Atrial Fibrillation (AF) is a heart's arrhythmia which, despite being often asymptomatic, represents an important risk factor for stroke, therefore being able to predict AF at the electrocardiogram exam, would be of great impact on actively targeting patients at high risk. In the present work we use Convolution Neural Networks to analyze ECG and predict Atrial Fibrillation starting from realistic datasets, i.e. considering fewer ECG than other studies and extending the maximal distance between ECG and AF diagnosis. We achieved 75.5% (0.75) AUC firstly increasing our dataset size by a shifting technique and secondarily using the dilation parameter of the convolution neural network. In addition we find that, contrarily to what is commonly used by clinicians reporting AF at the exam, the most informative leads for the task of predicting AF are D1 and avR. Similarly, we find that the most important frequencies to check are in the range of 5-20 Hz. Finally, we develop a net able to manage at the same time the electrocardiographic signal together with the electronic health record, showing that integration between different sources of data is a profitable path. In fact, the 2.8% gain of such net brings us to a 78.6% (std 0.77) AUC. In future works we will deepen both the integration of sources and the reason why we claim avR is the most informative lead.