Universiteit van Amsterdam
Methods for Non-invasive Cortico-cortical
Interactions: Recent Developments
Bachelor thesis
Student: Carmen Wolvius, 10208399
29-6-2014
Abstract
2.
Introduction
3.
1.1: Functional Imaging Techniques
4.
1.2 Methods for Data Analysis of Functional Imaging Techniques 5.
2.1 Recently Developed Methods and Solutions
9.
2.2 Remaining Problems
12.
Discussion
13.
Abstract
The field of neural imaging and recording techniques is growing rapidly and is widely discussed. This literature study addresses issues of M/EEG and fMRI/PET. Methods to reduce the effect of the inverse problem in M/EEG, the reference problem in EEG, and the lack of temporal resolution in fMRI/PET signals, are outlined. First regularly used methods are discussed, then methods that are recently developed or that recently received more attention are discussed, to see how these problem are handled. The issues seem to receive a reasonable amount of attention and researchers seems to be heading towards a way to get around some of the issues reasonably well. However, in some fields the issues seem underexposed or are not discussed in general. Therefore, it might be essential to indicate how these issues are handled in the flied in general, including applied research.
Introduction
In the last few decades there has been a large increase in our knowledge of the human brain. The interest in the processes that underlie our thoughts, produce our actions and link our perceptions to memory, have been nourished by multiple neural imaging techniques that seem to observe these processes. Brain activity can be measured and images of neuronal connections can be made. The development of neural imaging
techniques has been a rapidly growing process. Dr. Hans Berger first recorded human EEG in 1924. His insights at the time still apply today. For instance, he discovered the separate bands in electrical recordings and called them ‘alpha’ and ‘beta’, and his use of
30mm/second paper speed became standard for decades, until digitized computer
recordings took over. Scientists researching neural recordings have been prone to use the newest developments in technology from the first hour, which has always been a benefit (Collura, 1993).
In this field of research there is a distinction in cortical organization: functional segregation and functional integration. Where functional segregation focuses on the location of specific features in the brain, functional integration looks at the connections that may exists. Functional segregation emphasizes that a specific location in the brain has a specific function and tries to map these with different tasks. Functional integration builds on functional segregation, and implies that the specific functions of the brain depend on the connections between the functionally segregation areas. Were research has produced an abundance of examples of functional segregation, the assessment of
functional integration has proven more difficult. There are two main approaches to studying functional integration: those focussing on what has been called functional
connectivity, and those that have been called effective connectivity. The term functional connectivity refers to the basic correlation observed between measures of neural activity. The term effective connectivity on the other hand, is defined as a measure of the
influence of one neuron or neural assembly on the activity of another. Because in effective connectivity an influence is perceived, one could try to assess causal relationships between the neuronal connections (Friston, 2011).
The great enthusiasm of researchers and the rapid growth of research techniques have led to a burst of research based on recordings of brain activity or brain mapping. The fact that this area of research has made such progress, might be a reason to take a step back and look at some issues that might need reconsideration. Grasman et al. (2004) described the research methods used in the last decade and outlined all the problems that
arose using those methods. The purpose of this literature study is to assess the methods and the pitfalls of these methods currently used to address the prevailing questions in functional integration research (see Grasman et al., 2004, for an earlier assessment).
In the following, we will first give a short introduction to currently widely used functional brain imaging techniques, and then outline methods that are in common use to answer questions on functional integration. Subsequently, we discuss newly developed methods in the light of their interpretational problems. Finally, we discuss problems that seem to be solved and problems that remain an issue interpreting findings from functional integration studies.
1.1: Functional brain imaging techniques
There are several techniques that are used to record neural activity in the brain, of which we will briefly outline the most used the techniques. Electroencephalography (EEG) records electrical activity of the brain through electrodes that are attached to the scalp. The average electrical activity for the population of neurons that are situated under the electrode is recorded. This activity can result from spontaneous brain activity or activity evoked by a stimulus. Magneto encephalography (MEG) measures the magnetic fields that are generated by the electrical currents. The main advantage of EEG and MEG is the high temporal resolution: The signals recorded with both techniques are a direct reflection of neuronal electric activity, albeit an aggregate of many tens of thousands of neurons in the brain. The main limitation to both MEG and EEG is that only the surface of the scalp can be used to measure activity. Since there is activity in the whole brain volume, it is never completely clear where the signal comes from. Faint signals, for example can be faint signals that are close by or strong signals further from the surface (Kalat, 2009) .
Functional magnetic resonance imaging (fMRI) and Positron-emission tomography (PET) are neural imaging techniques. FMRI is a modified version of an MRI scan. A powerful magnetic field is applied in the scan, which allows for an image of the brain based on the haemoglobin in the blood. The characteristics of haemoglobin makes it suitable for fMRI because, oxygen-rich haemoglobin reacts differently to the magnetic field than oxygen-low haemoglobin does. An increase of neural activity in the brain is associated with an increase of oxygen in blood, so the fMRI detects changes in oxygen level in order to detect changes in activity.
Positron emission tomography (PET) also produces a high resolution image of brain activity, by recording emission of radioactivity in the brain. The radioactivity is induced by
injecting a substance into the body, glucose for example, that is radioactively labelled. The decay of a radioactive atom releases a positron, which reacts almost immediately with an electron. This produces gamma rays in exactly opposite directions. These gamma rays are recorded by the PET detectors and can be used to construct an image (Kalat, 2009).
The problem with both fMRI and PET scans can be seen as opposite to the main problem of M/EEG. Where M/EEG has a very high temporal resolution, but has issues creating a spatial image, PET/fMRI has high spatial resolution, but has a substantial problem with temporal resolution. Both fMRI and PET detect physiological changes: the course of oxygen-rich blood and the decay of radioactive atoms. These processes are an indication of neural activity, but they take ten seconds or more come about. Neural activity processes on the other hand, might only last for few milliseconds. Neural activity of fast cognitive processes cannot be indicated, only a ‘snap shot’ image can be made. The detection of activity with M/EEG takes only milliseconds, which corresponds to the actual time it takes neuronal assemblies to communicate (Kalat, 2009).
To summarize, currently two different techniques to look at neuronal connectivity are available, one with high temporal resolution and one with high spatial resolution, but none that contain both. Because of the technical structure of both techniques it is
impossible to completely solve these limitations, but researchers are looking for different ways to work around them. In the next section first the limitations of methods used in M/EEG analysis will be discussed and outlined with examples, than limitations of fMRI/PET will be addressed.
1.2 Methods for Data Analysis of Functional Imaging Techniques
Cross-correlation is a standard method used to assess a relationship between two series of data. In M/EEG these series consist of electrical or magnetic data picked up by two different sensors. The cross-correlation method uses the amplitudes of two signals at the same time interval and measures the covariance between these amplitudes. The co-existence of the amplitude is thought to indicate communication between neuralassemblies, which is based on the assumption that different brain areas hardly share the same amplitude by accident. When neural assemblies are thought to communicate though, it makes sense that one assembly reacts to the other. This implies that there is a cause and effect relationship between two assemblies, rather than a correlation between the assemblies. Because causes precede their effects, correlation is also assessed between different lags of time: a lagged correlation. When the amplitude of a neural assembly corresponds to the amplitude of another assembly a few milliseconds later, this may
suggest a cause-effect relationship. This can be seen as a phase difference, the amplitude is the same but the phase of the amplitude starts a few milliseconds later, (see figure 1). Since a direct relationship is observed in lagged correlations, it is suggested that lagged correlations are an implication of effective connectivity (Grasman et al. 2004).
Figure 1: The properties of a cause-effect relation have been suggested to indicate causal relations because they can be used to make a prediction. An example of an analysis that uses time series to predict causality is called
Granger causality, Granger causality however does not imply a direct causal influence (e.g., Friston, 2009). Another technique that is used to establish neuronal communication between different parts of the brain is coherence. The coherence function is related to cross-correlation and measures a cross-correlation between two signals in a frequency domain by means of the cross spectrum. Essentially, coherence is a correlation between the phases of signals instead of the correlation between amplitudes of the signals (Grasman et al., 2004). Where correlation and coherence can be interpreted almost equally well in M/EEG, coherence is not suitable for fMRI or PET scans. Coherence depends on the frequency domain of the signals, and the imaging techniques are not fast enough to indicate these domains. Time domain cross-correlation seems to be suitable for fMRI and PET. Then different fMRI time series are used to assess correlations.
There are several problems with the interpretation of (lagged) correlation and coherence in terms of neural communication, due to inherent characteristics of the relation between EEG and MEG signals and the underlying neural activity.
The first problem that arises when discussing methods used in M/EEG research such as correlation and coherence, is the problem that was briefly discussed in the introduction of M/EEG: The measurement of M/EEG is located only on the surface of the scalp, while there is activity in the whole brain volume. So, it is never completely clear where the signal comes from. This makes it difficult to indicate, given certain M/EEG measurements, what the electrical activity in the brain is. This is called the inverse problem and it
complicates making assumptions about the underlying connections based on the data observed. Sensors have overlap in sensitivity to signals that originate from multiple different locations. Winter et al. (2007) defined the inverse problem for EEG en MEG separately and compared the influence it had. In EEG the problem is mainly referred to as volume conduction problem which refers to the conduction characteristics of the brain.
Because the brain is highly electrically conductive, the signal of a small source in the head will be picked up by multiple sensors. In MEG it is a part of the field spread problem. Field spread is a little different from volume conduction, but overall it has the same
confounding effect. This difference is due to the different properties of EEG and MEG scans. Volume conduction and field spread have both been found to inflate or decrease coherence, but Winter et al. specified these findings. For instance they found that when channel separations are very large, the field-spread effect on MEG coherence appears smaller than the volume conduction effect on EEG coherence.
A second problem that specifically concerns EEG measurement comes from the fact that EEG measures electrical potential differences between different points on the scalp. EEG recordings require a reference point in order to indicate these potential differences. The other sensors are compared to the reference sensor, which remains the same over time, and the actual activity in the measuring sensor is used as data for, for example, correlational research. The problem is that finding a good reference point is difficult. The reference points that are commonly used seem to be too close to the original sources, so they might be influenced by sources of activity in the head. This could lead to a
correlation with the potentials measured by the sensors on the scalp and the reference point (Hagemann et al., 2001). Multiple methods have been suggested, but most are far from ideal. For example, a part of the head far from the signals recorded was used as a reference, such as the nose, or an average of all the signals the sensors picked up has been applied in different studies. The average reference, which can be computed by
re-referencing all EEG signals to the average of all electrodes, is not a well weighted average because it does not take the whole brain volume into account (Grasman et al. 2004). Different methods have been developed to analyse functional or effective connectivity with M/EEG data that suffer from these limitations or that are subject to other limitations. A few of them will be discussed below.
Phase-locking analysis tries to determine a comparison in phases of different signals in M/EEG. At particular frequencies, the phase differences are compared to see if they are constant (Grasman et al. 2004). When phases stay constant there is a phase relation between signals, this implies a high degree of coherence, which is suggestive of functional connectivity. To indicate these phases of the signals to imply coherence high temporal resolution is necessary, so phase-locking cannot be used for fMRI data.
Principle Component Analysis (PCA) is a method that is used for data analysis is many different ways. It is used to find a smaller set of variables in a set multivariate measurements in order to get a good explanation for the measurements that are found
(Bugli & Lambert, 2006). The method is used in M/EEG as well, but it was described in Grasman et al. (2004) that it is not suited for M/EEG data because the lead field, which influences the correlation pattern, influences the component structure in PCA (Grasman et al., 2004). It is used with fMRI/PET data as well, so it will be discussed later in this
section.
Event-Related Synchronization (ERS) and Event-Related De-synchronization (ERD) describe the short and regional localized amplitude increases or decreases in the alpha or beta band that occur after an event, such as opening the eyes. These increases and decreases can be described as a burst or a blocking in the frequency bands. It is thought that these changes are due to a decrease or increase in synchrony of the underlying neuronal populations (Pfurtscheller & Lopez da Silva, 1999; Grasman et al., 2004). That is why ERS and ERD are used to assess functional coupling and functional segregation, but it is used for different purposes than the other methods. ERS/ERD also suffers from the limitations related to M/EEG that are discussed above.
A model that is used for time series in EEG analysis is the vector autoregressive model (VAR). VAR models can be used to assess the lead-lag patterns discussed above and are parameter based. Another parameters based model is a Dynamic Factor Model (DFM). The latent processes in DFM may be indicative of functional connectivity or effective connectivity. The different processes are thought to be highly dense connected factors among a large scale network, which would indicate functional connectivity. But lead-lag patters can also be observed, which indicate effective connectivity (Grasman, 2004). Both
these models are however subject to the inverse problem of M/EEG discussed above.
A way to go around the inverse problem and the reference problem at once is by estimating a source in the brain based on the signals that are recorded. By estimating the activity at one location, there does not seem to be any influence of the fact that the measurement is solely on the scalp because the location is already known. The method has been used to determine one source or two sources, where correlation between the two can be assessed (Cuffin, 1998). However, when one location is determined, the activity
measured in that location can still originate from other locations. So the activity in one location can still be influenced by activity in other locations. When two sources are determined to indicate a correlation, they are in fact likely to influence each other. That way, when a correlation is found, it is indistinguishable whether the correlation is due to underlying connectivity or because the sources influence each other’s measurements (Cuffin, 1998).
Functional MRI and PET have their own sets of interpretational problems when used to establish functional connectivity relations between different brain sites. The main problem with fMRI/PET discussed briefly before, relates to the lack of temporal resolution. This implies that at least some form of communication in the brain cannot be recorded. However, some fMRI/PET methods have been developed to work around the mechanical limitations. Especially in fMRI, fMRI time series are used to assess correlations for
functional and effective connectivity. In functional connectivity correlations between the zero-lags of fMRI time series of different brain regions are computed (Despande et al. 2010). For example, a correlation is assessed between individual BOLD time points during resting conditions (Friston, 2009). With the fMRI time series, effective connectivity can be assessed as well. In effective connectivity analysis Granger causality is used. Granger Causality (GC) tries to infer a causal and directional influence of variable A to variable B. Unfortunately this method is not without its problems either. GC uses time series to assess a lagged correlation. This way it addresses different information than functional
connectivity, but it is likely that zero-lag relationships influences the time-lagged relationships. It has been suggested that this contaminates the causal correlation of GC. Deshpande, Sathian and Hu (2010) formulated it as: ‘zero-lag correlation “leaks” into estimates of time-lagged causality’. Functional connectivity reveals different information about the underlying networks, than effective connectivity does. Therefore, it follows that zero-lag correlations changes the interpretation of effective connectivity (Deshpande et al., 2010). Another issue with fMRI is that anatomic connections tend to correlate strongly with the patterns found in the analyses. Strong resemblance has been found between traditional anatomic structures and well-known brain systems (Buckner, 2010), and functional connectivity has been found where there are no anatomic connections, for example between the right and left primary visual cortex. When anatomic connections influence functional connectivity, functional connectivity is not merely a reflection of structural connections. So, it has been suggested that these anatomic correlations bias functional connectivity (Koene et al. 2010). Below, other methods that try to assess functional or effective connectivity with fMRI/PET are discussed.
Principal component analysis (PCA) has been used to established functional networks in fMRI by isolating functional patters in the functional imaging data (Mckeown, 1998). First PCA tries to indicate the tendency of the signals to covary and then finds the patterns that capture the largest source of variance. In fMRI analysis those components should be handled with care because the assumptions of PCA require the components to be uncorrelated. The components represent functional networks, which in theory are
connection is required between these networks (Grasman et al. 2004). So the uncorrelated components would be somewhat contradictory in this context.
A method that has been used to assess effective connectivity in fMRI or PET data is path-analysis or Structural Equation Modelling (SEM) (Grasman et al. 2004). Using a priori knowledge, a model, based on multiple regressions, is made. The parameters for the model are estimated in a way to ensure that the gap between the predicted covariance matrix and the observed covariance matrix is smallest. A problem found in all SEM models is that comparison with other models is difficult, and left out of the analysis, while there might be a model with the same or an even better fit (Kline, 2011). In the next section
methods that have been introduced into this field recently will be discussed.
2.1 Recently Developed Methods and Solutions
As seen in the previous section, a lot of different methods have been suggested to overcome the problems that arise with the use of neural imaging or recording techniques. But as far as those methods go, the problems have been too persistent to completely solve them. Over the last decade researchers have continued developing methods like the ones discussed above. In this section a few of those methods will be discussed to assess whether any changes have been made in the methods used or if the common interest has shifted to other areas of solving these problems.
Stam et al. (2007) introduced a new method to address the volume conduction and problem with reference electrodes in the assessment of functional connectivity, namely the Phase Lag Index (PLI). PLI is based on the assumption that when a consistent phase lag between two time series, that is nonzero, is indicated it cannot be explained by volume conduction of one strong source. Therefore, it must reflect a real interactions of the underlying variables. The main aim of PLI is to get reliable estimates of phase
synchronization that are invariant to common sources. In the case of EEG these common sources might be active reference electrodes that interfere or a single node that causes volume conduction. In their article, Stam et al. compared PLI to the frequently used measure of phase coherence and found that PLI was influenced less by their simulated volume conductions than the current methods. They also found that PLI was less sensitive to differences in coupling strength. It was able to detect synchronization with moderate coupling strength where other models fail to do so. PLI’s high sensitivity to noise also leads to less influence of reference points, but the researchers point out that PLI is not immune to volume conduction or influence of poor reference points. Peraza et al. (2012)
and also found that PLI is partially invariant to volume conduction. Partially, because when working with multiple sources, networks at electoral level were significantly different from networks found at source level, which indicates PLI estimation is influenced by volume conductions when multiple sources are used. Vinck et al. (2011) also indicated some issues with PLI. They posed that the methods sensitivity to noise and volume conduction might be influenced by the inconsistency of the index, because small perturbations have been known to turn phase lags into lead (and leads into lags) (Vinck et al. 2011). They introduced a Weighted Phase Lag Index (WPLI), which, compared to PLI, is said to be influenced less by additional noise sources. The WPLI weights the contributions of the phase leads and lags, which makes it as sensitive to noise as PLI but less prone to making type I or type II errors: unfairly rejecting or retaining the null hypothesis of detecting synchronization. Lau et al. (2012) tested WPLI the way Vinck et al. had described its use. They tested the ability of WPLI to detect the potentials of the cognitive response to walking. They found that WPLI detects the original cognitive reaction to walking, but the activity patterns found were more complex and variable than recorded by the traditional method. Though the fact that WPLI is more variable makes comparison across time or across subjects very difficult. Figure 2 shows the intervals of subject 1, and it can be seen that the intervals would be hard to compare. So WPLI might not be suitable for comparison across time or subjects, which means it is unsuitable recording for stimulated activity.
Figure 2:
A method that recently received a lot of attention represent a somewhat different way of trying to identify brain activity. This philosophy is based on the analogy of small and dense connected nodes in the brain that are connected with a few long connections, to make the transport of
information as efficient as possible. This theory has led researches to search for these nodes, they call ‘hubs’: small collections of neural assemblies that are very densely connected among each other, but have almost no connections outside of the hub. Fair et al. (2009) investigated these
networks in children and adults, under the assumption that these networks form during our childhood and adolescence. He found that brain regions of children communicate more locally with other brain regions and the connections seem to be more anatomical, but that communication becomes more distributed with age. Which suggest that maturation of the
brain makes some changes in the arrangement of the neurons, and these changes seem to lead to a network as expected from the network theory. Methods used for investigating these networks have been applied to mostly fMRI data, like Fair et al., but also to M/EEG data (Basset & Bullmore, 2010). A method that is used for the search for hubs in fMRI is Functional Connectivity Density Mapping (FCDM). Tomasi et al. (2010) gave a description of this very efficient method and praise it, for computational requirements of comprehensive analytical methods had been interfering with the search for neural networks. The ultrafast technique would give a functional connectivity map with very high spatial resolution, by using data from over 900 subjects. Buckner (2010), however, points out some issues with this approach. First he points out that with fMRI there is always the issue of anatomical connections interfering with the observed patterns of functional connections, and because anatomic connections might differ between subjects this might influence the findings negatively. Also this approach it would still not be possible to observe stimulated response initiated by a task because of the slow BOLD response.
In the analysis of brain networks in M/EEG the high temporal resolution is used. Micheloyannis et al. (2009) compared EEG data of children and students during a math test. They looked at different frequency bands in order to detect differences in activity during the cognitive task. They found that the theta and gamma band showed higher synchronization in adults during the cognitive task, as a result of higher working memory capacity, and a reduction of the connections in grey matters seems to occur in adults. This points to higher efficiency in adults, which also corresponds to the expectations of the theory. Because this analysis is based on comparison, one might opt that when a comparison is made between two signals the problem of volume conduction has less effect than it normally would. However aside from volume conduction, Micheloyannis et al. used two reference sensors placed behind the ear. As was discussed in the previous chapter, this sensor is too close to the activity to be neutral. In this case a source located close to the sensor might have enhanced or reduced the results that were found.
Previously, it has been pointed out that it is difficult to address causal modelling with fMRI data. As discussed in Deshpande et al. (2010) there are serious confounds found in coherence analysis when fMRI time series are used to assess lagged correlations. Friston (2009) devoted an article to the topic of causal modelling and brain connectivity in fMRI. He addresses two approaches, he calls ‘state-of-art approaches for understanding the
communication among distributed brain systems using neuroimaging’. These methods are based
on the distinction between the functional and effective connectivity models mentioned earlier, but both appeal to causality. The fundamental difference between dynamic causal
modelling (DCM) and Granger causal modelling (GCM) lies in the different ambitions of the models. Where DCM, based on effective connectivity models, uses a forward or generative model to see how the observed data are caused, GCM focuses on temporal dependencies among the data, without any inferences of how the data was caused. In DCM a model is fitted by optimizing the parameters with the evidence of the model. This should lead to a model that has the most likely parameters and model evidence. Altogether, this would enable one to explore the different models possible and handle the data in a parsimonious way. In GCM two models are compared: one with a direct mapping of brain areas and one without. There is evidence for the mapping between the areas when the model with mapping fits better than the model without. One can see that DCM has a method that can infer further after analyses are done: GCM usually stops after choosing one model over the other because there is not a great deal of biophysical meaning to the parameters.
Something to consider might be that the brain has multiple reciprocal polysynaptic interactions, which would not suggest that a model without connections is an adequate null model (Friston, 2009). David et al. (2008) have outlined empirical foundations for the issues of GCM and DCM. Using a well-defined animal model with known neuronal activity, they were able to compare both model outcomes to the true model. The study outlines GCM is a useful approach for model comparing when brain states are observed directly, as with EEG or in their case when the underlying neural activity is known, which makes it unsuitable for model comparison in fMRI. DCM on the other hand, can include multiple models on its own without the known neural activity pattern, so DCM seems suitable for fMRI model comparison.
2.2 Remaining Problems
In the former chapters a few topics have been discussed. The different focus of neural connectivity research, functional and effective connectivity, has been outlined. The inverse problem was discussed, which originates from the measurement of M/EEG that is located only on the surface of the scalp, while there is activity in the whole brain volume. What makes it difficult to indicate, given certain M/EEG measurements, what the
electrical activity in the brain is. Then the EEG reference problem was discussed, which includes the point of reference that is needed to compare the recorded signals with. The points that are used regularly are too close to the signals measured, which makes them unsuitable as a reference. The main issue of fMRI and PET scans is the slow response of both techniques. They form a high spatial 3D image, but are unable to picture the course of a signal stimulated by a task, that only takes a few milliseconds. To summarize, there are techniques with high spatial resolution that are not able to follow the course of a
signal in matter of milliseconds and there are techniques with high temporal resolution that can never be sure where the signals they measure originated from. These structural issues have been addressed with the methods discussed in the second chapter, but none of the methods seem to give a working solution. In the last chapter methods are discussed that have recently been addressed in scientific journals. WPLI addresses the issues of M/EEG directly, by focussing on detecting the bias in the recorded signals in order to detect the real interactions of the underlying variables. The results of the articles that tested WPLI indicated that it was influenced less by volume conductions than methods currently used (Lau et al. 2012; Vinck et al. 2010). However, WPLI is still not insusceptible for volume conduction. The high sensitivity to noise brought along an additional problem. The measures over time, which can be used to indicate the course of a signal, differ in a way that made the comparison of the measures difficult.
Both EEG and fMRI were used to indicate brain networks, which led to interesting findings. Two articles were discussed were both techniques were used to assess
connectivity differences in children and students. The results indicated that students show fewer connectivity overall than children but they presumably have more efficient
connections, which could be explained by the theory. There is a lot of attention for these ideas at the moment, but the methods used do not seem to take the problems in account that are discussed in the literature study. However, the lag of assessing effective
connectivity with fMRI is a topic discussed quite recently. DCM tries to assess effective connectivity in fMRI in a way SEM does, by fitting a model to the observed data. By
addressing effective connectivity, inferences could be made about causal relations in fMRI data. However, these inferences will not go as far as indicating the course of a signal during a task, a decision making task for example, as the BOLD signal is still not fast enough to indicate interactions in matter of milliseconds.
Discussion
In the field of research that focusses on improving the ways in which we determine neural connectivity, the inverse problem and reference problem are still a topic that is reasonably well discussed. The topics of PLI and WPLI have had publications quite recently and there are different researchers testing and commenting on the methods. These
methods seem to be heading towards a solution and researches seem to keep improving them. Fields that focus on other subjects however, like the network approach, sometimes seem to forget about the problems discussed. Just as Micheloyannis et al., discussed above, who investigated differences in connectivity between children and students, but used the international 10/20 system which uses ear references that might not be as neutral
as reference points should be. In fMRI/PET research a lot of new methods have been introduced. The DCM model addresses the measurement of effective connectivity in fMRI and tries to fit the data to a model in a way that causal relations can be assessed. Even though results that arise from this kind of research is really interesting it does not account for the slow BOLD signal which makes it impossible to see the course of stimulated
responses. A field that has hardly been addressed in this
literature study is the current field that uses M/EEG of fMRI/PET for research purposes. The literature used in this study is mainly methodological. It would be interesting to see how the problems addressed in this paper, are handled by researches that use neural imaging techniques as a tool to assess other research topics. For example, when
differences in neural connectivity in syndromes like Attention Deficit Hyperactive Disorder are studied. Then, a better image of the current view towards these problems and the way they are handled can be made. It is important to see if findings about the confounds in neural imaging research are picked up in the applied research field, for conclusions that are made might be confounded as well.
Another point of discussion is that the field of research that
addresses these issues is large and probably not all methods used to solve these issues will be mentioned in this literature study. It seems that the issues still receive a lot of
attention, an important question to address next is to what extent are these
methodological findings used in the applied field of neuroimaging, and what does that say about the conclusions made in that field of research.
Literature
Buckner, R.L. (2010). Human functional connectivity: new tools, unresolved questions.
PNAS vol. 107, no. 24.
Basset, D. S. & Bullmore, E. T. (2009). Human Brain Networks in Health and Disease. Curr Opin Neurol. 22, 4, 340–347.
Bugli, C. & Lambert, P. (2006). Comparison between Principal Component Analysis and Independent component Analysis in Electroencephalograms Modelling. Biometrical Journal, 48, 1-6.
Collura, T. F. (1993). History and Evolution of Electroencephalographic Instruments and Techniques. Journal of Clinical Neurophysiology, 10, 476-504
Cuffin, B. N. (1998). EEG dipole source localization. IEEE Engineering in Medicine and Biology, 98. 118- 122.
David O, Guillemain I, Saillet S, Reyt S, Deransart C, et al. (2008) Identifying neural drivers with functional MRI: An electrophysiological validation. PLoS Biology, 6, 12,
2683-2697.
Despande, G., Sathian, K., Hu, X. (2010). Assessing and Compensating for Zero-lag Correlation Effects in Time-lagged Granger Causality Analysis of fMRI. IEEE Trans Biomed Eng, 57, 6, 1446–1456.
Fair, D. A., Cohen, A. L., Power, J. D., Dosenbach, N. U. F., Church, J. A., Miezin, F. M., Schlaggar, B. L. & Petersen, S. E. (2009). PLoS Computational Biology 5, 5.
Friston, K. J. (2011). Functional and Effective Connectivity: A Review. Brain Connectivity, vol. 1, nr. 1.
Friston, K. J. (2009). Causal Modelling and Brain Connectivity in Functional Magnetic Resonance Imaging. PloS Biology, 7, 220-225.
Grasman, R. P. P. P. (2004). Sensor array signal processing and the neuro-electromagnetic inverse problem in functional connectivity analysis of the brain (Doctoral dissertation). Retrieved
from: http://dare.uva.nl/document/15082.
Hagemann, D., Naumann, E. & Thayer, J. F. (2001). The Quest for the EEG reference revisited: A glance from brain asymmetry research. Psychophysiology, 38, 847-857.
Kalat, J. W. (2009). Biological Psychology. Belmont: Wadsworth, Cengage Learning.
Koene, R. A., Hedden, T., Venkataraman, A., Evans, K. C., Lazar, S. W. & Buckner, R. L. (2010). Intrinsic Functional Connectivity As a Tool For Human Connectomics: Theory, Properties, and Optimization. Journal of Neurophysiology, 103, 297-321.
Kline, R. P. (2011). Principles and Practice of Structural Equation Modeling (3rd ed.) New York: The Guilford Press.
Lau, T. M., Gwin, J. T., McDowell, K. G. & Ferris, D. P. (2012). Weighted phase lag index stability as an artefact resistant measure to detect cognitive EEG activity during
locomotion. Journal of Neuroengineering and Rehabilitation, 9, 47.
Mckeown, M. J., Makeig, S., Brown, G. G., Jung, T-P., Kindermann, S. S., Bell, A. J. & Sejnowski, T. J. (1998). Analysis of fMRI Data by Blind Separation into Independent Spatial Components. Human Brain Mapping, 6, 160-188.
Micheloyannis, S., Vourkas, M., Tsirka, V., Karakonstantaki, E., Kanatsouli, K. & Stam, C. J. (2009). The Influence of Ageing on Complex Brain Networks: A Graph Theoretical Analysis. Human Brain Mapping, 30, 200-208.
Peraza, L. R., Asghar, A. U. R., Green, G., Halliday, D. M. (2012). Volume conduction effects in brain network inference from electroencephalographic recordings using phase lag index. Journal of Neuroscience Methods, 207, 189-199.
Pfurthscheller, G., Lopez da Silva, F. H. (1999). Event-related EEG/MEG synchronization and desynchronizatoin: basic principles. Clinical Neurophysiology, 110, 1842-1857.
Stam, C. J., Nolte, G. & Daffertshofer, A. (2007). Phase Lag Index: Assessment of
Functional Connectivity From Multi Channel EEG and MEG With Diminished Bias From Common Sources. Human Brain Mapping, 28, 1178-1193.
Tomasi, D. & Volkow, N.D., 2010. Functional connectivity density mapping. PNAS, 107, 21.
9885-9890.
Vinck, M., Oostenveld, R., van Wingerden, M., Battaglia, F., Pennartz, C. M. A., 2011. An improved index of phase-synchronization for electrophysiological data in the presence of volume- conduction, noise and sample-siza bias. Neuroimage, 55, 1548-1565.
Winter, W. R., Nunez, P. L., Ding, J. & Srinivasan, R., 2007. Comparison of the effect of volume conduction on EEG coherence with the effect of field spread on MEG coherence. Statist. Med. 26, 3946–3957.
