Aims: We propose to evaluate the performance of machine learning methods and especially the neural networks on the cardiac electrical mapping. Our idea is to built a specific function that directly constructs activation maps from the body surface potentials. In our study we use a dataset of 100 different simulations generated using the monodomain model and a realistic 3D atria-torso geometry segmented from CT-Scan images. The training dataset is obtained by time/space resampling of the simulation. The body surface measurements contain 264 nodes and the atria mesh 1994 nodes.
Methods: First, the simulated activation maps are constructed using the maximum negative slope rule. Then, we design a classic artificial neural network constituted of 3 linear layers and we use the rectified linear unit (RELu) as an activation function. We use the stochastic gradient descent as an optimization algorithm and the mean squared error between simulated and computed activation maps as an objective function. This neural network has as input Body surface potential temporal series and aims to reconstruct the corresponding activation map. The learning process has 3 phases: Training, validation to avoid overfitting and testing. The training stops when the validation absolute activation error stops improving. Finally, we evaluate the method performance by computing the mean absolute activation time error (ATE), the relative error (RE) and the correlation coefficient (CC) between the simulated and reconstructed activation maps.
Results: The mean absolute activation time error reached 7ms in the training phase and 9ms during the testing phase. Mean relative errors between simulated and reconstructed activation maps is 7% and mean correlation coefficient is 99%, results of the testing phase.
Conclusion: The neural network method is a good alternative for computing for the cardiac activation maps. It allows cardiac activation mapping directly from body surface potentials with a good accuracy.