You are on page 1of 9

Brain Multiphysics 2 (2021) 100021

Contents lists available at ScienceDirect

Brain Multiphysics
journal homepage: www.elsevier.com/locate/brain

Multivariate extension of phase synchronization improves the estimation of


region-to-region source space functional connectivity
Ricardo Bruña a,b,c,∗, Ernesto Pereda a,d
a
Laboratory for Cognitive and Computational Neuroscience, Canter for Biomedical Technology, Technical University of Madrid, Pozuelo de Alarcón, Madrid, Spain
b
Department of Experimental Psychology, School of Psychology, Complutense University of Madrid, Pozuelo de Alarcón, Madrid, Spain
c
Networking Research Center on Bioengineering, Biomaterials, and Nanomedicine (CIBER-BBN), Madrid, Spain
d
Electrical Engineering and Bioengineering Group, Departamento de Ingeniería Industrial, Instituto Universitario de Neurociencia (IUNE) e ITB, University of La Laguna,
San Cristobal de La Laguna, Tenerife, Spain

a r t i c l e i n f o a b s t r a c t

Keywords: The estimation of functional connectivity (FC) from noninvasive electrophysiological data recorded from sensors
Functional connectivity outside the skull requires transforming these data into a source space. As the number of sensors is much lower
Multivariate phase synchronization than the number of electrophysiological sources, the brain activity is usually parcellated into anatomical regions,
Source reconstruction
and the FC between each pair of regions is then estimated.
In this work, we generate a set of simulated scenarios with different configurations and coupling levels between
synthetic time series. Then, this simulated brain activity is converted into simulated MEG sensor-space data and
reconstructed back into the source space. Last, we estimated the FC between different regions using different
approaches commonly used in the literature and compared them with a novel approach.
Our results show that this novel approach, based on using all the information in each region, clearly outperforms
classical approaches based on a representative time series. The proposed approach is more sensitive to the level
of coupling and the extent of the area synchronized, and the resulting estimate better reflects the underlying FC.
Based on these results, we strongly discourage using a representative time series to summarize large brain areas’
activity when calculating FC.

Statement of significance

Functional connectivity is the current framework for understanding brain function. EEG and MEG are acquired
from outside the head, and when estimating source-level activity, the number of sources is much higher than
the number of sensors. As such, the source-level data is highly redundant, hampering the proper estimation of
functional connectivity. Using predefined brain areas instead of individual sources bypasses this problem, but
the leap from the former to the latter is far from trivial. Here we propose a novel approach that, according to
our simulations, correctly estimates the level of functional connectivity between predefined brain areas in several
scenarios. At the cost of a slight increase in computational burden, this solution can help unravel how the healthy
and pathological brain communicates.

1. Introduction resolution [3], allowing for the direct study of neural activations. It
is thus possible, under the hypothesis of phase synchronization (PS)
Electroencephalography (EEG) and magnetoencephalography [4], to look for brain areas whose patterns of neural activation and
(MEG) are two widespread techniques to measure electrophysiological deactivation are phase-locked, resulting in an estimate of functional
brain activity non-invasively [1]. While their spatial resolution is connectivity (FC) [5].
lower than other neuroimaging techniques such as functional mag- To go from the recording sensors to the brain generators, we need to
netic resonance [2], they provide a high millisecond-scale temporal perform some source reconstruction [6]. Initial attempts were based on


Corresponding author at: Department of Experimental Psychology, School of Psychology, Complutense University of Madrid, Pozuelo de Alarcón, Madrid, Spain
E-mail addresses: ricardo.bruna@ctb.upm.es (R. Bruña), eperdepa@ull.edu.es (E. Pereda).

https://doi.org/10.1016/j.brain.2021.100021
Received 30 September 2020; Received in revised form 17 December 2020; Accepted 8 January 2021
2666-5220/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

the hypothesis that only a reduced number of sources, termed dipoles,


were active at a given time [7]. The reason for this model has initially
been mathematical, as complex models had not yet been described, but
its popularity continued until the late nineties due to the limited compu-
tational capability available. Nowadays we use more complex models,
based on the hypothesis that multiple brain sources activate to produce
the signal measured at the sensors [8,9]. The source models typically
range from the thousands (for spatial filter approaches) to the tens of Fig. 1. Position of the simulated sources in the simulated brain. The simulated
thousands (for distributed sources models) sources. sources include: two clusters of sources (blue and red) in the left temporal pole;
The EEG and MEG electrodes arrays range from tens to a few hundred two clusters of sources (yellow and purple) in the right temporal pole; one cluster
sensors [1,10], and the resulting source-space activity is highly rank- of sources in the left occipital pole (green); and one cluster of sources in the right
occipital pole (cyan). The sources of either area that are not part of a cluster are
deficient. The estimation of FC over rank-deficient data generates many
displayed as smaller gray dots. TP: Temporal pole. OP: Occipital pole.
ghost interactions [11], decreasing the usefulness of this approach. Also,
the amount of information increases with the square of the number of
signals. As a result, the size of the estimated whole-brain FC matrices
In (1), the sub-index 𝑅 indicates the master (Rössler) oscillator, the sub-
is huge and redundant. Instead of using this approach, researchers usu-
index 𝑖 indicates the 𝑖-th slave (Lorenz) oscillator, 𝑎 is a scaling parame-
ally deal with FC estimations between pre-defined brain areas, resulting
ter determining the frequency of the Rössler oscillator, 𝑏𝑖 is a parameter
in smaller FC matrices. One common approach is to use an atlas with
determining the frequency of the 𝑖-th Lorenz oscillator, and 𝐶𝐿 is the
several regions comparable to the number of sensors [12–14], so the
coupling. For our simulations, 𝑎 was fixed to 0.293, 𝐶 was fixed to 0.19,
matrix’s redundancy is minimized.
and the values of 𝑏 for each of the three slave systems were set to 23,
There are several different approaches to go from the whole source
24, and 25.
model to the reduced FC. Most of the recent works use approaches based
To achieve synchronization between two different sets of sources,
on representative time series, where the whole activity of each area
we modified the definition of the Rössler oscillators, so they are pair-
is represented by one single time series [15–19]. However, these ap-
wise coupled. The equation defines this pair of synchronized Rössler
proaches may be oversimplifications of the problem and thus carry a loss
oscillators:
of information. Some authors have developed fully multivariate exten-
sions of well-known FC metrics, integrating all the information in both ⎧ 𝑥̇ = −1 ⋅ 𝑦 − 𝑧 + 𝐶 ⋅ (𝑦 − 𝑥 )
⎪ 𝑅1 𝑅1 𝑅1 𝑅 𝑅2 𝑅2
regions into a single FC estimate [20–22]. The use of these multivariate ⎨ 𝑦̇ 𝑅1 = 𝑥𝑅1 + 𝑎 ⋅ 𝑦𝑅(1 )
approaches has, for example, been proven to improve the accuracy of ⎪ 𝑧̇ 𝑅1 = 0.1 + 𝑧𝑅1 ⋅ 𝑥𝑅1 − 8.5

⎧ 𝑥̇ = −1 ⋅ 𝑦 − 𝑧 + 𝐶 ⋅ (𝑦 − 𝑥 )
the area-to-area FC estimation in functional magnetic resonance imag- (2)
ing [23]. Recently Basti and colleagues [24] provided a comprehensible ⎪ 𝑅2 𝑅2 𝑅2 𝑅 𝑅1 𝑅1
review of these methods, together with a small toolbox with examples. ⎨ 𝑦̇ 𝑅2 = 𝑥𝑅2 + 𝑎 ⋅ 𝑦𝑅(2 )
⎪ 𝑧̇ 𝑅2 = 0.1 + 𝑧𝑅2 ⋅ 𝑥𝑅2 − 8.5
However, the authors highlight the fact that no such extension exists, at ⎩
the moment, for PS methods.
In (2), 𝐶𝑅 indicates the coupling between the two Rössler oscillators,
In this work, we propose a multivariate extension of one of the
and ranges between 0 (completely independent clusters) and 0.18 (com-
most popular metrics for PS, namely the phase-locking value (PLV)
pletely synchronized clusters).
[25,26]. Then, we evaluate different approaches, among them our pro-
The coupling between the two Rössler oscillators is bidirectional and
posed method, to go from the whole set of source positions to a re-
symmetrical. Thus, the phase difference is zero. The coupling between
duced area-by-area FC matrix. We part from a set of simple simulations
the Rössler and the Lorenz oscillators is unidirectional, and the Lorenz
of phase-coupled signals in the brain, then we generate synthetic sensor
oscillators follow the Rössler one. As both sets of systems are identical,
measurements of these signals and, from them, we estimate FC using
the Lorenz oscillators from different clusters also present a zero-phase
a PS metric. Finally, we compare the estimated FC with the real con-
difference. We finally shifted the time series of the clusters by two sam-
nectivity introduced in the model. We hypothesize that our proposed
ples in order to avoid high correlations that would affect the quality of
method will outperform those based on representative time series.
the beamformer reconstruction [28].
2. Materials and methods
2.2. Simulated brain activity
2.1. Simulated signals
The head model consisted of a realistic single-shell delimitating the
To evaluate any FC estimator’s accuracy to describe a real interac- brain compartment extracted from the T1 MRI of a real subject, us-
tion, we need to know the actual strength of the interaction itself. This ing the unified segmentation algorithm [29] in SPM12. The MEG sen-
fact is only possible using simulations. In this study, we simulated dif- sors were simulated as the real positions from a recording of the same
ferent sets of synchronized brain areas (i.e., sources) coupled (or not) by subject, taken in an Elekta Vectorview system with 306 sensors (102
different levels of connectivity, and we estimated the signal that would magnetometers and 204 planar gradiometers). As the simulations are
be captured by an array of MEG sensors placed in realistic positions. interference-free, only the magnetometers were used.
Each independent brain activity was defined as a set of three nearby As a source model, we created a three-dimensional homogenous grid
sources. The overall activity was the product of a Rössler oscillator, and of sources, spaced 10 mm in MNI space, and labeled each source ac-
each of the three sources was a Lorenz oscillator highly coupled to it. cording to the automated anatomical labeling (AAL) atlas [12]. Only
The equation defining the system is [27]: those sources placed in cortical gray matter were taken into account. We
⎧ 𝑥̇ = −1 ⋅ 𝑦 − 𝑧 selected four areas of interest: the bilateral temporal poles, with nine
⎪ 𝑅 𝑅 𝑅
sources each, and the bilateral occipital poles, with six sources each.
⎨ 𝑦̇ 𝑅 = 𝑥𝑅 + 𝑎 ⋅ 𝑦𝑅( )
⎪ 𝑧̇ 𝑅 = 0.1 + 𝑧𝑅 ⋅ 𝑥𝑅 − 8.5 The position of the sources of interest is depicted in Fig. 1. We split the
⎩ (1) areas of interest in sub-areas of three continuous sources (marked as
⎧ 𝑥̇ = −1 ⋅ 𝑦 − 𝑧 colored, thicker points in the Figure) that will form a cluster and thus
⎪ 𝑅 𝑅 𝑅
⎨ 𝑦̇ 𝑅 = 𝑥𝑅 + 𝑎 ⋅ 𝑦𝑅( ) show similar behavior in the simulations (i.e., each of the three Lorenz
⎪ 𝑧̇ 𝑅 = 0.1 + 𝑧𝑅 ⋅ 𝑥𝑅 − 8.5 oscillators).

2
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

In (3), 𝑇 is the length of the signals in samples, 𝜑𝛼 is the instantaneous


phase of signal 𝛼, and 𝑖 is the imaginary unit. As PLV measures the
level of phase locking, 𝑥 and 𝑦 must be signals originated by oscillatory
processes, so the instantaneous phase can be meaningful described. To
achieve this, the simulated signals were filtered between 0.2𝜋 and 0.4𝜋
radians per second (see the Supplementary materials for details on how
this band was defined) using a 500th order FIR filter designed with Ham-
ming window and applied in a two-pass procedure. The instantaneous
phase was obtained using Hilbert’s analytical signal. The band-pass and
Hilbert’s filter were applied using 1000 samples of data at each side as
padding.
PLV is sensitive to zero-lag synchronization, making it extremely sen-
sitive to source leakage. In scenarios like the one simulated here (see 2.1.
Fig. 2. Different scenarios tested for multivariate phase synchronization. The Simulated signals section), the effects of source leakage can dominate
image on the left is the original scenario, presenting an inter-hemispherical con- over the real FC (see Supplementary Methods for details). To solve this
nection between left and right temporal poles. The other six images show differ- problem, instead of using PLV, we use its zero-lag-insensitive extension,
ent variations over the original scenario. A dot indicates the existence of an ac- corrected imaginary part of PLV (ciPLV) [31]. The mathematical defini-
tive cluster. A solid line connecting two dots indicates the presence of a coupling
tion for ciPLV is:
between the connected clusters. The blue solid line indicates the inter-temporal
link under evaluation. | {∑𝑇 −𝑖(𝜑 (𝑡)−𝜑 (𝑡)) }|
|ℑ
1 |
𝑡=1 𝑒
𝑥 𝑦
|𝑇 |
𝑐𝑖𝑃 𝐿𝑉 = √ | | (4)
( { )2
∑𝑇 −𝑖 𝜑𝑥 (𝑡)−𝜑𝑦 (𝑡)) }||
We transformed the grid to subject space using a linear transforma- | (
tion based on the T1 MRI and calculated the lead field at each source 1 − 𝑇1 ||ℜ 𝑡=1 𝑒 |
| |
position using a modified spherical solution [30]. Then, we simulated
several distributions of sources oriented according to each source’s max-
In (4), ℜ{𝛼} and ℑ{𝛼} stand for the real and imaginary part of 𝛼, re-
imal radiation direction to maximize the output power. This activity was
spectively. For the rest of this work, ciPLV will be the metric used to
reconstructed using a linearly constrained minimum variance (LCMV)
extract bivariate PS between pairs of signals.
beamformer [9] with no constrain in the dipoles’ direction. The solu-
tion for each source position was projected over its principal direction.
Last, we used different multivariate PLV approaches (see below) to cal-
culate multivariate PS and compared the results with the ground truth. 2.5. Approaches based on representative time series

2.3. Simulated scenarios


The most common approach to evaluate FC between pairs of large
brain areas is to take one time series and designate it to represent the
We tested seven different scenarios with different connectivity pat-
area’s overall activity. This representative time series can either be as-
terns, as depicted in Fig. 2. The active clusters of sources are represented
sociated with a specific source position or an assemble of different time
with a dot, and the functional couplings are represented with a line con-
series corresponding to different source positions.
necting the two coupled clusters. To avoid sparseness in the source solu-
When using a real-time series, common approaches use the source
tion, those sources not part of an active cluster were simulated as white
located in, or closer to, the center of mass of the area [16] or the source
noise (see Supplementary Methods for details).
with the time series more correlated to the rest of the area [19,32]. In the
The solid blue line in Fig. 2 indicates the inter-temporal connection,
first case, the method tries to comprise all the possible information into
and we will try to estimate the strength of this connection using differ-
a single time series. In the second one, the method tries to extract the
ent approaches. The coupling parameter was varied in ten increments,
dominant activity of the area. Other methods, like selecting the source
from 0 (no connection) to 0.18 (complete connection). Instead of di-
with the higher power, usually suffer from depth biases derived from
rectly using the coupling parameter for the rest of this work, we will
the source reconstruction method applied (e.g., external sources in min-
discuss the resulting connectivity, ranging from 0% to 100%. Each sce-
imum norm estimates [33] or deep sources in beamformers [34]).
nario and coupling were simulated ten times in order to evaluate the
The methods of using information from all the time series in the area
variability of each approach.
pursue similar goals. To combine all the area information into a single
time series, some authors use the average of all the source activities
2.4. Phase synchronization
in the area as representative time series [17,18]. When the goal is ob-
taining the dominant activity in the area, a popular choice is the use
Phase synchronization (PS) is a type of synchronization where the
of principal component analysis (PCA) to detect the weighted combina-
instantaneous phase of oscillatory processes is synchronized instead of
tion of source activities, which explains the largest possible amount of
their direct amplitude. The concept of PS was introduced by Rosenblum
variance in the whole area [35–37].
and colleagues in 1996 [4] and was quickly accepted as a model for
A third possible way to comprise all the area’s information into one
synchronization in the brain [25,26]. A critical point in favor of PS as a
time series is using the Kosambi-Hilbert torsion (KHT) [38]. This method
model for brain synchronization is that a phase lock can be established
is an extension of PCA optimized to extract the dominant phase in a set of
with minimal energy, while an amplitude lock requires a more consid-
noise measurements. Supposing the activity reconstructed in the area’s
erable amount.
sources is indeed a noisy measurement of the area’s real oscillatory ac-
One of the most intensively used metrics to study PS in the brain is
tivity, KHT extracts the underlying oscillatory activity with the higher
the phase-locking value (PLV) [26]. PLV quantifies the amount of syn-
signal-to-noise ratio (see Supplementary Methods for details).
chronization between two signals on a scale ranging from zero (absence
In this work, we will evaluate the performance of five methods based
of synchronization) to 1 (complete phase-locking). Mathematically, PLV
on representative time series: the use of the centroid of the area; the use
is calculated as:
of the average activity of all the sources in the area; the use of the source
|𝑇 ) ||
1 |∑ ( activity associated with the largest power; the use of the first component,
𝑃 𝐿𝑉 = || 𝑒−𝑖 𝜑𝑥 (𝑡)−𝜑𝑦 (𝑡) || (3)
𝑇 | 𝑡=1 | according to PCA; and the use of the first component, according to KHT.
| |

3
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

Fig. 3. Estimated multivariate PS for different cou-


plings in scenario 1. (a) Multivariate PS matrix be-
tween the four areas considered. The left panel
shows the theoretical model. The right panel shows
the multivariate PS matrix calculated using the
RMS approach for ten levels of coupling ranging
from 10% to 100%, in reading order. (b) Multivari-
ate PS between left and right temporal poles esti-
mated with every approach for ten levels of cou-
plings ranging from 0% to 100%. PS-AVG: Aver-
age of the bivariate PS matrix. PS-RMS: Root-mean-
square of the bivariate PS matrix. 1PCA: First com-
ponent, according to PCA. All the lines are rep-
resented as average across simulations (solid line)
and standard deviation (shadowed area).

2.6. Multivariate PS approaches In this work, we will evaluate the performance of two multivariate
methods: the average of all the bivariate PS values; and the root-mean-
A second family of approaches base on the algebraic properties of the square of all the bivariate PS values.
pairwise synchronization matrix. If one area contains 𝑀 sources, and the
other contains 𝑁 sources, all the relations between one source from one
3. Results
area and one source from the other can be expressed as an 𝑀 × 𝑁 matrix
of bivariate PS. Operations over this matrix have been used, for example,
As indicated in the Materials and methods section, we have evaluated
to extend the Phase Slope Index [39] into its multivariate counterpart
seven different scenarios with different connectivity profiles, each of
[22].
the scenarios with eleven different real connectivity levels. Each pair of
A straightforward approach to account for all the possible interac-
scenarios and the level of coupling was evaluated using seven different
tions between both areas is to get a single value equal to the average of
approaches. The results obtained using the KHT were identical to those
all the links connecting the areas. That is the average of the entries in
obtained using the PCA and are not displayed here (see Supplementary
the bivariate PS matrix [40,41]. We will term this approach PS-AVG.
Methods for a possible explanation). Here we will only report the most
In the last years, several authors have developed methods to esti-
significant results, but all seven scenarios are depicted, in full, in the
mate the overall intra-area PS from the symmetrical 𝐿 × 𝐿 bivariate
Supplementary Results.
PS matrix describing all the possible pairwise interactions in one area
The first scenario consisted of only two clusters of sources, located
with 𝐿 sources. Using cluster analysis over the PS matrix, it is possi-
bilaterally in the temporal poles and connected with different coupling
ble to identify the clusters of synchronized sources in the area and each
levels. The results for the six different approaches are depicted in Fig. 3.
source’s participation coefficient in each cluster [42]. A straightforward
When only the inter-hemispheric material connection is present, all the
approach to identify this cluster is the eigendecomposition of the matrix,
metrics behave adequately. Fig. 3(a) shows the resulting synchroniza-
where the eigenvectors indicate the participation coefficients of each
tion matrices for the RMS of the bivariate PS (PS-RMS), and Fig. 3(b)
source in each cluster, and the eigenvalues indicate the global variance
shows the inter-area synchronization between temporal lobes for each
explained by the cluster [43,44]. We can then define a global PS (GPS)
approach and coupling. In the bottom part of the Figure (and, in the
measure as the squared summation of the normalized eigenvalues:
same fashion, in the following Figures), the x-axis represents the cou-
√ √ pling level between the simulated signals, ranging from 0 to 100%, while
√ 𝐿 ( )2 √𝐿
√∑ 𝜆 1√∑
𝐺𝑃 𝑆 = √ 𝑖
= √ 𝜆𝑖 2 (5)
the y-axis represents the estimated FC between both temporal poles. We
𝑖=1
𝐿 𝐿 𝑖=1 would expect a straight line for an "ideal" metric, starting at a value
of zero and going up to the maximal PS value between the areas. This
In (5), 𝜆𝑖 is the 𝑖-th eigenvalue of the pairwise PS matrix. maximal value might be one or lower, depending on how we interpret
From this definition, GPS is equal to the normalized Frobenius norm the number of synchronized sources in the area. We will discuss this
of the pairwise PS matrix. As the Frobenius norm can be calculated for interpretation in the Discussion section.
rectangular nonsymmetrical matrices, it is possible to extend the GPS The increase of the measured synchronization with the coupling is al-
definition to get an index of inter-area PS: most linear in all cases. It is possible to observe three groups of metrics:
√ the average PS (PS-AVG), the RMS of the PS (PS-RMS), and the syn-
√ chronization between centroids increase up to synchronization values
√ 𝑁𝐴 𝑁𝐵
∑ ∑
√ 1
𝐺𝑃 𝑆𝐴−𝐵 =√ 𝑃 𝑆𝐵𝑗 −𝐴𝑖 2 (6) of around 0.3; the synchronization of the principal components (PCA)
𝑁𝐴 𝑁𝐵 𝑖=1 𝑗=1 increases up to around 0.5, and the synchronization between the aver-
age time series increases up to around 0.8. The synchronization of the
In (6), 𝑁𝐴 is the number of sources in area 𝐴, and 𝐴𝑖 is the 𝑖-th source sources with the higher power has a behavior similar to PCA, but its
of area 𝐴. As the result is equal to the root-mean-square (RMS) of all the variability is much higher, suggesting this approach might be unreli-
elements in the bivariate PS matrix, we will term this approach PS-RMS. able. Note that, while only the variability (represented as a shadowed

4
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

Fig. 4. Estimated multivariate PS for different cou-


plings in scenario 2. (a) Multivariate PS matrix be-
tween the four areas considered. The left panel
shows the theoretical model. The right panel shows
the-multivariate PS matrix calculated using the
PCA approach for ten levels of coupling ranging
from 10% to 100%. (b) Multivariate PS between
left and right temporal poles estimated with every
approach for ten levels of coupling ranging from
0% to 100%. PS-AVG: Average of the bivariate PS
matrix. PS-RMS: Root-mean-square of the bivari-
ate PS matrix. 1PCA: First component, according
to PCA. All the lines are represented as average
across simulations (solid line) and standard devi-
ation (shadowed area).

area) of this approach is visible to the naked eye, this area is represented the synchronization of the centroids drops when the right temporo-
for all the approaches, but its value is much smaller. occipital connection is active, implying that this approach could be un-
The second scenario includes secondary foci of activity in the tem- reliable in some scenarios.
poral poles but not connected to any other region of the brain. This Finally, Fig. 7 shows the value of inter-temporal synchronization
new model’s interest is that the temporal poles have sets of sources with in the seven analyzed scenarios for intermediate (50%) and complete
different activities, simulating heterogeneous brain regions. The results (100%) couplings. It is noteworthy that the approaches based on the
are depicted in Fig. 4, where two different issues arise. First, the repre- root-mean-squared and the average of the bivariate PLV matrix are sta-
sentative time series results using the source with higher power behave ble across the scenarios, while other approaches vary in their estima-
inconsistently and with high variability. We decided to this approach as tions. The approach based on the average time series of the area gives
unreliable from the first and second scenarios and excluded it from the very different results for the first scenario than for the rest.
analysis. Secondly, the approach based on PCA fails to detect the con-
nectivity between the two temporal lobes. As the activity at the tem-
4. Discussion
poral lobes is composed of two different clusters, the first PCA seems
to capture the non-connected activity (likely due to a higher power on
The current source reconstruction methods from EEG and MEG pro-
those sources, as their position is more in-depth into the brain [34]) and
vide detailed spatial information, at the scale of a few millimeters, of
discard the activity related to the coupled sources.
electromagnetic brain activity. Such sources, whose activities are highly
It is possible to avoid this problem by taking several components in-
redundant, are naturally grouped into regions of interest, most often us-
stead of only the first one. Fig. 5(b) shows the measured inter-temporal
ing anatomical landmarks. The reduction of the problem’s dimension-
synchronization using the first component, the RMS of the first two com-
ality is usually accomplished by taking a single time series per region,
ponents (2PCA), and the RMS of the components explaining up to 95%
for example, using an anatomical atlas. However, this popular approach
of the variance of the area (PCA95). In this case, the two first compo-
overlooks that these regions are very different in size, shape, or number
nents explain more than 95% of the brain area’s variance, so the second
of sources. Logically, this heterogeneity suggests that using the same
and third approaches give the same result. In a more general scenario,
method to generate the representative time series for each region (or
the use of a threshold in variance would be superior, as it does not re-
the very idea itself of reducing the dimensionality in this way) may not
quire any a priori knowledge of the number of signals of interest in each
be the best approach to, for example, estimate the FC between two re-
area. Fig. 5(c) depicts the inter-temporal synchronization as measured
gions. In this framework, multivariate FC estimators are a natural al-
using each approach for each level of coupling, after replacing the PCA
ternative as, by definition, they incorporate the information from all
approach for the PCA95 and removing the representative time series
sources within each region. Moreover, these approaches can be applied
with the largest amount of power. PCA95 and RMS behave closely to
reasonably, given the current computational capabilities of a standard
the first scenario, while the rest of the approaches present a drop in the
neuroimaging lab.
measured synchronization.
In this work, and using simulated data, we have analyzed the per-
The next scenarios include one or both temporo-occipital links.
formance of two of such multivariate approaches for different scenarios
Fig. 6(a) to Fig. 6(c) shows the inter-temporal synchronization with the
and coupling levels and then compared them with bivariate estimators
coupling in scenarios 3, 5, and 7. It is important to note that, to reduce
that calculate PS from different versions of representative time series.
the simulated data’s dimensionality, the value of the coupling for the
The simulation results for the basic scenario show that, in the naïve case
temporo-occipital connection was the same that for the inter-temporal
of areas perfectly homogeneous in their activity, all the approaches eval-
one.
uated perform adequately. However, the PS calculated using the source
The behavior of the average and RMS PS are similar in all the scenar-
with the greatest power as a representative of the whole area showed
ios. The methods based on PCA95 and the average time series give sim-
inconsistent results. This result suggests that the selection of this repre-
ilar results when either temporo-occipital connections are present but
sentative time series is strongly affected by noise, and should thus be
drop when both connections appear simultaneously. On the other hand,
avoided.

5
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

Fig. 5. Estimated multivariate PS for different cou-


plings in scenario 2. (a) Multivariate PS matrix be-
tween the four areas considered. The left panel
shows the theoretical model. The right panel shows
the multivariate PS matrix calculated using the
PCA95 approach for ten levels of coupling ranging
from 10% to 100%. (b) Multivariate PS between
left and right temporal poles estimated with dif-
ferent PCA-based approaches for ten levels of cou-
pling ranging from 0% to 100%. (c) Same as (b)
but using the five different approaches. PS-AVG:
Average of the bivariate PS matrix. PS-RMS: Root-
mean-square of the bivariate PS matrix. 1PCA: First
component, according to PCA. 2PCA: RMS of the
two first components, according to PCA. PCA95:
RMS of the components adding up to 95% of the
variance, according to PCA. All the lines are rep-
resented as average across simulations (solid line)
and standard deviation (shadowed area).

The average time series choice shows a behavior slightly different virtually identical, indicate that the approach might be unreliable and
from the rest of the approaches when increasing coupling. While, in highly dependent on the details (likely the location of the activity inside
general, all the approaches steadily increment their estimation with in- the area).
creasing coupling, this approach seems to reach a plateau at around Last, both multivariate approaches (average and RMS of the bivariate
80%. The approach seems, thus, unable to distinguish between differ- PS matrix) show similar behavior, with the average estimating lower
ent values of high coupling. However, this plateau never takes place for connectivity than the RMS. This effect is mathematically congruent, as
the first scenario, where only the coupled activity is present. This be- in the root mean square of a set of numbers the larger numbers dominate
havior is evident in Fig. 7 and could be interpreted as an inconsistency. more than in the average. Both approaches show a steady linear increase
While this approach is not very commonly used for the study of PS, we with coupling, and their behavior is consistent across scenarios, showing
strongly discourage it. high stability in their estimates. While both approaches are similar, the
The (commonly chosen) use of the first PCA as representative time properties of the PS metrics used here (PLV and ciPLV), similarly to
series for the whole area behaves correctly for the basic scenario but coherence, behave more consistently with the squared value than with
fails when the brain areas show a more complex mixture of activities. the original one. For example, PLV is, as all the sample-based estimates,
This result is to be expected, as the first PCA successfully extracts only biased up with small sample sizes, and this bias is proportional to the
one of the region’s activities, ignoring the rest. This limitation can be squared value of PLV [45]. In this light, it makes sense to use the root
overcome using the root-mean-squared value of several components, mean square of the bivariate PLV values to obtain the between-areas
but this approach requires selecting an arbitrary threshold (e.g., 95% PLV value, or even work directly with the squared PLV instead of the
of the total variance, in our case). In the extreme case where we take all raw value (as it is common to work with the squared value of coherence
the components, the result should converge to the bivariate PS matrix’s [46]).
RMS. Analyzing each scenario individually, all the approaches (except for
As representative time series, the centroid shows attractive advan- the source with the greatest power) show a high correlation with the
tages, being one of the most important the simplified source model. As coupling (i.e., a straight increase in estimated PS with increasing cou-
in the first PCA case, this approach behaves correctly in simple sce- pling). However, when comparing the same coupling in different sce-
narios, where only the coupled activity is present. However, in more narios, the calculated synchronization varies considerably for the ap-
complex scenarios, where that amount of information increases, the es- proaches based on representative sources. On the other hand, both the
timated connectivity decreases. Indeed, the centroid gives the lowest PS mean and the RMS of the bivariate PS matrix show high consistency
values of all the approaches in all the scenarios, except for the third one. across scenarios. This consistency might be desired or not, depending on
Moreover, the diverging results in scenarios three and five, which are the interpretation assigned to multivariate PS. If multivariate PS should

6
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

Fig. 6. Estimated multivariate PS for different cou-


plings in scenarios 3, 5 and 7. (a) Multivariate PS
between left and right temporal poles estimated
with every approach for ten levels of coupling rang-
ing from 0% to 100% in scenario 3. (b) Same as
(a) but for scenario 5. (c) Idem but for scenario 7.
PS-AVG: Average of the bivariate PS matrix. PS-
RMS: Root-mean-square of the bivariate PS matrix.
PCA95: RMS of the components adding up to 95%
of the variance, according to PCA. All the lines are
represented as average across simulations (solid
line) and standard deviation (shadowed area).

represent the net amount of synchronization between both areas, simi- den introduced by them is an essential factor to consider. The AAL atlas
lar couplings in different scenarios should give the same synchronization used in this work comprises 80 cortical regions and 1210 source po-
levels, as in the case of the mean or the RMS of the bivariate PS matrix. sitions. This configuration gives rise to 3160 pairs of areas and over
On the other hand, if multivariate PS should represent the fraction of 730,000 pairs of sources. The multivariate approaches require the cal-
activity of an area synchronized with other areas, its value should vary culation of all the source-to-source PS, dramatically increasing the re-
in different scenarios, as in the synchronization between average time quired computational power. However, the new computationally algo-
series. Both interpretations are equally valid, but we will take part in rithms for calculating PS allows for estimating whole-brain source to
the first one. source FC in only a few minutes using an average computer [31].
Another property of the multivariate approaches open to interpre- In fairness, some limitations of this work must be highlighted. First,
tation is the amount of synchronized activity in each of the simulated the source model used here is straightforward, with only six sets of
scenarios. In our case, we selected two areas (both temporal poles) with sources generating measurable activity (and the rest of the sources gen-
nine sources each, and defined a cluster of activity comprising three con- erating low-level noise). In a realistic scenario the number of sources
tiguous sources in each area. When the coupling was set to the maximum is much higher, in the order of thousands, and the limited number of
value, it could be argued that the expected synchronization should be sensors is usually unable to disentangle all of them. The selected simu-
one third, as one-third of the sources of one brain area is connected to lations also include a set of chaotic oscillators that, although nonlinear,
one-third of the sources in the other. The RMS of the bivariate PS ma- produce clear phase synchronizations (see the Supplementary Materials
trix seems to give a result close to that. On the other hand, PCA95 cal- for a thorough evaluation of these couplings). In a real-world scenario,
culates the root-mean-squared of a reduced bivariate PS matrix, where the couplings between different parts of the brain will be more involved,
each element represents the synchronization between components, and and the PS algorithms might be unable to capture them correctly. Last, in
the information of how many sources take part in this component is re- our simulations we used a single orientation per source position, namely
moved, giving as a result a higher synchronization value. Again, both the higher signal-to-noise ratio in the forward model, so the source ori-
interpretations are valid, but we will take part in the first one. entation is determined. In a realistic scenario the sources’ orientation
While our interpretation of area to area FC favors the estimations ob- is unknown, and FC estimations should not be affected by the selected
tained using the purely multivariate scenarios, the computational bur- coordinate system (this is, should be insensitive to rotations of the coor-

7
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

Fig. 7. Estimated multivariate PS for fixed val-


ues of coupling in different scenarios. (a) Multi-
variate PS between the left temporal pole and the
right temporal pole estimated with the five differ-
ent approaches for 50% couplings in several sce-
narios. (b) Same as (a) but for 100% coupling. All
the lines are represented as average across simula-
tions (solid line) and standard deviation (shadowed
area).

dinate framework [21,22]). PLV (and its extension ciPLV) are nonlinear References
metrics and are therefore sensitive to orientation changes, but a linear
extension has been recently developed, namely Hilbert coherence [31], [1] R. Hari, A. Puce, MEG-EEG Primer, Oxford University Press, 2017.
10.1093/med/9780190497774.001.0001.
which could be modified to behave correctly in this scenario. [2] G.H. Glover, Overview of functional magnetic resonance imaging, Neurosurg. Clin.
N. Am. 22 (2011) 133–139, doi:10.1016/j.nec.2010.11.001.
5. Conclusion [3] M.S. Hämäläinen, R. Hari, R.J. Ilmoniemi, J. Knuutila, O.V. Lounasmaa, Mag-
netoencephalography—Theory, instrumentation, and applications to noninvasive
studies of the working human brain, Rev. Mod. Phys. 65 (1993) 413–497,
According to the results, and based on our inter-area synchronization doi:10.1103/RevModPhys.65.413.
interpretation, the simulations support using the approaches based on [4] M.G. Rosenblum, A.S. Pikovsky, J. Kurths, Phase synchronization of chaotic oscilla-
tors, Phys. Rev. Lett. 76 (1996) 1804–1807, doi:10.1103/PhysRevLett.76.1804.
the bivariate synchronization matrix. In particular, the approach that [5] K.J. Friston, Functional and effective connectivity in neuroimaging: a synthesis,
gives results more similar to those expected is the root mean square Hum. Brain Mapp. 2 (1994) 56–78, doi:10.1002/hbm.460020107.
of the values in the bivariate PS matrix. However, depending on the [6] J.C. Mosher, R.M. Leahy, P.S. Lewis, EEG and MEG: forward solutions for inverse
methods, IEEE Trans. Biomed. Eng. 46 (1999) 245–259, doi:10.1109/10.748978.
interpretation of the concept of inter-area PS, the use of a summary of [7] M. Scherg, Fundamentals of dipole source potential analysis, in: F. Grandori,
the area’s activity based on PCA (for example, components explaining M. Hoke, G.L. Romani (Eds.), Audit. Evoked Magn. Fields Electr. Potentials. Adv.
up to 95% of the variance) could give more acceptable results. Audiol., S. Karger AG, 1990, pp. 40–69.
[8] M.S. Hämäläinen, R.J. Ilmoniemi, Interpreting magnetic fields of the brain:
On the other hand, using a representative time series has proven to
minimum norm estimates, Med. Biol. Eng. Comput. 32 (1994) 35–42,
be overly dependent on the scenario and thus highly inconsistent. These doi:10.1007/BF02512476.
approaches are strongly discouraged by the authors. [9] B.D. van Veen, W. van Drongelen, M. Yuchtman, A. Suzuki, Localization of brain
electrical activity via linearly constrained minimum variance spatial filtering, IEEE
Trans. Biomed. Eng. 44 (1997) 867–880, doi:10.1109/10.623056.
Funding [10] M. Junghöfer, T. Elbert, D.M. Tucker, B. Rockstroh, Statistical control of arti-
facts in dense array EEG/MEG studies, Psychophysiology 37 (2000) 523–532,
doi:10.1017/S0048577200980624.
This work was partially funded by the Spanish Ministry of Sci-
[11] J.M. Palva, S.H. Wang, S. Palva, A. Zhigalov, S. Monto, M.J. Brookes, J.-
ence (Grants TEC2016-80063-C3-2-R, PSI2017-91955-EXP, PID2019- .M. Schoffelen, K. Jerbi, Ghost interactions in MEG/EEG source space: a note
111537GB-C22, and RTI2018-098,762-B-C31), the government of Ca- of caution on inter-areal coupling measures, Neuroimage 173 (2018) 632–643,
nary Islands (Grant ProID2017010100), and Comunidad de Madrid doi:10.1016/j.neuroimage.2018.02.032.
[12] N. Tzourio-Mazoyer, B. Landeau, D. Papathanassiou, F. Crivello, O. Etard, N. Del-
(Grant B2017/BMD-3760 (NEUROCENTRO-CM)). croix, B. Mazoyer, M. Joliot, Automated anatomical labeling of activations in SPM
using a macroscopic anatomical parcellation of the MNI MRI single-subject brain,
Declaration of Competing Interest Neuroimage 15 (2002) 273–289, doi:10.1006/nimg.2001.0978.
[13] R.S. Desikan, F. Ségonne, B.R. Fischl, B.T. Quinn, B.C. Dickerson, D. Blacker,
R.L. Buckner, A.M. Dale, R.P. Maguire, B.T. Hyman, M.S. Albert, R.J. Killiany,
None. An automated labeling system for subdividing the human cerebral cortex on
MRI scans into gyral based regions of interest, Neuroimage 31 (2006) 968–980,
doi:10.1016/j.neuroimage.2006.01.021.
Supplementary materials [14] M.F. Glasser, T.S. Coalson, E.C. Robinson, C.D. Hacker, J. Harwell, E. Yacoub,
K. Ugurbil, J. Andersson, C.F. Beckmann, M. Jenkinson, S.M. Smith, D.C. Van Essen,
Supplementary material associated with this article can be found, in A multi-modal parcellation of human cerebral cortex, Nature 536 (2016) 171–178,
doi:10.1038/nature18933.
the online version, at doi:10.1016/j.brain.2021.100021.

8
R. Bruña and E. Pereda Brain Multiphysics 2 (2021) 100021

[15] O. Korhonen, S. Palva, J.M. Palva, Sparse weightings for collapsing inverse solu- subjective cognitive decline to mild cognitive impairment, Int. J. Neural Syst. 27
tions to cortical parcellations optimize M/EEG source reconstruction accuracy, J. (2017) 1750041, doi:10.1142/S0129065717500411.
Neurosci. Methods. 226 (2014) 147–160, doi:10.1016/j.jneumeth.2014.01.031. [33] A.M. Dale, A.K. Liu, B.R. Fischl, R.L. Buckner, J.W. Belliveau, J.D. Lewine,
[16] H. Luckhoo, J.R. Hale, M.G. Stokes, A.C. Nobre, P.G. Morris, M.J. Brookes, E. Halgren, Dynamic statistical parametric mapping: combining fMRI and MEG
M.W. Woolrich, Inferring task-related networks using independent compo- for high-resolution imaging of cortical activity, Neuron 26 (2000) 55–67,
nent analysis in magnetoencephalography, Neuroimage 62 (2012) 530–541, doi:10.1016/S0896-6273(00)81138-1.
doi:10.1016/j.neuroimage.2012.04.046. [34] E.L. Hall, M.W. Woolrich, C.E. Thomaz, P.G. Morris, M.J. Brookes, Using variance
[17] R.C. Craddock, G.A. James, P.E. Holtzheimer, X.P. Hu, H.S. Mayberg, A whole brain information in magnetoencephalography measures of functional connectivity, Neu-
fMRI atlas generated via spatially constrained spectral clustering, Hum. Brain Mapp. roimage (2013), doi:10.1016/j.neuroimage.2012.11.011.
33 (2012) 1914–1928, doi:10.1002/hbm.21333. [35] K.J. Friston, P. Rotshtein, J.J. Geng, P. Sterzer, R.N. Henson, A cri-
[18] G.A. James, O. Hazaroglu, K.A. Bush, A human brain atlas derived via n-cut parcel- tique of functional localisers, Neuroimage 30 (2006) 1077–1087,
lation of resting-state and task-based fMRI data, Magn. Reson. Imaging. 34 (2016) doi:10.1016/j.neuroimage.2005.08.012.
209–218, doi:10.1016/j.mri.2015.10.036. [36] D.R. King, M. de Chastelaine, R.L. Elward, T.H. Wang, M.D. Rugg, Recollection-
[19] P. Garcés, M.C. Martín-Buro, F. Maestú, Quantifying the test-retest reliability of mag- related increases in functional connectivity predict individual differences in
netoencephalography resting-state functional connectivity, Brain Connect 6 (2016) memory accuracy, J. Neurosci. 35 (2015) 1763–1772, doi:10.1523/JNEU-
448–460, doi:10.1089/brain.2015.0416. ROSCI.3219-14.2015.
[20] H. Hotelling, Relations between two sets of variates, Biometrika 28 (1936) 321–377, [37] L.E. Hughes, R.N. Henson, E. Pereda, R. Bruña, D. López-Sanz, A.J. Quinn,
doi:10.1093/biomet/28.3-4.321. M.W. Woolrich, A.C. Nobre, J.B. Rowe, F. Maestú, Biomagnetic biomarkers for
[21] A. Ewald, L. Marzetti, F. Zappasodi, F.C. Meinecke, G. Nolte, Estimat- dementia: a pilot multicentre study with a recommended methodological frame-
ing true brain connectivity from EEG/MEG data invariant to linear and work for magnetoencephalography, Alzheimer’s Dement. Diagn. Assess. Dis. Monit.
static transformations in sensor space, Neuroimage 60 (2012) 476–488, (2019), doi:10.1016/j.dadm.2019.04.009.
doi:10.1016/j.neuroimage.2011.11.084. [38] J.T.C. Schwabedal, H. Kantz, Optimal extraction of collective oscillations from un-
[22] A. Basti, V. Pizzella, F. Chella, G.L. Romani, G. Nolte, L. Marzetti, Disclosing large- reliable measurements, Phys. Rev. Lett. 116 (2016) 104101, doi:10.1103/Phys-
scale directed functional connections in MEG with the multivariate phase slope in- RevLett.116.104101.
dex, Neuroimage 175 (2018) 161–175, doi:10.1016/j.neuroimage.2018.03.004. [39] G. Nolte, A. Ziehe, V.V. Nikulin, A. Schlögl, N. Krämer, T. Brismar, K.R. Müller,
[23] L. Geerligs, R.N.H. Cam-CAN, Functional connectivity and structural co- Robustly estimating the flow direction of information in complex physical systems,
variance between regions of interest can be measured more accurately Phys. Rev. Lett. 100 (2008) 234101, doi:10.1103/PhysRevLett.100.234101.
using multivariate distance correlation, Neuroimage 135 (2016) 16–31, [40] M.E. López, R. Bruña, S. Aurtenetxe, J.Á. Pineda-Pardo, A. Marcos, J. Arrazola,
doi:10.1016/j.neuroimage.2016.04.047. A.I. Reinoso, P. Montejo, R. Bajo, F. Maestú, Alpha-band hypersynchronization in
[24] A. Basti, H. Nili, O. Hauk, L. Marzetti, R.N. Henson, Multi-dimensional connec- progressive mild cognitive impairment: a magnetoencephalography study, J. Neu-
tivity: a conceptual and mathematical review, Neuroimage 221 (2020) 117179, rosci. 34 (2014) 14551–14559, doi:10.1523/JNEUROSCI.0964-14.2014.
doi:10.1016/j.neuroimage.2020.117179. [41] D. López-Sanz, R. Bruña, P. Garcés, M.C. Martín-Buro, S. Walter, M.L. Delgado,
[25] J.-.P. Lachaux, E. Rodriguez, J. Martinerie, F.J. Varela, Measuring phase M. Montenegro, R.L. Higes, A. Marcos, F. Maestú, Functional connectivity disruption
synchrony in brain signals, Hum. Brain Mapp. 8 (1999) 194–208, in subjective cognitive decline and mild cognitive impairment: a common pattern of
doi:10.1002/(SICI)1097-0193(1999)8:4<194::AID-HBM4>3.0.CO;2-C. alterations, Front. Aging Neurosci 9 (2017) 109, doi:10.3389/fnagi.2017.00109.
[26] F. Mormann, K. Lehnertz, P. David, C.E. Elger, Mean phase coherence as a measure [42] C. Allefeld, J. Kurths, An approach to multivariate phase synchronization analysis
for phase synchronization and its application to the EEG of epilepsy patients, Phys. and its application to event-related potentials, Int. J. Bifurc. Chaos. 14 (2004) 417–
D Nonlinear Phenom. 144 (2000) 358–369, doi:10.1016/S0167-2789(00)00087-7. 426, doi:10.1142/S0218127404009521.
[27] R. Quian-Quiroga, J. Arnhold, P. Grassberger, Learning driver-response relationships [43] G. Lohmann, D.S. Margulies, A. Horstmann, B. Pleger, J. Lepsien, D. Goldhahn,
from synchronization patterns, Phys. Rev. E. Stat. Phys. Plasmas. Fluids. Relat. In- H. Schloegl, M. Stumvoll, A. Villringer, R. Turner, Eigenvector centrality mapping
terdiscip. Topics. 61 (2000) 5142–5148, doi:10.1103/PhysRevE.61.5142. for analyzing connectivity patterns in fMRI data of the human brain, PLoS One 5
[28] B. Widrow, K.M. Duvall, R.P. Gooch, W.C. Newman, Signal cancellation phenomena (2010) e10232, doi:10.1371/journal.pone.0010232.
in adaptive antennas: causes and cures, IEEE Trans. Antennas Propag. 30 (1982) [44] A. Groth, M. Ghil, Multivariate singular spectrum analysis and the road to phase
469–478, doi:10.1109/TAP.1982.1142804. synchronization, Phys. Rev. E - Stat. Nonlinear Soft Matter Phys. 84 (2011) 036206,
[29] J. Ashburner, K.J. Friston, Unified segmentation, Neuroimage 26 (2005) 839–851, doi:10.1103/PhysRevE.84.036206.
doi:10.1016/j.neuroimage.2005.02.018. [45] S. Aydore, D. Pantazis, R.M. Leahy, A note on the phase locking value and its prop-
[30] G. Nolte, The magnetic lead field theorem in the quasi-static approximation and its erties, Neuroimage 74 (2013) 231–244, doi:10.1016/j.neuroimage.2013.02.008.
use for magnetoencephalography forward calculation in realistic volume conductors, [46] P.L. Nunez, R. Srinivasan, A.F. Westdorp, R.S. Wijesinghe, D.M. Tucker,
Phys. Med. Biol. 48 (2003) 3637–3652, doi:10.1088/0031-9155/48/22/002. R.B. Silberstein, P.J. Cadusch, EEG coherency I: statistics, reference elec-
[31] R. Bruña, F. Maestú, E. Pereda, Phase locking value revisited: teaching new tricks to trode, volume conduction, Laplacians, cortical imaging, and interpretation at
an old dog, J. Neural Eng. 15 (2018) 056011, doi:10.1088/1741-2552/aacfe4. multiple scales, Electroencephalogr. Clin. Neurophysiol. 103 (1997) 499–515,
[32] D. López-Sanz, P. Garcés, B. Álvarez, M.L. Delgado-Losada, R. López-Higes, doi:10.1016/S0013-4694(97)00066-7.
F. Maestú, Network disruption in the preclinical stages of Alzheimer’s disease: from

You might also like