Supraventricular tachyarrhythmias, in particular Atrial Fibrillation (AF), are the most commonly cardiac diseases encountered in the routine clinical practice. The prevalence of AF is less than 1% among population under 60 years old, but it increases significantly among those over 70, approximating to 10% in those older than 80.  Undergoing a sutained AF episode is related to a higher mortality ratio and to a rising probability of suffering thromboembolisms, myocardial infarction, and stroke. On the other hand,  paroxysmal AF (PAF), which is characterized by its  spontaneous termination, is frequently the  precursor to sustained AF. This provokes a great interest among the scientific community in disclosing the mechanisms which lead to AF perpetuation or to spontaneous AF termination.    

The analysis of the surface electrocardiogram (ECG) is the most extended noninvasive technique in medical diagnosis of cardiac pathologies. In order to use the ECG as a tool for the AF study,  the atrial activity (AA) must be separated  from other cardioelectric signals. In this sense,  Blind Source Separation (BSS) techniques are able to perform a multi-lead statistical analysis with the aim to obtain a set of independent sources that include  the AA. When the  BSS problem is tackled, it becomes necessary to  consider a source mixing model near to the real mixing process in order to develop mathematical algorithms that solve the problem. A feasible model consists of  assuming  the linear  mixture of sources. Within this linear mixing model it can be made the additional assumption of instantaneous mixture. This instantaneous linear mixing model is the one used in Independent Component Analysis (ICA). An alternative mixing model is considered by convolutive BSS (CBSS) algorithms, where a more realistic process in the generation of ECG leads is taken into account with delayed contributions of cardioelectric sources.  

In this thesis, a performance study of CBSS algorithms applied to AA extraction from ECG recordings has been carried out for the first time. With this aim, the most relevant CBSS algorithms have been compared with the instantaneous  algorithm FastICA, the effectivity of which is extensively proved.  This comparison will allow to know which CBSS algorithms are useful for AA extraction from ECG recordings of AF episodes. On the other hand, CBSS algorithms have the problem of requiring a minimum number of observed signals for their suitable application. Here a new AA extraction algorithm is presented, which is based on the convolutive mixing model and solves the problem of lack of available leads  from Holter ECG recordings. The high likeliness level between original and estimated AA, measured by different performance indicators, demonstrates the suitability of this new method for the AA extraction from Holter recordings and,  furthermore, a higher robustness against noise of the convolutive mixing model is highlighted.  
  
The most common cause of undergoing an AF episode is attributed to the reentry mechanism, which consists of multiple wavelet fronts that propagate through the atrial tissue. Recent studies have proved a relation between the number of simultaneous reentries and atrial electrical activity organization of the atria.  In addition, it is also known the progressive deterioration of the atrial electrical activity organization after the PAF onset and the organization increase prior to its return to normal sinus rhythm. Furthermore, the maintenance of AF episodes is related to frequency dispersion variability  of AA. From a clinic point of view, it is interesting to evaluate the atrial electrical activity by means of regularity indicators applied to the AA extracted from the surface ECG with the aim to predict the evolution of PAF episodes. This can be carried out from two different perspectives.  Firstly, a regularity estimation can be applied to the raw AA  obtained by extraction algorithms. An alternative way to tackle the problem is to study the regularity of certain spectral feature throughout time. This last aspect has never been considered before and is a matter of study in this thesis. 
 
The work of this thesis finishes with the presentation of a new method for predicting the early termination or the maintenance of PAF episodes. This new method is based on the regularity analysis of AA spectral features. During the method design, the regularity of twelve numerical series of spectral features was analyzed. In order to construct these series, the spectrogram of the AA was previously computed. The series regularity was estimated by the nonlinear regularity estimator Sample Entropy (SampEn). The SampEn of six spectral features were revealed as statistically relevant to PAF characterization with p<0.05 for all features. This study was complemented with a multivariate regularity analysis that executes a joint study of the spectral features series and the AA in time domain   The multivariate analysis discloses the combination of features that optimizes the prediction  with 100% of correctly classified recordings   for the learning set and 93.33% for the test set. Consequently, the presented method can be reliably used for predicting PAF termination.