Analysis of EEG and MEG Signals with Matching Pursuit

Introduction: A wide spectrum of methods has been used for analysis of biomedical signals. A relatively new and promising approach to signal analysis is matching pursuit (MP). The main advantage of the MP algorithm is its ability to create a concise description of an analyzed signal with a relatively small number of components localized in the time-frequency space with maximal possible, adaptive resolution. We applied MP to real MEG and EEG signals in order to present its possible applications. Methods: MP creates signal approximations based on a linear combination of a small number of vectors (atoms) chosen from a bigger, redundant set (dictionary). The atoms are iteratively chosen in order to best match the signal structure. In the first iteration the best matching atom is subtracted from the signal. This creates a signal residuum. The next step is to find the best matching atom with the residuum. This procedure is repeated until the desired approximation energy level is reached. With a dictionary of Gabor time-frequency atoms (scaled, translated and modulated Gauss functions) MP is an adaptive time-frequency transformation. The Gabor atoms are described by five parameters. These parameters localize the atom in the time-frequency space. We applied MP to real MEG and EEG data. Neuromagnetic fields were recorded over the contralateral somatosensory cortex with a 31-channel biomagnetometer. Constant current 0.2 ms square wave pulses were delivered to the right wrist at a stimulation rate of 4Hz. The individual sensory and motor thresholds were determined and the stimulation strength was set to a value equal to the sum of sensory and motor thresholds. Simultaneously electric scalp potentials were recorded using 32 electrodes. We analyzed the so-called 600Hz component and filtered out artefact components. Conclusions: The description of the 600Hz component obtained using MP was consistent with the one found in a previous study, where a Gabor filter was used for the analysis of the same data. The advantage of MP is that the signals were well described with only 15 components. The parameterization of the signals makes further processing very simple. The possible applications of MP are filtering, i.e., artefact rejection (stimulus artefacts or 50Hz powerline noise) or separation of the signal components that are of interest (e.g., 600Hz components) and automatic signal recognition through detection of characteristic signal components or component groups.