Abstract : Atrial fibrillation (AF), the most common arrhythmia in adults, is still considered as the last great frontier of cardiac electrophysiology, since its mechanisms are not completely understood. Analysis of the atrial activity (AA) signal contained in electrocardiograms during AF episodes is a noninvasive and inexpensive solution for obtaining useful information about AF. This work presents tensor decompo-sitions as an alternative to classic blind source separation methods based on matrix decompositions due to their appealing uniqueness properties and considers in particular the block term decomposition (BTD). The practical usefulness of BTD is evaluated by comparing its AA estimation quality, measured by spectral concentration, to those of two benchmark methods, revealing that BTD presents a better performance. The results presented in this work motivate further investigation of tensor decompositions for AF analysis.