Background. Atrial fibrillation (AF) is characterized by complex and irregular propagation patterns, and AF onset locations and drivers responsible for its perpetuation are main targets for ablation procedures.
Aim. We propose to estimate the zone where AF drivers are found from body surface potentials (BSP) and Convolutional Neural Networks (CNN).
Methods. We used realistic computerized models of atria (N=2039 nodes) and torso (M=659 nodes). Atria and torso models were used to simulate a normal sinus rhythm and 13 different AF propagation patterns in both left atria (LA) and right atria (RA), with different complexity and driver positions.
We addressed the localization problem as a supervised classification problem. We divided the atria into 7 regions. Target variable, AF driver localization, represents seven classes corresponding to each atria region (class 0 if there is no driver found). AF drivers, if exist, were manually classified into one region for each time-instant. Input data are tridimensional matrices obtained from 64 BSP signals with a SNR = 20dB (see Fig. 1).
We trained a Convolutional Neural Network with convolutional layers of size (32, 32, 3), (64, 64, 3) and (64, 64, 3). Max-pooling is applied after convolutional layers ((2,2) window size). We added two dense layers of size (128, 64) units. Output layer is composed by 8 units to perform the classification. We used ReLu activation function in convolutional and dense layers and Softmax in the output.
Results. We were able to locate the 98% and 89% of drivers in the training and test sets, respectively. In the same scenario, Cohen's Kappa metric was 0.97 and 0.85.
Conclusions. The proposed CNN-based method could help to identify the area where AF drivers are found using body surface potential mapping (BSPM), avoiding the use of ECGI.