A K-means multivariate approach for clustering independent components from magnetoencephalographic data

Independent component analysis (ICA) is typically applied on functional magnetic resonance imaging, electroencephalographic and magnetoencephalographic (MEG) data due to its data-driven nature. In these applications, ICA needs to be extended from single to multi-session and multi-subject studies for interpreting and assigning a statistical significance at the group level. Here a novel strategy for analyzing MEG independent components (ICs) is presented, Multivariate Algorithm for Grouping MEG Independent Components K-means based (MAGMICK). The proposed approach is able to capture spatio-temporal dynamics of brain activity in MEG studies by running ICA at subject level and then clustering the ICs across sessions and subjects. Distinctive features of MAGMICK are: i) the implementation of an efficient set of "MEG fingerprints" designed to summarize properties of MEG ICs as they are built on spatial, temporal and spectral parameters; ii) the implementation of a modified version of the standard K-means procedure to improve its data-driven character. This algorithm groups the obtained ICs automatically estimating the number of clusters through an adaptive weighting of the parameters and a constraint on the ICs independence, i.e. components coming from the same session (at subject level) or subject (at group level) cannot be grouped together. The performances of MAGMICK are illustrated by analyzing two sets of MEG data obtained during a finger tapping task and median nerve stimulation. The results demonstrate that the method can extract consistent patterns of spatial topography and spectral properties across sessions and subjects that are in good agreement with the literature. In addition, these results are compared to those from a modified version of affinity propagation clustering method. The comparison, evaluated in terms of different clustering validity indices, shows that our methodology often outperforms the clustering algorithm. Eventually, these results are confirmed by a comparison with a MEG tailored version of the self-organizing group ICA, which is largely used for fMRI IC clustering.     

Localization of individual area neuronal activity

A family of methods, collectively known as independent component analysis (ICA), has recently been added to the array of methods designed to decompose a multi-channel signal into components. ICA methods have been applied to raw magnetoencephalography (MEG) and electroencephalography (EEG) signals to remove artifacts, especially when sources such as power line or cardiac activity generate strong components that dominate the signal. More recently, successful ICA extraction of stimulus-evoked responses has been reported from single-trial raw MEG and EEG signals. The extraction of weak components has often been erratic, depending on which ICA method is employed and even on what parameters are used. In this work, we show that if the emphasis is placed on individual "independent components," as is usually the case with standard ICA applications, differences in the results obtained for different components are exaggerated. We propose instead the reconstruction of regional brain activations by combining tomographic estimates of individual independent components that have been selected by appropriate spatial and temporal criteria. Such localization of individual area neuronal activity (LIANA) allows reliable semi-automatic extraction of single-trial regional activations from raw MEG data. We demonstrate the new method with three different ICA algorithms applied to both computer-generated signals and real data. We show that LIANA provides almost identical results with each ICA method despite the fact that each method yields different individual components.     

Validating the independent components of neuroimaging time series via clustering and visualization

Recently, independent component analysis (ICA) has been widely used in the analysis of brain imaging data. An important problem with most ICA algorithms is, however, that they are stochastic; that is, their results may be somewhat different in different runs of the algorithm. Thus, the outputs of a single run of an ICA algorithm should be interpreted with some reserve, and further analysis of the algorithmic reliability of the components is needed. Moreover, as with any statistical method, the results are affected by the random sampling of the data, and some analysis of the statistical significance or reliability should be done as well. Here we present a method for assessing both the algorithmic and statistical reliability of estimated independent components. The method is based on running the ICA algorithm many times with slightly different conditions and visualizing the clustering structure of the obtained components in the signal space. In experiments with magnetoencephalographic (MEG) and functional magnetic resonance imaging (fMRI) data, the method was able to show that expected components are reliable; furthermore, it pointed out components whose interpretation was not obvious but whose reliability should incite the experimenter to investigate the underlying technical or physical phenomena. The method is implemented in a software package called Icasso.     

Selection of independent components representing event-related brain potentials: a data-driven approach for greater objectivity

Following the development of increasingly precise measurement instruments and fine-grain analysis tools for electroencephalographic (EEG) data, analysis of single-trial event-related EEG has considerably widened the utility of this non-invasive method to investigate brain activity. Recently, independent component analysis (ICA) has become one of the most prominent techniques for increasing the feasibility of single-trial EEG. This blind source separation technique extracts statistically independent components (ICs) from the EEG raw signal. By restricting the signal analysis to those ICs representing the processes of interest, single-trial analysis becomes more flexible. Still, the selection-criteria for in- or exclusion of certain ICs are largely subjective and unstandardized, as is the actual selection process itself. We present a rationale for a bottom-up, data-driven IC selection approach, using clear-cut inferential statistics on both temporal and spatial information to identify components that significantly contribute to a certain event-related brain potential (ERP). With time-range being the only necessary input, this approach considerably reduces the pre-assumptions for IC selection and promotes greater objectivity of the selection process itself. To test the validity of the approach presented here, we present results from a simulation and re-analyze data from a previously published ERP experiment on error processing. We compare the ERP-based IC selections made by our approach to the selection made based on mere signal power. The comparison of ERP integrity, signal-to-noise ratio, and single-trial properties of the back-projected ICs outlines the validity of the approach presented here. In addition, functional validity of the extracted error-related EEG signal is tested by investigating whether it is predictive for subsequent behavioural adjustments.     

Independent component analysis of short-time Fourier transforms for spontaneous EEG/MEG analysis

Analysis of spontaneous EEG/MEG needs unsupervised learning methods. While independent component analysis (ICA) has been successfully applied on spontaneous fMRI, it seems to be too sensitive to technical artifacts in EEG/MEG. We propose to apply ICA on short-time Fourier transforms of EEG/MEG signals, in order to find more “interesting” sources than with time-domain ICA, and to more meaningfully sort the obtained components. The method is especially useful for finding sources of rhythmic activity. Furthermore, we propose to use a complex mixing matrix to model sources which are spatially extended and have different phases in different EEG/MEG channels. Simulations with artificial data and experiments on resting-state MEG demonstrate the utility of the method.     

Co-modulatory spectral changes in independent brain processes are correlated with task performance

This study investigates the independent modulators that mediate the power spectra of electrophysiological processes, measured by electroencephalogram (EEG), in a sustained-attention experiment. EEG and behavioral data were collected during 1-2 hour virtual-reality based driving experiments in which subjects were instructed to maintain their cruising position and compensate for randomly induced drift using the steering wheel. Independent component analysis (ICA) applied to 30-channel EEG data separated the recorded EEG signals into a sum of maximally temporally independent components (ICs) for each of 30 subjects. Logarithmic spectra of resultant IC activities were then decomposed by principal component analysis, followed by ICA, to find spectrally fixed and temporally independent modulators (IM). Across subjects, the spectral ICA consistently found four performance-related independent modulators: delta, delta-theta, alpha, and beta modulators that multiplicatively affected the spectra of spatially distinct IC processes when the participants experienced waves of alternating alertness and drowsiness during long-hour simulated driving. The activation of the delta-theta modulator increased monotonically as subjects' task performances decreased. Furthermore, the time courses of the theta-beta modulator were highly correlated with concurrent changes in driving errors across subjects (r=0.77±0.13).     

Enhanced detection of artifacts in EEG data using higher-order statistics and independent component analysis

Detecting artifacts produced in EEG data by muscle activity, eye blinks and electrical noise is a common and important problem in EEG research. It is now widely accepted that independent component analysis (ICA) may be a useful tool for isolating artifacts and/or cortical processes from electroencephalographic (EEG) data. We present results of simulations demonstrating that ICA decomposition, here tested using three popular ICA algorithms, Infomax, SOBI, and FastICA, can allow more sensitive automated detection of small non-brain artifacts than applying the same detection methods directly to the scalp channel data. We tested the upper bound performance of five methods for detecting various types of artifacts by separately optimizing and then applying them to artifact-free EEG data into which we had added simulated artifacts of several types, ranging in size from thirty times smaller (-50 dB) to the size of the EEG data themselves (0 dB). Of the methods tested, those involving spectral thresholding were most sensitive. Except for muscle artifact detection where we found no gain of using ICA, all methods proved more sensitive when applied to the ICA-decomposed data than applied to the raw scalp data: the mean performance for ICA was higher and situated at about two standard deviations away from the performance distribution obtained on raw data. We note that ICA decomposition also allows simple subtraction of artifacts accounted for by single independent components, and/or separate and direct examination of the decomposed non-artifact processes themselves.     

Neuromagnetic measurement of unilateral temporo-parietal theta rhythm in patients with internal carotid artery occlusive disease

Rhythmic theta activity detected by electroencephalography (EEG) may be correlated with cerebrovascular brain diseases. Magnetoencephalography (MEG) has higher sensitivity and spatial resolution than conventional scalp EEG, so may be a better method to detect theta rhythm in patients with internal carotid artery (ICA) occlusive disease. Simultaneous EEG and MEG were performed in the awake state in 48 patients with unilateral (n = 42) or bilateral (n = 6) stenotic lesions (more than 60% occlusion) of the ICA (n = 47) or middle cerebral artery (n = 7), and in 27 age-matched healthy normal subjects. No subject had severe neurological deficits. MEG detected the theta rhythm (6-8 Hz) in 14 of 48 patients: ipsilateral to the stenotic or occluded side in 13 hemispheres and bilaterally in one patient with unilateral lesion. The source of the MEG theta rhythm was estimated in the dorsolateral temporo-parietal area, regardless of the location of infarct foci or the stenotic portion of the ICA system. The temporo-parietal theta rhythm was separated from the occipital alpha rhythm by frequency and distribution in MEG. The theta rhythm was found in only two patients by EEG, as well as by MEG. MEG provided better separation of this theta rhythm from the occipital alpha rhythm. Neither MEG nor EEG detected this theta rhythm in the normal subjects. Unilateral temporo-parietal theta rhythm is correlated with the hemisphere with ICA occlusive disease. This rhythm may indicate mild or subclinical abnormalities in the ICA system. MEG is superior to EEG for the detection and localization of theta rhythm.     

Performance comparison of machine learning algorithms and number of independent components used in fMRI decoding of belief vs. disbelief

Machine learning (ML) has become a popular tool for mining functional neuroimaging data, and there are now hopes of performing such analyses efficiently in real-time. Towards this goal, we compared accuracy of six different ML algorithms applied to neuroimaging data of persons engaged in a bivariate task, asserting their belief or disbelief of a variety of propositional statements. We performed unsupervised dimension reduction and automated feature extraction using independent component (IC) analysis and extracted IC time courses. Optimization of classification hyperparameters across each classifier occurred prior to assessment. Maximum accuracy was achieved at 92% for Random Forest, followed by 91% for AdaBoost, 89% for Na?ve Bayes, 87% for a J48 decision tree, 86% for K*, and 84% for support vector machine. For real-time decoding applications, finding a parsimonious subset of diagnostic ICs might be useful. We used a forward search technique to sequentially add ranked ICs to the feature subspace. For the current data set, we determined that approximately six ICs represented a meaningful basis set for classification. We then projected these six IC spatial maps forward onto a later scanning session within subject. We then applied the optimized ML algorithms to these new data instances, and found that classification accuracy results were reproducible. Additionally, we compared our classification method to our previously published general linear model results on this same data set. The highest ranked IC spatial maps show similarity to brain regions associated with contrasts for belief > disbelief, and disbelief < belief.     

Testing the ICA mixing matrix based on inter-subject or inter-session consistency

Independent component analysis (ICA) is increasingly used for analyzing brain imaging data. ICA typically gives a large number of components, many of which may be just random, due to insufficient sample size, violations of the model, or algorithmic problems. Few methods are available for computing the statistical significance (reliability) of the components. We propose to approach this problem by performing ICA separately on a number of subjects, and finding components which are sufficiently consistent (similar) over subjects. Similarity is defined here as the similarity of the mixing coefficients, which usually correspond to spatial patterns in EEG and MEG. The threshold of what is "sufficient" is rigorously defined by a null hypothesis under which the independent components are random orthogonal components in the whitened space. Components which are consistent in different subjects are found by clustering under the constraint that a cluster can only contain one source from each subject, and by constraining the number of the false positives based on the null hypothesis. Instead of different subjects, the method can also be applied on different recording sessions from a single subject. The testing method is particularly applicable to EEG and MEG analysis.     

MEG/EEG sources of the 170-ms response to faces are co-localized in the fusiform gyrus

The 170-ms electrophysiological processing stage (N170 in EEG, M170 in MEG) is considered an important computational step in face processing. Hence its neuronal sources have been modelled in several studies. The current study aimed to specify the relation of the dipolar sources underlying N170 and M170. Whole head EEG and MEG were measured simultaneously during the presentation of unfamiliar faces. An Independent Component Analysis (ICA) was applied to the data prior to localization. N170 and M170 were then modelled with a pair of dipoles in a four-shell ellipse (EEG)/homogeneous sphere (MEG) arranged symmetrically across midline. The dipole locations were projected onto the individual structural MR brain images. Dipoles were localized in fusiform gyri in ten out of eleven individuals for EEG and in seven out of eleven for MEG. N170 and M170 were co-localized in the fusiform gyrus in six individuals. The ICA shifted some of the single-subject dipoles up from cerebellum to fusiform gyrus mainly due to the removal of cardiac activity. The group mean dipole locations were also found in posterior fusiform gyri, and did not differ significantly between EEG and MEG. The result was replicated in a repeated measurement 3 months later.     

Validation of SOBI components from high-density EEG

Second-order blind identification (SOBI) is a blind source separation (BSS) algorithm that can be used to decompose mixtures of signals into a set of components or putative recovered sources. Previously, SOBI, as well as other BSS algorithms, has been applied to magnetoencephalography (MEG) and electroencephalography (EEG) data. These BSS algorithms have been shown to recover components that appear to be physiologically and neuroanatomically interpretable. While some proponents of these algorithms suggest that fundamental discoveries about the human brain might be made through the application of these techniques, validation of BSS components has not yet received sufficient attention. Here we present two experiments for validating SOBI-recovered components. The first takes advantage of the fact that noise sources associated with individual sensors can be objectively validated independently from the SOBI process. The second utilizes the fact that the time course and location of primary somatosensory (SI) cortex activation by median nerve stimulation have been extensively characterized using converging imaging methods. In this paper, using both known noise sources and highly constrained and well-characterized neuronal sources, we provide validation for SOBI decomposition of high-density EEG data. We show that SOBI is able to (1) recover known noise sources that were either spontaneously occurring or artificially induced; (2) recover neuronal sources activated by median nerve stimulation that were spatially and temporally consistent with estimates obtained from previous EEG, MEG, and fMRI studies; (3) improve the signal-to-noise ratio (SNR) of somatosensory-evoked potentials (SEPs); and (4) reduce the level of subjectivity involved in the source localization process.     

Spectral spatiotemporal imaging of cortical oscillations and interactions in the human brain

This paper presents a computationally efficient source estimation algorithm that localizes cortical oscillations and their phase relationships. The proposed method employs wavelet-transformed magnetoencephalography (MEG) data and uses anatomical MRI to constrain the current locations to the cortical mantle. In addition, the locations of the sources can be further confined with the help of functional MRI (fMRI) data. As a result, we obtain spatiotemporal maps of spectral power and phase relationships. As an example, we show how the phase locking value (PLV), that is, the trial-by-trial phase relationship between the stimulus and response, can be imaged on the cortex. We apply the method to spontaneous, evoked, and driven cortical oscillations measured with MEG. We test the method of combining MEG, structural MRI, and fMRI using simulated cortical oscillations along Heschl's gyrus (HG). We also analyze sustained auditory gamma-band neuromagnetic fields from MEG and fMRI measurements. Our results show that combining the MEG recording with fMRI improves source localization for the non-noise-normalized wavelet power. In contrast, noise-normalized spectral power or PLV localization may not benefit from the fMRI constraint. We show that if the thresholds are not properly chosen, noise-normalized spectral power or PLV estimates may contain false (phantom) sources, independent of the inclusion of the fMRI prior information. The proposed algorithm can be used for evoked MEG/EEG and block-designed or event-related fMRI paradigms, or for spontaneous MEG data sets. Spectral spatiotemporal imaging of cortical oscillations and interactions in the human brain can provide further understanding of large-scale neural activity and communication between different brain regions.     

Inferring task-related networks using independent component analysis in magnetoencephalography

A novel framework for analysing task-positive data in magnetoencephalography (MEG) is presented that can identify task-related networks. Techniques that combine beamforming, the Hilbert transform and temporal independent component analysis (ICA) have recently been applied to resting-state MEG data and have been shown to extract resting-state networks similar to those found in fMRI. Here we extend this approach in two ways. First, we systematically investigate optimisation of time-frequency windows for connectivity measurement. This is achieved by estimating the distribution of functional connectivity scores between nodes of known resting-state networks and contrasting it with a distribution of artefactual scores that are entirely due to spatial leakage caused by the inverse problem. We find that functional connectivity, both in the resting-state and during a cognitive task, is best estimated via correlations in the oscillatory envelope in the 8-20 Hz frequency range, temporally down-sampled with windows of 1-4s. Second, we combine ICA with the general linear model (GLM) to incorporate knowledge of task structure into our connectivity analysis. The combination of ICA with the GLM helps overcome problems of these techniques when used independently: namely, the interpretation and separation of interesting independent components from those that represent noise in ICA and the correction for multiple comparisons when applying the GLM. We demonstrate the approach on a 2-back working memory task and show that this novel analysis framework is able to elucidate the functional networks involved in the task beyond that which is achieved using the GLM alone. We find evidence of localised task-related activity in the area of the hippocampus, which is difficult to detect reliably using standard methods. Task-positive ICA, coupled with the GLM, has the potential to be a powerful tool in the analysis of MEG data.     

Pattern analysis of EEG responses to speech and voice: influence of feature grouping

Pattern recognition algorithms are becoming increasingly used in functional neuroimaging. These algorithms exploit information contained in temporal, spatial, or spatio-temporal patterns of independent variables (features) to detect subtle but reliable differences between brain responses to external stimuli or internal brain states. When applied to the analysis of electroencephalography (EEG) or magnetoencephalography (MEG) data, a choice needs to be made on how the input features to the algorithm are obtained from the signal amplitudes measured at the various channels. In this article, we consider six types of pattern analyses deriving from the combination of three types of feature selection in the temporal domain (predefined windows, shifting window, whole trial) with two approaches to handle the channel dimension (channel wise, multi-channel). We combined these different types of analyses with a Gaussian Naïve Bayes classifier and analyzed a multi-subject EEG data set from a study aimed at understanding the task dependence of the cortical mechanisms for encoding speaker's identity and speech content (vowels) from short speech utterances (Bonte, Valente, & Formisano, 2009). Outcomes of the analyses showed that different grouping of available features helps highlighting complementary (i.e. temporal, topographic) aspects of information content in the data. A shifting window/multi-channel approach proved especially valuable in tracing both the early build up of neural information reflecting speaker or vowel identity and the late and task-dependent maintenance of relevant information reflecting the performance of a working memory task. Because it exploits the high temporal resolution of EEG (and MEG), such a shifting window approach with sequential multi-channel classifications seems the most appropriate choice for tracing the temporal profile of neural information processing.     

Performance evaluation of the Champagne source reconstruction algorithm on simulated and real M/EEG data

In this paper, we present an extensive performance evaluation of a novel source localization algorithm, Champagne. It is derived in an empirical Bayesian framework that yields sparse solutions to the inverse problem. It is robust to correlated sources and learns the statistics of non-stimulus-evoked activity to suppress the effect of noise and interfering brain activity. We tested Champagne on both simulated and real M/EEG data. The source locations used for the simulated data were chosen to test the performance on challenging source configurations. In simulations, we found that Champagne outperforms the benchmark algorithms in terms of both the accuracy of the source localizations and the correct estimation of source time courses. We also demonstrate that Champagne is more robust to correlated brain activity present in real MEG data and is able to resolve many distinct and functionally relevant brain areas with real MEG and EEG data.     

Independent component analysis of interictal fMRI in focal epilepsy: comparison with general linear model-based EEG-correlated fMRI

The general linear model (GLM) has been used to analyze simultaneous EEG-fMRI to reveal BOLD changes linked to interictal epileptic discharges (IED) identified on scalp EEG. This approach is ineffective when IED are not evident in the EEG. Data-driven fMRI analysis techniques that do not require an EEG derived model may offer a solution in these circumstances. We compared the findings of independent components analysis (ICA) and EEG-based GLM analyses of fMRI data from eight patients with focal epilepsy. Spatial ICA was used to extract independent components (IC) which were automatically classified as either BOLD-related, motion artefacts, EPI-susceptibility artefacts, large blood vessels, noise at high spatial or temporal frequency. The classifier reduced the number of candidate IC by 78%, with an average of 16 BOLD-related IC. Concordance between the ICA and GLM-derived results was assessed based on spatio-temporal criteria. In each patient, one of the IC satisfied the criteria to correspond to IED-based GLM result. The remaining IC were consistent with BOLD patterns of spontaneous brain activity and may include epileptic activity that was not evident on the scalp EEG. In conclusion, ICA of fMRI is capable of revealing areas of epileptic activity in patients with focal epilepsy and may be useful for the analysis of EEG-fMRI data in which abnormalities are not apparent on scalp EEG.     

Towards natural stimulation in fMRI--issues of data analysis

In search for suitable tools to study brain activation in natural environments, where the stimuli are multimodal, poorly predictable and irregularly varying, we collected functional magnetic resonance imaging data from 6 subjects during a continuous 8-min stimulus sequence that comprised auditory (speech or tone pips), visual (video clips dominated by faces, hands, or buildings), and tactile finger stimuli in blocks of 6-33 s. Results obtained by independent component analysis (ICA) and general-linear-model-based analysis (GLM) were compared. ICA separated in the superior temporal gyrus one independent component (IC) that reacted to all auditory stimuli and in the superior temporal sulcus another IC responding only to speech. Several distinct and rather symmetric vision-sensitive ICs were found in the posterior brain. An IC in the V5/MT region reacted to videos depicting faces or hands, whereas ICs in the V1/V2 region reacted to all video clips, including buildings. The corresponding GLM-derived activations in the auditory and early visual cortices comprised sub-areas of the ICA-revealed activations. ICA separated a prominent IC in the primary somatosensory cortex whereas the GLM-based analysis failed to show any touch-related activation. "Intrinsic" components, unrelated to the stimuli but spatially consistent across subjects, were discerned as well. The individual time courses were highly consistent in sensory projection cortices and more variable elsewhere. The ability to differentiate functionally meaningful composites of activated brain areas and to straightforwardly reveal their temporal dynamics renders ICA a sensitive tool to study brain responses to complex natural stimuli.     

Synchronization likelihood with explicit time-frequency priors

Cognitive processing requires integration of information processed simultaneously in spatially distinct areas of the brain. The influence that two brain areas exert on each others activity is usually governed by an unknown function, which is likely to have nonlinear terms. If the functional relationship between activities in different areas is dominated by the nonlinear terms, linear measures of correlation may not detect the statistical interdependency satisfactorily. Therefore, algorithms for detecting nonlinear dependencies may prove invaluable for characterizing the functional coupling in certain neuronal systems, conditions or pathologies. Synchronization likelihood (SL) is a method based on the concept of generalized synchronization and detects nonlinear and linear dependencies between two signals (Stam, C.J., van Dijk, B.W., 2002. Synchronization likelihood: An unbiased measure of generalized synchronization in multivariate data sets. Physica D, 163: 236-241.). SL relies on the detection of simultaneously occurring patterns, which can be complex and widely different in the two signals. Clinical studies applying SL to electro- or magnetoencephalography (EEG/MEG) signals have shown promising results. In previous implementations of the algorithm, however, a number of parameters have lacked a rigorous definition with respect to the time-frequency characteristics of the underlying physiological processes. Here we introduce a rationale for choosing these parameters as a function of the time-frequency content of the patterns of interest. The number of parameters that can be arbitrarily chosen by the user of the SL algorithm is thereby decreased from six to two. Empirical evidence for the advantages of our proposal is given by an application to EEG data of an epileptic seizure and simulations of two unidirectionally coupled Hénon systems.