Atrial fibrillation is the most common sustained cardiac rhythm disturbance, increasing in prevalence with age. During the past 20 years, there has been a 66% increase in hospital admissions related to atrial fibrillation. Neither the natural history of atrial fibrillation nor its response to therapy is sufficiently predictable from clinical and echocardiographic parameters. Treatment of atrial fibrillation is mainly based on trial and error. Thus, it seems appropriate to develop tests that quantify the state of the disease and guide its management. Standard 12-lead electrocardiogram recordings are commonly required for clinical evaluation. Therefore, possible prognostic information contained within the electrocardiogram provides a great interest. The goal of this thesis is to help clinicians treat atrial fibrillation by developing information from the standard 12-lead electrocardiogram on atrial fibrillation substrates, dynamics, and to predict the success of different treatments. Due to the much higher ventricular activity amplitude, the characterization of atrial fibrillation based on surface electrocardiogram signals requires that the ventricular activity first be cancelled. Average beat subtraction and independent component analysis algorithms are the most frequently used techniques in atrial activity extraction. To solve this problem with the best quality results, five different techniques were studied and compared in the first part of this thesis : the use of two different independent component analysis algorithms, a refined version of the average beat subtraction technique, a novel technique that treats each cardiac cycle in an independent manner, and a novel approach based on the use of atom dictionaries dedicated to atrial and ventricular activities. The performance of these five techniques was evaluated by using simulated and clinical electrocardiogram signals. The simulated signals were created by using models of the atria and the thorax. The refined version of the average beat subtraction technique and the technique that treats each cardiac cycle outperformed the other ones and produced high quality results on all leads. Their main advantage lies in the treatment of the ventricular depolarization and repolarization waves in an independent manner. The second part of this thesis aims at exploring the potential of the "clean" atrial electrocardiogram signals. So far, the complexity of the electrical atrial activity during atrial fibrillation has mainly been assessed through the analysis of the frequency spectrum or of the time-frequency analysis (spectrogram) of the electrocardiogram signal of lead V1. We propose three different approaches to characterize the atrial fibrillation that exploit the information of multiple leads in various and independent manners. Each of them considers different features that require different signal processing techniques. The first approach is based on clinician's observations on electrocardiograms, to characterize the disorganization of fibrillatory waves. We developed the processing tools to quantify these observations. Based on these clinical features, we were able to predict self-termination of atrial fibrillation with a high accuracy. We also observed an interesting correlation between the increase in the percentage of atrial fibrillations identified as non-terminating and the increase in atrial fibrillation duration. The second approach is based on dominant frequencies observed on multiple electrocardiogram signals. Firstly, two simulated cases were analyzed to understand the correspondence between the dominant frequencies observed on the thorax and the atrial fibrillation dynamics. We observed that multiple lead signals yield more information than a single lead in terms of atrial fibrillation dynamic and that lead V1 and V5 were good candidates for the observation of the left and right atrial dynamics, respectively. In order to obtain accurate dominant frequency estimation on clinical signals that contain ventricular artifacts, a recently introduced signal processing technique was tested. We observed a positive left-to-right dominant frequency gradient preference in the overall atrial fibrillations when a negative gradient preference was to be expected based on other invasive studies. These dominant frequency features also permitted us to obtain a good prediction of the response to pharmacological cardioversion attempts. Our third approach was to extract spatial information from the dipole estimated from body surface potentials. This dipole provides a global representation of the electrical atrial activity. Then, features were used to express the complexity of the atrial fibrillation dynamics. This analysis extracted spatial information that was generally lacking in typical non-invasive atrial fibrillation studies. We observed discrimination between atrial fibrillations with both single and recurrent episodes, characterized by modal and uniform dipole distributions, respectively. This discrimination could link the different atrial fibrillation types to the electrical, contractile, and structural remodelings. This approach permitted us to identify and localize stable and single atrial activity sources in the simulations. We hypothesize that this approach, together with patient's anatomical data, could have the power to help clinicians in ablation procedures ; by helping them to predict where ablation lines can be effective.