Cardiac mapping is a technique for diagnosing and treating arrhythmias by characterizing the electrical activity of the heart. Electrocardiographic Imaging (ECGI) uses electrodes on the torso to capture Body Surface Potentials (BSPs) to reconstruct electrical potentials and repolarization patterns from the surface of the heart non-invasively. Current ECGI algorithms, however, lack noise robustness and stability. This paper proposes a method o reconstruct the electrical activity of the heart using a combination of two architectures: a convolutional autoencoder to extract features from BSPs and a CNN-LSTM Network to reconstruct the electrograms (EGMs) from the extracted features. The proposed strategy was evaluated on 44 atrial fibrillation computational models and provided maximum mean values of correlation of 0.5 and minimum mean RMSE values of 0.48. The reconstruction, latent space and real EGMs maintain the same spectral information, yielding a promising method to address this problem.