neural network. A binary label identifying whether the mapped site contains STD pattern according to the interventional cardiologist is combined to raw EGMs as classifiers’ input. This additional input can be assimilated to a prior probability in STD-guided ablation. Data augmentation is used to handle the dataset imbalance. We opt for random oversampling, previously proven to enhance the classification AUC by 30%. Dropout is also used to prevent overfitting.

Results: Accuracy, AUC, sensitivity and specificity are, respectively, 0.74, 0.79, 0.76 and 0.73 for MLR and 0.77, 0.84, 0.74, 0.79 for CNN9. The CNN9 yields a slightly better performance. More elaborate ML architectures were also tested but did not provide improved results.

Conclusions: ML techniques can automatically identify ablated sites and are thus expected to guide cardiologists in patient-tailored catheter ablation procedures for treating persistent AF.