5 EEG Source Imaging and Multimodal Neuroimaging
115
arately, which was then linked using an Empirical Bayesian model [46]. This, in
essence, performs both of the previous asymmetric methods and uses a Bayesian
approach to fuse them as a newer combination. Alternative methods have extracted
EEG and fMRI features (BOLD and ERP peak latency and amplitude, BOLD percent signal change, RMS measure, etc.), establishing the probability distribution of
each and using these to determine the information shared by the components [63].
Others have created spatially adaptive priors for use within Bayesian frameworks,
developed by implementing measures of Total Variation [53]. Finally, some have
sought to introduce measures of graph theory or connectivity into the framework.
An early method for this performed functional network analysis on the fMRI and
EEG signals using spatial ICA and Granger Causality, again linking these within an
Empirical Bayesian Framework [47]. The important message to derive from all of
these is not just the individual approaches, but also the numerous ways that they can
be combined, adjusted, or refined to improve on current technology. Though much
has been accomplished, it is certain that many more methods will be developed with
the potential to both advance multimodal imaging and translate across fields.
Model-Based Approaches
Model-based approaches, on the other hand, are founded upon the development of
realistic models of the biological and physical factors that give rise to the detected
BOLD and EEG signals [71]. In general, this means that data-driven approaches are
simpler and better suited to naïve cases while model-driven approaches are more
conceptually complete and informative, albeit with an increased computational cost.
It also means that model-based approaches require a somewhat deeper understanding
of the factors connecting EEG to fMRI. For example, we have established that neurovascular coupling as the underlying mechanism that connects the EEG and fMRI
modalities. This coupling is not constant, however—decoupling between the BOLD
and EEG signals can be informative depending on the situation. Decoupling has been
observed in a variety of specific cases, including decoupling in the frontal lobe during locomotion and in cases of cerebral amyloid angiopathy [67] or seizure [75]. An
ideal model will also need to deal with the dynamics of whole neuronal populations
and once again address how these dynamics are represented at the scalp level. These
populations may show complex activity, including both inhibitory and excitatory
interactions within context of a population firing pattern that may be conditionally
rhythmic or transient. True neurophysiologic models are then difficult to generate;
the core principles for them are highly complex and any errors will be amplified as
the model is built outward from the cell populations and their properties [71]. Thankfully, our current discussion has provided a firm basis in EEG source localization and
reconstruction, which accounts for one major aspect of neurogenerative modelling.
The general pattern of the forward model and inverse problem, as explored above,
reappears here. Forward models will serve to represent the processes that generate
EEG and fMRI data while the inverse calculation will identify the model conditions
responsible for observed data [42]. Early single and double columnar models of
neuronal population activity were generated following a biological representation
of excitatory pyramidal cells modulated by inhibitory interneurons and excitatory
115
arately, which was then linked using an Empirical Bayesian model [46]. This, in
essence, performs both of the previous asymmetric methods and uses a Bayesian
approach to fuse them as a newer combination. Alternative methods have extracted
EEG and fMRI features (BOLD and ERP peak latency and amplitude, BOLD percent signal change, RMS measure, etc.), establishing the probability distribution of
each and using these to determine the information shared by the components [63].
Others have created spatially adaptive priors for use within Bayesian frameworks,
developed by implementing measures of Total Variation [53]. Finally, some have
sought to introduce measures of graph theory or connectivity into the framework.
An early method for this performed functional network analysis on the fMRI and
EEG signals using spatial ICA and Granger Causality, again linking these within an
Empirical Bayesian Framework [47]. The important message to derive from all of
these is not just the individual approaches, but also the numerous ways that they can
be combined, adjusted, or refined to improve on current technology. Though much
has been accomplished, it is certain that many more methods will be developed with
the potential to both advance multimodal imaging and translate across fields.
Model-Based Approaches
Model-based approaches, on the other hand, are founded upon the development of
realistic models of the biological and physical factors that give rise to the detected
BOLD and EEG signals [71]. In general, this means that data-driven approaches are
simpler and better suited to naïve cases while model-driven approaches are more
conceptually complete and informative, albeit with an increased computational cost.
It also means that model-based approaches require a somewhat deeper understanding
of the factors connecting EEG to fMRI. For example, we have established that neurovascular coupling as the underlying mechanism that connects the EEG and fMRI
modalities. This coupling is not constant, however—decoupling between the BOLD
and EEG signals can be informative depending on the situation. Decoupling has been
observed in a variety of specific cases, including decoupling in the frontal lobe during locomotion and in cases of cerebral amyloid angiopathy [67] or seizure [75]. An
ideal model will also need to deal with the dynamics of whole neuronal populations
and once again address how these dynamics are represented at the scalp level. These
populations may show complex activity, including both inhibitory and excitatory
interactions within context of a population firing pattern that may be conditionally
rhythmic or transient. True neurophysiologic models are then difficult to generate;
the core principles for them are highly complex and any errors will be amplified as
the model is built outward from the cell populations and their properties [71]. Thankfully, our current discussion has provided a firm basis in EEG source localization and
reconstruction, which accounts for one major aspect of neurogenerative modelling.
The general pattern of the forward model and inverse problem, as explored above,
reappears here. Forward models will serve to represent the processes that generate
EEG and fMRI data while the inverse calculation will identify the model conditions
responsible for observed data [42]. Early single and double columnar models of
neuronal population activity were generated following a biological representation
of excitatory pyramidal cells modulated by inhibitory interneurons and excitatory