Atrial fibrillation (AF) is the most common sustained arrhythmia encountered in clinical practice, and one of the main causes of stroke. Yet its thorough characterization and treatment remain an open issue. Despite the increasing popularity of the radiofrequency catheter ablation (CA) therapy, very little is known about its impact on heart substrate, leading to rather uncertain success rates. This calls for advanced signal processing tools for quantitatively assessing CA outcome. The surface 12-lead electrocardiogram (ECG), a noninvasive and cost-effective cardiac activity recording modality, provides valuable information about AF. However, some issues affect most of the standard CA outcome predictors, e.g., manual computation and limited single-lead perspective. This thesis aims at explicitly exploiting the ECG’s multilead character through multivariate decomposition tools, so as to enhance the role of some ECG features as CA outcome predictors. Fibrillatory wave amplitude is correlated with CA success in a multilead framework through principal component analysis (PCA). Multivariate approaches also enhance AF spatiotemporal variability measured on the ECG (e.g., weighted PCA, nonnegative matrix factorization), evidencing that the less repetitive the AF pattern, the less likely CA success. Information theory also quantifies interlead similarity between AF patterns, and is linked with CA outcome in a multilead framework. Another perspective focuses on the ventricular response as reflected on heart rate variability (HRV). Point process modeling can highlight certain HRV properties typical of AF in a parametric probabilistic context, helping AF pattern recognition.
