IEEE Signal Processing - May 2018 - 91

Reference Network

Identified Network
Network Similarity?
+

Simulated Epileptic Spikes
250 ms
Forward
Problem

Functional
Connectivity and
Network Analysis

Source 1

Reconstructed Brain Sources

Simulated Scalp EEG

~1,500
Regional Time
Series

Source n

256 Channels

Inverse
Problem

Time (s)

FIGURE 5. A model-based evaluation of EEG source connectivity methods aimed at identifying epileptogenic networks from scalp recordings. First, a spatially distributed epileptogenic network is generated by a physiological model (coupled neural masses generating epileptic spikes). This network is considered to be the ground truth. By solving the forward problem, synthetic dense-EEG data are generated. These simulated signals are then used to evaluate
the performance of EEG source connectivity methods according to their ability to recover the reference network. Different combinations of methods were
used to solve the inverse problem and reconstruct the dynamics of the cortical sources. For each combination, the identified network is compared with the
original network using a similarity index accounting for topological features (the 3-D position of nodes and edges) of matched networks.

the connectivity patterns and the neuropsychological results
obtained in patients.
Here, we present a model-based evaluation of EEG source
connectivity methods aimed at identifying epileptogenic networks from scalp recordings (Figure 5). We performed a joint
comparison of two inverse solution algorithms [wMNE and
dynamic statistical parametric mapping (dSPM)] and two connectivity measures (PLV and r 2 ) using data simulated from a
biophysical/physiological model that allows for the generation
at the cortical level of realistic interictal epileptic spikes that
also reflect in scalp EEG signals. We used a network-based similarity index to compare the network identified by each inverse/
connectivity combination with the original network simulated
in the model. The main advantage of this algorithm, called
SimiNet, is that it takes into consideration the physical locations
of the nodes to compute network similarity, which is a crucial
element when dealing with brain networks. The nodes shown
in Figure 6(b) represent the physical locations of the generated
sources, while Figure 6(a) represents the nodes with the 5%
highest strength values (the most important nodes in the network). Edges are not shown to enhance visualization.
Globally, the results revealed that the choice of the inverse/
connectivity combination can have a significant impact on
the networks identified from scalp EEG signals [Figure 6(a)].

They also showed that methods based on PS (e.g., PLV) combined with the wMNE inverse algorithm show better performance in terms of similarity between the reference network
and identified network as compared with other combinations
[Figure 6(c)]. Other methods and other network scenarios were
tested in [38]. Interestingly, the same combination exhibited
the best performance. Finally, applying this combination on
real dense (256 channels) EEG data recorded in epileptic candidates for surgery showed excellent matching between scalp
EEG-based networks and intracerebral EEG-based networks,
as reported in [38].

Neurodegenerative diseases
Neurodegenerative diseases are associated with distinct
patterns of functional network dysfunction [37]. The main
motivation for using EEG source connectivity here is to find
an association between the degree of cognitive deficit on the
one hand and the alterations in the functional brain networks
on the other. The hypothesis is that cognitive impairment
gradually worsens with the progressive alteration of brain
functional connectivity. Besides its utilization with neurodegenerative diseases, EEG source connectivity was also used
in other applications, such as with schizophrenia [44], major
depression [45], pain [46], and obsessive compulsive disorder

IEEE Signal Processing Magazine

May 2018

Table of Contents for the Digital Edition of IEEE Signal Processing - May 2018