Maciej Gratkowski

Inverse modeling on decomposed electroencephalographic data: a way forward?

Inverse modeling is typically applied to instantaneous electroencephalogram signals. However, this approach has several shortcomings including its instability to model multiple and deep located dipole sources and the interference of background noise may hamper the sensitivity, stability, and precision of the estimated dipoles. This article validates different dipole estimation techniques to find the most optimal combination of different analysis principles using both simulations and recordings. Electroencephalogram data were simulated with six known source locations. First, a dataset was simulated with sources chosen to mimic somatosensory-evoked potentials to electrical stimuli. Additionally, 20 further datasets were simulated each containing six randomly located and oriented sources. The simulated sources included superficial, deep, and simultaneously active sources. Furthermore, somatosensory-evoked potentials to median nerve stimuli were recorded from one subject. On both simulated and recorded evoked potential data, three different methods of signal decomposition were compared: independent component analysis (ICA), second-order blind identification (SOBI), and multichannel matching pursuit (MMP). For inverse modeling of the brain sources, the DIPFIT function of the EEGLAB software was used on raw and decomposed data. MMP was able to separate all simulated components that corresponded to superficial, deep, and simultaneously active sources. ICA and SOBI were only able to find components that corresponded to superficial sources. For the 20 randomized simulations, the results from the evoked potential simulation were reproduced. Inverse modeling on MMP components (atoms) was better than on ICA or SOBI components (P < 0.001). DIPFIT on MMP atoms localized 99.2% of the simulated dipoles in correct areas with their correct time/frequency distribution. DIPFIT on ICA and SOBI components localized 35% and 39.6%, respectively of the simulated dipoles in correct areas. As for the real-evoked potentials recorded in one subject, DIPFIT on MMP atoms allowed us to build a dipole model closer to the current physiological knowledge than dipole modeling of ICA and SOBI components. The results show that using MMP before inverse modeling is a reliable way to noninvasively estimate cortical activation.

Decomposition of biomedical signals in spatial and time-frequency modes.

OBJECTIVES The purpose of this paper is to introduce a new method for spatial-time-frequency analysis of multichannel biomedical data. We exemplify the method for data recorded with a 31-channel Philips biomagnetometer. METHODS The method creates approximations and decompositions of spatiotemporal signal distributions using elements (atoms) chosen from a very large and redundant set (dictionary). The method is based on the Matching Pursuit algorithm, but it uses atoms that are distributed both in time and space (instead of only time-distributed atoms in standard Matching Pursuit). The time-frequency distribution of signal components is described by Gabor atoms and their spatial distribution is modeled by spatial modes. The spatial modes are created with the help of Bessel functions. Two versions of the method, differing in the definition of spatial properties of the atoms, are presented. RESULTS The technique was validated on simulated data and real magnetic field data. It was used for removal of powerline noise from multichannel magnetoencephalography data, extraction of high-frequency somatosensory evoked fields and for separation of partially overlapping T- and U-waves in magnetocardiography. CONCLUSIONS The method allows for parameterization of multichannel data in the time-frequency as well as in the spatial domains. It thus can be used for signal preserving filtering simultaneously in time, frequency, and space. Applications are e.g. the elimination of artifact components, extraction of components with biological meaning, and data exploration.

Time-frequency filtering of MEG signals with matching pursuit

Time-frequency signal analysis based on various decomposition techniques is widely used in biomedical applications. Matching Pursuit is a new adaptive approach for time-frequency decomposition of such biomedical signals. Its advantage is that it creates a concise signal approximation with the help of a small set of Gabor atoms chosen iteratively from a large and redundant set. In this paper, the usage of Matching Pursuit for time-frequency filtering of biomagnetic signals is proposed. The technique was validated on artificial signals and its performance was tested for varying signal-to-noise ratios using both simulated and real MEG somatic evoked magnetic field data.

Topographic analysis of engagement and disengagement of neural oscillators in photic driving: a combined electroencephalogram/magnetoencephalogram study.

Objective A coupled system of nonlinear neural oscillators with an individual resonance frequency is assumed to form the neuronal substrate for the photic driving phenomenon. The aim was to investigate the spatiotemporal stability of these oscillators and quantify the spatiotemporal process of engagement and disengagement of the neuronal oscillators in both multitrial and single-trial data. Methods White light-emitting diode flicker stimulation was used at 15 frequencies, which were set relative to the individual &agr; frequency of each of the 10 healthy participants. Simultaneously, the electroencephalogram (EEG) and the magnetoencephalogram (MEG) were recorded. Subsequently, spatiotemporal matching pursuit (MP) algorithms were used to analyze the EEG and MEG topographies. Results Intraindividually similar topographies were found at stimulation frequencies close to (1) the individual &agr; frequency and (2) half the individual &agr; frequency in the multitrial and the single-trial cases. In both stimulation frequency ranges, the authors observed stable topographies 5 to 10 stimuli after the beginning of the stimulation and lasting three nonexisting periods after the end of the stimulation. This was interpreted as the engaging/disengaging effect of the observed oscillations, because especially the frequency parameter adopted before and after stable topographies were observed. Topographic entrainment was slightly more pronounced in MEG as compared with that in EEG. Conclusions The results support the hypothesis of nonlinear information processing in human visual system, which can be described by nonlinear neural oscillators.

Comparison of wavelet transform and Matching Pursuit in the analysis of EEG and MEG signals

INTRODUCTION Time-frequency analysis of biomedical signals is an important problem. There have been several jpproaches to this matter, for example Wavelet Transtbrm (WT) or Short Time Fourier Transform (STFT). Al l of these methods have both advantages and disadvantages. For instance, for many applications WT has insufticient frequency resolution at higher trequencies. Adaptive algorithms, such äs Matching Pursuit (MP), represent a relatively new approach to time-frequency analysis. One advantage of the MP algorithm is its ability to create a concise description of the analysed signal and an adaptive time-frequency resolution. We implemented a MP algorithm and applied it to artificial test signals and real EEG and MEG signals and compared it to WT.

Multichannel matching pursuit validation and clustering: A simulation and empirical study

INTRODUCTION Multichannel matching pursuit (MMP) is a relatively new method that can be applied to electroencephalogram (EEG) signals in combination with inverse modelling. However, limitations of MMP have not been adequately tested. The aims of this study were to investigate how the accuracy of MMP algorithm is altered due to increased number of brain sources and increased noise level, and to implement and test a modified K-means clustering algorithm in order to group similar MMP atoms in time-frequency and space between subjects together. METHODS Four groups of 20 EEG signals were simulated. The groups consisted of simulations with 5, 10, 15, and 20 brain sources. The accuracy of MMP algorithm was first tested on increasing number of sources. Then, different levels of noise were added to the simulations and accuracy of the algorithm was tested on increasing noise level. K-means clustering algorithm was tested on 4 datasets (5, 10, 15, and 20 sources) of 10 similar phantom subjects. Finally, the clustering algorithm was tested on empirical somatosensory evoked potential and brainstem evoked potential data. RESULTS The MMP accuracy decreased as the number of sources increased and MMP accuracy was robust to noise. Furthermore, we found that when applying the clustering method to a subject groups MMP data, the clustering method grouped the similar atoms between subjects correctly. CONCLUSION The MMP and clustering method proved to be an efficient way to group similar brain activity and thus study differences in brain activation sequence to sensory stimulation between groups of subjects.

Power spectral density changes and language lateralization during covert object naming tasks measured with high-density EEG recordings.

Our objective was to study changes in EEG time-domain power spectral density (PSDt) and localization of language areas during covert object naming tasks in human subjects with epilepsy. EEG data for subjects with epilepsy were acquired during the covert object naming tasks using a net of 256 electrodes. The trials required each subject to provide the names of common objects presented every 4 seconds on slides. Each trial comprised the 1.0 second before and 3.0 seconds after initial object presentation. PSDt values at baseline and during tasks were calculated in the theta, alpha, beta, low gamma, and high gamma bands. The spatial contour plots reveal that PSDt values during object naming were 10-20% higher than the baseline values for different bands. Language was lateralized to left frontal or temporal areas. In all cases, the Wada test disclosed language lateralization to the left hemisphere as well.

Time-Frequency-Space Components of Somatosenory Evoked Potentials in Rats

This paper presents a description of components extracted from somatosensory evoked potentials recorded from rats after electrical stimulation of the forelimb. For the decomposition of SEPs Topographic Matching Pursuit method was used. This method is able to decompose multichannel data into spatio-temporal atoms defined by only a few parameters. Such parameterization allows convenient processing and analysis of data components. In the present study four distinct components could be identified in the SEP data.