Nonnegative matrix factorization for noninvasive prediction of catheter ablation outcome in persistent atrial fibrillation

Despite the increasing popularity of catheter ablation (CA) for treating atrial fibrillation (AF), the identification of patients who would actually benefit from the therapy remains a challenging open issue. This study aims at noninvasively predicting CA outcome by quantifying the spatio-temporal variability of the atrial activity (AA) signal measured on the standard 12-lead electrocardiogram. The normalized mean square error (NMSE) between consecutive atrial segments and their principal component approximations is computed for each lead, as a recent noninvasive index of AF organization. In the present work, the multilead NMSE array is decomposed by means of a nonnegative matrix factorization (NNMF) with two different initializations. The reconstruction error between the original NMSE matrix and its low-rank NNMF approximation is taken as a classification feature. A dataset of persistent AF patients undergoing CA reveals that the proposed feature is able to predict the therapy outcome with a notably higher level of statistical significance than recent single-lead indices.