Consensus Matching Pursuit of Multi-Trial Biosignals, with Application to Brain Signals

Time-frequency representations are commonly used to analyze the oscillatory nature of bioelectromagnetic signals. There is a growing interest in sparse representations, where the data is described using few components. In this study, we adapt the Matching Pursuit of Mallat and Zhang for biosignals consisting of a series of variations around a similar pattern, with emphasis on multi-trial datasets encountered in MEG and EEG. The general principle of Matching Pursuit (MP) is to iteratively subtract from the signal its projection on the atom selected from a dictionary. The originality of our method is to select each atom using a voting technique that is robust to variability, and to subtract it by adapting the parameters to each trial. Because it is designed to handle inter-trial variability using a voting technique, the method is called Consensus Matching Pursuit (CMP). The method is validated on both simplified and realistic simulations, and on two real datasets (intracerebral EEG and scalp EEG ). We also compare our method to two other multi-trial MP algorithms: Multivariate MP (MMP) and Induced activity MP (IMP). CMP is shown to be able to sparsely reveal the structure present in the data, and to be robust to variability (jitter) across trials.