Brain Topography manuscript No.

(will be inserted by the editor)

Conditionally Gaussian Hierarchical Bayesian

Inversion of the P20-N20 Component of SEP/SEF
measurements with FEM Forward Modeling

A. Rezaei · M. Antonakakis · M.
C. Piastra · Q. He · A. Koulouri ·
V. Rimpiläinen · C.H. Wolters · S.
Pursiainen

the date of receipt and acceptance should be inserted later

Atena Rezaei
Laboratory of Mathematics, Tampere University, P.O. Box 692, 33101 Tampere, Finland
E-mail: atena.rezaei@tuni.fi
Marios Antonakakis
Institute of Biomagnetism and Biosignalanalysis, University of Münster, Germany, Malme-
dyweg 15, D-48149 Münster, Germany
Maria Carla Piastra
Institute of Biomagnetism and Biosignalanalysis, University of Münster, Germany, Malme-
dyweg 15, D-48149 Münster, Germany
Cognitive Neuroscience, Donders Institute for Brain, Cognition and Behaviour, Radboud
University Nijmegen Medical Centre, Nijmegen, The Netherlands
Qin He
Faculty of Information Technology and Communication Sciences, Tampere University, P.O.
Box 692, 33101 Tampere, Finland
Laboratory of Signal Processing, Tampere University, P.O. Box 553, 33101 Tampere, Fin-
land
Alexandra Koulouri
Faculty of Information Technology and Communication Sciences, Tampere University, P.O.
Box 692, 33101 Tampere, Finland Department of Physics, Aristotle University of Thessa-
loniki, Thessaloniki 54124, Greece
Ville Rimpiläinen
Department of Physics, University of Bath, Claverton Down, BA27AY Bath, United King-
dom
Carsten H. Wolters
Institute of Biomagnetism and Biosignalanalysis, University of Münster, Germany, Malme-
dyweg 15, D-48149 Münster, Germany
Otto Creutzfeldt Center for Cognitive and Behavioral Neuroscience, University of Münster,
Münster, Germany
Sampsa Pursiainen
Faculty of Information Technology and Communication Sciences, Tampere University, P.O.
Box 692, 33101 Tampere, Finland
2 A. Rezaei et al.

Abstract This paper advances computational inversion methods, where the

neural activity of the brain is detected using electro- and magnetoencephalog-
raphy. We focus on the conditionally Gaussian hierarchical Bayesian model
(CG-HBM) which has been shown to reconstruct networks with focal sources
without indicating the number of sources and without depth localization deficits.
The goal of this proof-of-concept study is to, for the first time, reconstruct
the activity corresponding to somatosensory evoked potential (SEP) and field
(SEF) measurements with the CG-HBM coupled with an advanced finite ele-
ment method forward approach. We parameterize the CG-HBM using a syn-
thetic reference source placed in the cytoarchitectonic Brodmann area 3b of the
central sulcus. We compare maximum a posteriori optimization and Markov
chain Monte Carlo sampling techniques as well as the gamma and inverse
gamma hypermodels in detecting the activity in the sulcal wall. Furthermore,
we apply the CG-HBM to reconstruct SEP and SEF activity obtained in elec-
tric wrist stimulation for a healthy human subject. We show that the generator
of the P20-N20 component, 20 ms post-stimulus, is localized in the area 3b.
While our method is theoretically able to also localize deep focal sources in
the presence of lateral ones, here and in this subject it finds the P20-N20 peak
without an overlaid thalamic activity for both EEG and MEG, supporting the
notion that our approach to reconstruct a single region might be appropriate
at least for the present experimental setup.
Keywords Electroencephalography (EEG) · Magnetoencephalography
(MEG) · Evoked Somatosensory Activity · Iterative Alternating Sequential ·
Hierarchical Bayesian Method · Measured Data

1 Introduction

This paper investigates computational inversion methods in electro- and mag-

netoencephalography (EEG and MEG) (Hämäläinen et al. 1993; Niedermeyer
and da Silva 2004; Brette and Destexhe 2012). We focus on the condition-
ally Gaussian hierarchical Bayesian model (CG-HBM) (O’Hagan and Forster
2004) which, based on numerical simulations Calvetti et al. (2009); Lucka et al.
(2012), has been suggested to reconstruct networks with focal sources with-
out indicating the number of sources and without depth localization deficits.
The goal of this proof-of-concept study is to, for the first time, reconstruct
the activity corresponding to somatosensory evoked potential (SEP) and field
(SEF) measurements by coupling the CG-HBM with an advanced finite el-
ement method (FEM) forward approach (de Munck et al. 2012; Pursiainen
et al. 2016). Our objective is to detect the P20-N20 component response to
the median nerve (wrist) stimulation and to recover the corresponding current
density without limiting the number of the active source components in the
model (Haueisen et al. 2007; Buchner et al. 1994; Hari and Puce 2017; Allison
et al. 1991). The activity is reconstructed in the hand-knob of the primary
somatosensory cortex (SI), in the Brodmann area 3b which is located in the
posterior bank of the central sulcus. We investigate the peak of the P20-N20
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 3

component occurring 20 ms after the stimulus, because it is the most stable

transient SEP/SEF component corresponding to an excellent signal-to-noise
ratio (SNR) in especially MEG, but also in EEG (Buchner et al. 1994; Fuchs
et al. 1998). The motivation to explore this relatively simple single-component
scenario is to provide a suitable framework for parameterizing the present
inversion algorithms and evaluating them with the real dataset.
Reconstructing the primary current field of the neurons as a 3D distribution
restricted to the grey matter of the brain is an ill-conditioned inverse problem
in which the applied prior model and reconstruction technique have a major
effect on the final result (Brette and Destexhe 2012). Consequently, a priori
information such as a physiologically accurate head model is needed. In this
case, finding the activity in the 3b area of SI, i.e., in the posterior wall of
the central sulcus (Figure 1), is challenging, since EEG and, specifically, MEG
are classically known to be more sensitive to the superficial parts than to the
deeper lying areas in the sulci (Ahlfors et al. 2010; Hari et al. 2018).
We compare a gradient-based maximum a posteriori (MAP) optimization
with a sampling-based conditional mean (CM) estimation technique as well as
the gamma (G) and inverse gamma (IG) hypermodels in detecting the activ-
ity in the sulcal wall. The iterative alternating sequential (IAS) and Markov
chain Monte Carlo (MCMC) sampling technique presented in Calvetti et al.
(2009) are applied. The present divergence conforming H(div) finite element
method (FEM) forward modeling approach (Pursiainen et al. 2016) allows for
creating an accurate volumetric discretization of a multi-compartment head
segmentation regarding its conductivity distribution and strongly folded tissue
structures (Braess 2001; Haueisen et al. 2002; Ramon et al. 2006; Beltrachini
2018; Vorwerk et al. 2014). The computations were performed with the Zef-
firo interface1 (ZI) software pipeline (He et al. 2018) which couples the H(div)
forward model with CG-HBM.

Fig. 1 Left: A schematic illustration depicting the sagittal cut of the primary somatosen-
sory cortex (SI). The P20-N20 component of the somatosensory activity occurs in the Brod-
mann area 3b of SI, which is located in the posterior wall of the central sulcus (Allison
et al. 1991). Right: The orientation of the primary currents (SEP/SEF components) in the
cerebral cortex is normal with respect to the surface due to the normal alignment of the
pyramidal cells (Hari et al. 2018).

1 Zeffiro is Italian for a ’gentle breeze’. The source code is available on-line at https:

//github.com/sampsapursiainen/zeffiro_interface/graphs/traffic
4 A. Rezaei et al.

CG-HBM was found to reconstruct the generator of the P20-N20 compo-

nent in the area 3b via both the IAS and MCMC technique in altogether 32
inversion tests. While our method is theoretically able to localize deep focal
sources in the presence of lateral ones without additional assumptions on the
unknown number of sources parameter, the reconstructions found for the P20-
N20 peak without an overlaid thalamic activity, suggesting that our approach
to reconstruct a single region might be appropriate at least for the present
experimental setup. The simulation results obtained for the known synthetic
source suggest that through MAP estimation, one can reconstruct the P20-
N20 component without a priori limiting the region of the activity. MCMC
sampling was observed to allow adapting the posterior to the structure and
resolution of the underlying numerical model and geometry, to avoid a numer-
ical bias, e.g., overly focal results. A sampling-based CM estimate obtained
for a region of interest (ROI) was found to enhance the distinction of the
reconstructed activity distribution in the sulcal wall compared to the MAP
optimization results, while the center of mass for the reconstructed distribu-
tion is essentially the same for both approaches.

2 Methods

This section briefly reviews the mathematical CG-HBM approach and its im-
plementation in this study. We denote by x the primary current density of
the neural activity, that is the unknown of the inverse problem, and by y the
measurement data vector. In both EEG and MEG, the dependence of y on x,
i.e., the forward model, can be formulated via the lead field matrix equation
of the form y = Lx + n, where n is a noise vector and L is the so-called
lead field matrix (Hämäläinen et al. 1993). Here L is obtained via the FEM
discretization of the classical field equations following from the quasi-static
approximation of the Maxwell’s equations as described in Miinalainen et al.
(2019); Pursiainen et al. (2016); Pursiainen (2012).

2.1 Conditionally Gaussian Hierarchical Bayesian model

For a single given data set y, the classical Bayes formula for subjective condi-
tional probabilities can be written as
p(x) p(y | x)
p(x | y) = ∝ p(x) p(y | x). (1)
p(y)
That is, the posterior probability density p(x | y) of the unknown primary
current distribution x in the brain is proportional to the product between the
prior density p(x), i.e., the a priori knowledge of x, and the likelihood function
p(y | x) following from the measurement noise model (Schmidt et al. 1999).
The measurement error is here assumed to be a Gaussian zero mean random
vector n = y − Lx with independent entries. Consequently, the likelihood is
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 5

of the form p(y | x) ∝ exp(−(2σ 2 )−1 kLx − yk2 ), where σ is the standard
deviation of the noise. In the hierarchical Bayesian approach, one assumes the
prior to be a joint density
p(x, θ) ∝ p(θ) p(x | θ) (2)
of x and a hyperparameter Rθ. That is, the actual prior is, in fact, a marginal
density of the form p(x) ∝ p(θ) p(x | θ) dθ.
In CG-HBM O’Hagan and Forster (2004); Calvetti et al. (2009); Calvetti
and Somersalo (2007), the conditional part p(x | θ) is also a zero mean Gaus-
sian density. Its diagonal covariance matrix is predicted by a long-tailed hyper-
prior p(θ), meaning that the variance vector, i.e., the set of diagonal entries,
is likely to contain outliers. Thus, it is implicitly assumed that x is a sparse
vector with a small subset of entries, which are considerably large in absolute
value compared to the others (Sato et al. 2004). The number and intensity of
these outliers are controlled by the hyperprior (Calvetti and Somersalo 2007).
The resulting impulse-like prior model for the unknown is particularly advan-
tageous to obtain a focal reconstruction for the brain activity, since one can
control how many the entries in x essentially differ from zero.

2.1.1 Gamma and inverse gamma hyperprior

As the hyperprior, we test both the gamma G(θ | β, θ0 ) and inverse gamma
IG(θ | β, θ0 ) distribution, whose density is supported on the set of non-negative
real numbers with a structure determined by the scale and shape parameter
θ0 and β, respectively. The first one of these controls the hyperprior’s effective
support size as well as its distance from the origin: the greater value the larger
the support and the greater the maximum. Again, the second one steers the
decay rate of its tail.

2.1.2 Parameter selection

We use an approach in which β ≥ 1.5 and θ0 > 0 are subjectively chosen by

the scientist. The scale parameter θ0 can here be interpreted as the sensitivity
of detecting brain activity which grows as θ0 tends to zero. A lower sensitivity
means also a greater vulnerability to noise effects and numerical posterior
estimation errors. Namely, the posterior density for a very low value of θ0 will
be highly concentrated around a certain region, which can be difficult to be
found numerically, when exploring the posterior. In such a situation, to ease up
the reconstruction process, one can increase the value of the shape parameter
β, which should, nevertheless, be as small as possible, if the goal is to obtain
a focal reconstruction.

2.1.3 Posterior exploration

Given the posterior, the actual reconstruction can be found via several different
approaches. The most common ones can be divided to optimization and sam-
pling techniques. The former ones include the MAP algorithms, which target at
6 A. Rezaei et al.

finding the maximizer of the posterior density, i.e., xM AP = argmax ppost (x, θ |
y). MAP estimation usually provides the faster but less robust way to obtain a
reconstruction than the sampling
R techniques that approximate the conditional
mean xCM = E(x, θ | y) = (x, θ)ppost (x, θ | y) dx dθ (Kaipio and Somersalo
2004). The detailed descriptions for the IAS MAP estimation method and the
MCMC sampler employed for CM estimation in this study can be found in
Calvetti et al. (2009).

2.2 Numerical model

2.2.1 Head segmentation

Fig. 2 1st and 2nd from the left: A top view of the positions for the successfully
recorded 72 electrode (left) and 271 magnetometer (right) channels. The magnetometer ori-
entations are also shown. 3rd and 4th from the left: A sagittal cut of the six compartment
tetrahedral FE mesh for 1 and 2 mm resolution (left and right, respectively). The modeled
compartments are the following: skin (brown), compact bone (beige), spongious bone (blue),
cerebrospinal fluid (green), grey matter (grey) and white matter (light grey).

The FE implementation approach presented in Pursiainen et al. (2012)

was applied including the formula for the EEG and MEG lead field matrix: a
tetrahedral finite element mesh for the 49-year old healthy male subject was
generated by subdividing each voxel in a surface based regular hexahedral
segmentation into six tetrahedra. The FE meshes were generated using a six-
layer surface segmentation based T1-weighted and T2-weighted MRI sequences
recorded with a 3T MRI scanner. The surfaces (level-sets) of skin, compact
bone (skull), spongious bone (skull), cerebrospinal fluid (CSF), grey matter
and white matter were included in the model. A FE mesh was generated for
both 1 and 2 mm resolution (voxel size). The first one of these included 3.8
M nodes and 22 M tetrahedra and the second one 0.47 M nodes and 2.7 M
tetrahedra. A sagittal cut of each mesh is illustrated in Figure 2.

2.2.2 Source space

To generate the source space of x, we used the quadratic H(div) approach

presented in Pursiainen et al. (2016), employing the Position Based Optimiza-
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 7

tion (PBO) interpolation with the 10-source (8-point) stencil. That is, a given
dipolar current source was estimated via the 4 linear face and 6 quadratic edge
vector basis functions associated with the barycenter of the tetrahedron con-
taining the source position. The sources were placed in the interior part of the
grey matter compartment in the elements with a full set (four) of neighbors
belonging to the same compartment. The rest of the compartment forming the
boundary layer of the grey matter contained no sources, since the modeling
accuracy is known to be reduced for the boundary layer (Miinalainen et al.
2019).
To obtain a uniform (mesh independent) source density, 100,000 points
were distributed randomly in the grey matter for each FE mesh. A uniform
point spread was obtained through a straightforward random permutation due
to the uniform mesh structure. The points placed on the boundary layer in the
initial stage were filtered out of the eventual distribution consisting of 76,000
and 61,000 positions for 1 and 2 mm FE mesh, respectively. The lower source
count for the 2 mm case was caused by the thicker boundary layer following
from the larger element size. Each position comprised three sources oriented
along the three Cartesian coordinate axes. Hence, the total number of sources
was 228,000 and 183,000 for 1 and 2 mm FE mesh, respectively.
The Cartesian set of sources was used in inverting the data. After the
inversion process, the distribution obtained was projected using the neuro-
physiological normal constraint (Figure 1): the component coinciding with the
direction of the surface normal constituted the final reconstruction.

2.3 Measured data

The inversion computations were conducted using a (non-open) data set which
was obtained for a right-handed, 49-year-old, healthy male subject, whose right
median nerve was stimulated, when lying in a supine position in a magnetically
shielded room. Simultaneous SEP/SEF measurements were performed using
80 AgCl sintered ring electrodes (EASYCAP GmbH, Herrsching, Germany)
including 74 EEG channels with additional 6 channels for detecting eye move-
ments together with an MEG system including 275 axial gradiometers and 29
reference coils (OMEGA2005, VSM MedTech Ltd). Four magnetometers and
two EEG sensors were reported as defective channels. Measurements from 72
electrodes and 271 magnetometers (Figure 2) were used in the eventual data
set. The total number of 1,200 stimuli were obtained during a 10 minute mea-
surement session. The electric pulse duration was 0.5 ms, and, to determine the
magnitude, the stimulus strength was increased until a clear movement of the
thumb was visible. Each measurement had a 300 ms total duration, which was
subdivided into 100 ms pre-stimulus, stimulus and post-stimulus sub-intervals.
The inter-stimulus interval varied between 350 and 450 ms. The measurements
were averaged and pre-processed using a notch filter for the frequency 50 Hz
and for its harmonics to remove the power-line noise. The response measured
8 A. Rezaei et al.

EEG and MEG data

Fig. 3 The data obtained in the simultaneous EEG (left) and MEG (right) SEP/SEF
measurements. The actual stimulus response was measured for a 100 ms interval between
the pre-stimulus and post-stimulus phase. The P20-N20 component investigated in this
study corresponds to the peaks corresponding to the 20 ms post-stimulus time point which
is indicated by the vertical line.

for the different stimuli were averaged to produce the SEP/SEF data (Figure
3) the amplitude of which was normalized to one.
The eventual signal corresponded to the P20-N20 activity peak occurring
at the 20 ms post-stimulus time point (Figure 3). The data vector y for the
inversion computation was obtained for the 20 ms post-stimulus time point
after filtering the data with a low- and high-pass filter to obtain the frequency
range 20–250 Hz.

2.4 Synthetic data

To enable a comparison between the measured and synthetic data, a normally-

oriented synthetic dipolar source was placed in the hand-knob of the 3b area
in the posterior wall of the central sulcus (Figure 4). The position in the MNI
(Montreal Neurological Institute) coordinate system was (-38 mm, -28 mm,
56 mm) with ±2 mm accuracy. The data for the source was estimated using
the present FEM forward simulation and additive zero mean Gaussian noise
with 3 % standard deviation with respect to the maximal signal amplitude.
The same noise model was applied also to obtain the likelihood function for
the measured data.

2.5 Inverse estimates

The MAP estimate was found via three IAS iteration steps. For CM, a sample
of 10,000 points was created with the MCMC sampler. Of these, 1,000 points
in the beginning of the sequence were neglected as a burn-in phase. The CM
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 9

Table 1 The values of the scale parameter θ0 for different reconstructions specified by
the type of the data (Measured/Synthetic), measurement modality (EEG/MEG), estimate
(MAP/CM), and the hyperprior (G/IG).
Data Modality Estimate Hyperp. Scale
Meas. EEG MAP G 1E-12
IG 1E-11
CM G 1E-12
IG 1E-12
MEG MAP G 1E-12
IG 1E-11
CM G 1E-12
IG 1E-12
Synth. EEG MAP G 1E-12
IG 1E-11
CM G 1E-14
IG 1E-14
MEG MAP G 1E-11
IG 1E-11
CM G 1E-14
IG 1E-14

evaluation was performed by limiting the activity within a spherical ROI which
was defined as a single sub-domain based on preliminary MAP estimates (Fig-
ure 4) suggesting that the P14-N14 (thalamus) and P22-N22 (Brodmann area
1) components, which are peaked earlier and later with respect to P20-N20
(Buchner et al. 1994, 1995), were not distinguishable. The placement of the
ROI (Figure 4) was selected referring to the literature the hand-knob within
the Brodmann area 3b (Hari et al. 2018). The difference vector pointing from
the synthetic source position to the center point of the ROI was (4 mm, -1
mm, 2 mm).
The scale parameter values used in finding the inverse estimates can be
found in Table 1. Based on the preliminary tests, the scale parameter was
chosen to be θ0 = 1E-12 and θ0 = 1E-14 for the measured and synthetic data,
respectively. The main goal in selecting the scale parameter value was to make
the posterior probability model sensitive enough to detect the activity deep in
the sulcal wall. The value for the MAP estimation process was set to be larger
than for CM evaluation, if the gradient-based IAS algorithm did not otherwise
find the peak of the posterior.
The shape parameter value was fixed to β = 3. The main criterion for
selecting selecting this value was the shape of the posterior density, whose
peak (the credibility interval containing 90 % of the probability mass around
its maximum) was assumed to be have a few millimeters diameter w.r.t. the
mass center of x, i.e., comparable to the mutual distance of the lead field
source space density, in order to avoid the failure of the sampling process due
to an extremely peaked posterior structure.

2.5.1 Implemetation in Zeffiro interface

The present forward and inverse methods have been implemented in the Zeffiro
interface (ZI) (He et al. 2018) toolbox which uses the Matlab (The MathWorks
10 A. Rezaei et al.

Fig. 4 Left: The hand knob (green line) is a part of the Brodmann area 3b in the central
sulcus (between the blue and red areas) Hari et al. (2018). Center: General overview of the
P20-N20 component reconstruction in the posterior bank of the central sulcus, Brodmann
3b area. Right: The region of interest (ROI) utilized in the CM computations is the region
restricted by the 24 mm diameter grey spherical surface approximately covering the hand-
knob of the left hemisphere. The number of A MAP estimate for synthetic EEG data is
visualized on the cortex.

Inc.) platform. ZI aims to provide an user-friendly tool for advanced forward

and inverse computations, e.g., accurate lead field matrix construction, source
localization and time-lapse data analysis. ZI’s on-line code repository1 includes
the methods used in the present study.
To speed up the computations, ZI utilizes a Graphics Processing Unit
(GPU) in the following processes: (1) segmenting the FE grid, (2) creating
the lead field, (3) source space interpolation, and (4) inverting the data. The
ability to utilize the GPU is particularly important regarding the MATLAB
environment, which currently does not parallelize the sparse matrices result-
ing from the FE discretization in CPU. The following computation times were
obtained for 1 mm resolution 6-compartment test mesh with 36M elements,
6M nodes and ∼ 0.5M sources using Lenovo P910 ThinkStation equipped with
2 x Intel Xeon E5-2697A v4 CPU (RAM 256 GB) and 2 x NVIDIA Quadro
P6000 GPU (RAM 24 GB): (1) FE mesh generation: 1329 s, (2) EEG lead
field for 128 electrodes: 2362 s, (3) MEG lead field for 154 magnetometers:
4826 s, (4) Source space interpolation: 212 s.

2.6 Data and code availability statement

The ZI software utilized in this study is openly accessible at GitHub1 . The

data set is currently not open.

3 Results

The hierarchical Bayesian approach was found to successfully detect the peak
of the P20-N20 component in the area 3b of the left hemisphere in altogether
32 inversion tests. An overlaid P14-P14 (thalamus) or P22-N22 (Brodmann
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 11

Table 2 The values obtained for the spread of reconstructed source as well as for the
position and orientation difference between the synthetic source and the mass center of the
reconstructed one for each inverse estimate.
Spread Orientation Position
FE Mesh Data Modality Estimate Hyperp. (mm2 ) ∆ (deg) ∆ (mm)
1 mm Meas. EEG MAP G 44.7 9.7 3.4
IG 43.9 9.5 3.4
CM G 71.8 13.3 3.8
IG 37.7 14.4 3.8
MEG MAP G 32.0 10.8 3.4
IG 32.2 10.8 3.4
CM G 48.8 11.7 3.7
IG 42.0 12.5 3.8
Synth. EEG MAP G 42.0 5.8 3.5
IG 40.3 5.7 3.5
CM G 64.8 11.5 3.7
IG 69.9 12.1 3.7
MEG MAP G 22.0 12.8 3.5
IG 13.7 14.3 3.5
CM G 52.5 12.8 3.7
IG 52.6 12.6 3.7

2 mm Meas. EEG MAP G 34.8 9.7 3.4

IG 30.9 9.7 3.4
CM G 43.4 10.8 3.5
IG 67.7 7.0 3.4
MEG MAP G 87.9 11.5 3.5
IG 87.9 11.5 3.5
CM G 99.4 15.8 3.5
IG 115.9 15.6 3.5
Synth. EEG MAP G 64.2 6.7 3.3
IG 64.2 6.7 3.3
CM G 39.1 13.7 3.5
IG 8.3 14.2 3.3
MEG MAP G 60.2 10.2 3.5
IG 51.0 10.0 3.5
CM G 109.0 16.1 3.5
IG 110.2 16.1 3.5

area 1) activity was not detected. Table 2 includes a numeric measure for the
spread (focality) of the reconstructed and normally restricted activity defined
as the area of the a set in which its intensity exceeds 80 % of the maximum
value. The position and orientation difference between the synthetic source
and the mass center of the reconstructed activity are also given. Figures 5–8
visualize the reconstructed activity on the white matter surface (i.e. on the
inner surface of the grey matter) in the vicinity of the ROI. Each distribution
shows the activity in the direction of the outward-pointing surface normal.

3.1 MAP estimation

The MAP estimates are shown in Figure 5 and 7 for 1 and 2 mm FE mesh,
respectively. All the inverse estimates localized the activity in the sulcal wall
with the position and orientation difference varying less than 0.4 mm and
10.4 degrees within the complete set of reconstructions. In the 1 mm case,
they were maximally 3.8 mm and 14.4 degrees, respectively; MEG provided a
12 A. Rezaei et al.

Fig. 5 The MAP estimation results obtained with the IAS iteration corresponding to the
measured and synthetic data and the 1 mm FE mesh. The placement and orientation of the
synthetic source is shown by the red pin and the mass center of the reconstruction by the
green one.

slightly more focal reconstructions compared to EEG; and the IG hyperprior

led to more focal outcome than G for three out of four reconstructions. Based
on both visual and numerical inspection, the 1 mm FE mesh resolution yielded
a greater similarity between the MAP estimates obtained with the measured
and synthetic data as compared to the 2 mm case, where both the intensity
and spread of the reconstructions vary more than with the 1 mm mesh.

3.2 CM estimation

Figure 6 and 7 illustrate the CM estimates obtained for the ROI together with
the corresponding marginal densities (histograms) for the volumetric mass
center of the posterior. Overall, CM had a higher maximum intensity in a
mutual comparison to MAP regarding both real and synthetic data, while
the location of the maximum was virtually similar. The marginal densities
obtained show that maximum length of the 90 % credibility interval for the
marginal posterior mass center was in the range 0.9–1.4 and 1.2–2.3 mm for
the 1 and 2 mm FE mesh resolution, matching with the targeted range, and
that the mutual differences in the median, for each coordinate direction, were
less than 0.2 and 1.1 mm, respectively. Akin to the MAP, the spread of the CM
obtained with the 2 mm FE mesh resolution varied more between the different
reconstructions than in the case of the 1 mm mesh, while the maximal intensity
of the CM was observed to vary less than that of the MAP.
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 13

Fig. 6 The CM estimation results obtained with the sampler corresponding to the measured
and synthetic data and the 1 mm FE mesh. The marginal density (histograms) of the
(volumetric) posterior mass centre is illustrated for each case and Cartesian coordinate
component, including the median (red dashed line) and the 90 % credibility interval (blue
solid line). The placement and orientation of the synthetic source is shown by the red pin
and the mass center of the reconstruction by the green one.

4 Discussion

In this proof-of-concept study, we applied the conditionally Gaussian hierar-

chical Bayesian model (CG-HBM) (Calvetti et al. 2009; Lucka et al. 2012) for
the first time for source analysis of the P20-N20 component of electric wrist
stimulated somatosensory evoked potentials (SEP) and fields (SEF). We pre-
sented a process for parameter selection and showed via 32 inversion tests that
without a priori indication of the number of sources, the underlying sources of
the P20-N20 component were detected at and limited to the Brodmann area
3b in the primary somatosensory cortex. The computations were performed
using the Zeffiro interface (ZI) software tool (He et al. 2018).
The activation found in this study, was limited only to the area 3b, while
other works suggest that the thalamic or brainstem areas might also contribute
(Buchner et al. 1994, 1995, 1997; Haueisen et al. 2007; Götz et al. 2014). For
example, in Götz et al. (2014), the activations found at the P20-N20 peak were
14 A. Rezaei et al.

Fig. 7 The MAP estimation results obtained with the IAS iteration corresponding to the
measured and synthetic data and the 2 mm FE mesh.

Fig. 8 The CM estimation results obtained with the sampler corresponding to the measured
and synthetic data and the 2 mm FE mesh. The marginal density (histograms) of the
(volumetric) posterior mass centre is illustrated for each case and Cartesian coordinate
component.
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 15

exclusively limited in the area 3b in ten subjects and the thalamus was found
to be activated in two. Consequently, even if limiting the exploration to the
3b area might be appropriate for the majority of the subjects, it might be also
necessary to include the thalamus in some cases. One of the most important
future investigations will, therefore, be to apply our method in a somatosen-
sory group study in which more areas and their interaction as well as other
stimulation types, e.g., tactile stimulation, will be investigated. This future
direction is motivated by our CG-HBM implementation which is, in principle,
able to also reconstruct also deep sources in the presence of superficial activity,
as is suggested by the earlier computer simulations (Calvetti et al. 2009; Lucka
et al. 2012).
Based on the results, especially the position difference with respect to the
mass center, it seems that a source localization accuracy of around 4 mm could
be achieved in the area 3b. This coincides the maximal spatial accuracy found
for the MEG, i.e., 2-4 mm for superficial areas (Tarkiainen et al. 2003; Cohen
and Cuffin 1991), and even surpasses that of EEG which for superior locations
is in average about 9 mm (Cuffin et al. 2001a,b; Buchner et al. 1994). A
significant factor affecting the accuracy of EEG is the uncertainty related to the
conductivity distribution (Wang et al. 2009). However, taking into account the
total estimated 32-116 mm2 areal spread of the inverse estimates, the accuracy
found here does not exceed the suggested maximal accuracy limits. Here, the
spread of the estimates found follows from the current numerical framework
as the maximal achievable focality without a potential numerical bias. The
relationship between the estimates found and the actual physiological spread
of the source is not evaluated here and would need further work.
The values selected for the shape and scale parameter are crucial with
respect to the inversion outcome. Relying on the experience obtained from
this study, the scale parameter θ0 for detecting the P20-N20 component needs
to be selected so that the following two conditions will be satisfied. Firstly,
(i) θ0 has to be small enough to guarantee that the activity is reconstructed
in the sulcal wall instead of the gyrus part. At the same time, (ii) to avoid
a bias due to the numerical modeling accuracy, the posterior peak-width of
the posterior density should be maintained above the source the source space
density. The marginal densities obtained suggest that the condition (i) could
be satisfied for both EEG and MEG, when the the scale parameter was set to
be 1E-12 and 1E-14 for measured and synthetic data, respectively. A larger
value was found to be necessary to ease up the convergence of some IAS MAP
estimates, especially, in the case of the IG hyperprior. In order to fulfill (ii), the
shape parameter was given the value 3 which provided overall an appropriate
peak-width (the larger the scale parameter, the wider the peak). A lower value
was likely to result in a narrow peak which was not necessarily found by the
MCMC sampler.
CG-HBM, as a statistical model, might be advantageous for obtaining ro-
bust inverse estimates. When the parameters are chosen appropriately, the
sampler-based approach seems to provide a robust technique for estimating
the marginal posterior and the CM, giving information of the posterior distri-
16 A. Rezaei et al.

bution shape and structure. Here, it allowed us to adapt the it according to

the underlying numerical model and geometry, whereas the IAS MAP estima-
tion technique alone did not completely reveal the posterior shape, thereby,
leaving the possibility to obtain, e.g., overly focal estimates. It also seems that
estimating the CM via a sampling approach and defining a ROI for the sam-
pler is beneficial with respect to the distinguishability of the activity obtained.
We deem that controlling β and θ0 in the above described manner might be
essential for robust statistical detection of the sulcal activity with CG-HBM.
Since the IAS MAP estimation technique can be associated with the classical
minimum current and minimum support estimate (MCE and MSE) (Uutela
et al. 1999; Nagarajan et al. 2006; Calvetti et al. 2009), CG-HBM can also be
seen as a potential way to enhance the outcome obtained with these classical
methods. In particular, the present approach to select the shape and scale
parameter suggests that the optimal outcome of CG-HBM might be obtained
with β > 1.5, when it does not coincide with MSE or MCE.
The present forward simulation approach was found to perform adequately
with both 1 and 2 mm resolution. Agreeing with the existing knowledge on
physiologically accurate volumetric head modeling and forward simulation
(de Munck et al. 2012; Rullmann et al. 2009), the FE mesh resolution of 1
mm was observed to be advantageous for obtaining an overall even recon-
struction quality. The present GPU-based approach to the forward simulation
was found to be essential in order to achieve a feasibly short computation
time for the 1 mm mesh generation and lead field matrix evaluation processes
with our ZI implementation. Further ways to improve the present forward ap-
proach might include using the neurophysiological normal constraint in the
inverse computation stage. In this study, the normal constraint was applied
after inverting the data for a set of Cartesian source orientations.
In future work, it will be necessary to apply the present inversion technique
for more data sets and subjects, in order to learn more about the practical
localization capability of the CG-HBM. Potential directions for the mathe-
matical method development include inverstigating depth-localization, e.g.,
based on physiological knowledge (Calvetti et al. 2015, 2018) as well as the
difference between the G and IG hyperprior which performed here in a similar
manner without few exceptions. Another interesting direction is to implement
and develop inversion approaches using ZI. Following this, an ongoing topic
is to combine the presented CG-HBM source localization technique with an
automatically generated parcellation of the brain areas. A separate tool which
allows importing a parcellation generated with FreeSurfer2 has already been
implemented in ZI.

Acknowledgments

AR, QH, and SP were supported by the Academy of Finland Centre of Excel-
lence in Inverse Modelling and Imaging 2018–2025. AK was supported by the
2 https://surfer.nmr.mgh.harvard.edu
Conditionally Gaussian Inversion of the P20-N20 SEP/SEF Component 17

Academy of Finland Postdoctoral Researcher grant number 316542. MA, MCP

and CHW were supported by EU project ChildBrain (Marie Curie Innovative
Training Networks, grant agreement 641652), by the Deutsche Forschungsge-
meinschaft (DFG), project WO1425/7-1 and by the Priority Program 1665 of
the DFG (WO1425/5-2).

References

Ahlfors SP, Han J, Belliveau JW, Hämäläinen MS (2010) Sensitivity of MEG and EEG to
source orientation. Brain topography 23(3):227–232
Allison T, Wood CC, McCarthy G, Spencer DD (1991) Cortical somatosensory evoked po-
tentials. ii. effects of excision of somatosensory or motor cortex in humans and monkeys.
Journal of neurophysiology 66(1):64–82
Beltrachini L (2018) Sensitivity of the projected subtraction approach to mesh degenera-
cies and its impact on the forward problem in eeg. IEEE Transactions on Biomedical
Engineering
Braess D (2001) Finite Elements. Cambridge University Press, Cambridge
Brette R, Destexhe A (2012) Handbook of Neural Activity Measurement. Cambridge Uni-
versity Press, New York, DOI http://dx.doi.org/10.1017/CBO9780511979958
Buchner H, Fuchs M, Wischmann HA, Dössel O, Ludwig I, Knepper A, Berg P (1994)
Source analysis of median nerve and finger stimulated somatosensory evoked potentials:
multichannel simultaneous recording of electric and magnetic fields combined with 3d-
mr tomography. Brain topography 6(4):299–310
Buchner H, Adams L, Müller A, Ludwig I, Knepper A, Thron A, Niemann K, Scherg
M (1995) Somatotopy of human hand somatosensory cortex revealed by dipole source
analysis of early somatosensory evoked potentials and 3d-nmr tomography. Electroen-
cephalography and Clinical Neurophysiology/Evoked Potentials Section 96(2):121–134
Buchner H, Knoll G, Fuchs M, Rienäcker A, Beckmann R, Wagner M, Silny J, Pesch J
(1997) Inverse localization of electric dipole current sources in finite element models of
the human head. Electroencephalogr Clin Neurophysiol 102(4):267–78
Calvetti D, Somersalo E (2007) A gaussian hypermodel to recover blocky objects. Inverse
problems 23(2):733
Calvetti D, Hakula H, Pursiainen S, Somersalo E (2009) Conditionally Gaussian hypermod-
els for cerebral source localization. SIAM J Imaging Sci 2(3):879–909
Calvetti D, Pascarella A, Pitolli F, Somersalo E, Vantaggi B (2015) A hierarchical krylov–
bayes iterative inverse solver for MEG with physiological preconditioning. Inverse Prob-
lems 31(12):125,005
Calvetti D, Pascarella A, Pitolli F, Somersalo E, Vantaggi B (2018) Brain activity mapping
from MEG data via a hierarchical bayesian algorithm with automatic depth weighting.
Brain topography pp 1–31
Cohen D, Cuffin BN (1991) Eeg versus meg localization accuracy: theory and experiment.
Brain topography 4(2):95–103
Cuffin BN, Schomer DL, Ives JR, Blume H (2001a) Experimental tests of eeg source
localization accuracy in realistically shaped head models. Clinical neurophysiology
112(12):2288–2292
Cuffin BN, Schomer DL, Ives JR, Blume H (2001b) Experimental tests of eeg source local-
ization accuracy in spherical head models. Clinical neurophysiology 112(1):46–51
Fuchs M, Wagner M, Wischmann HA, Köhler T, Theißen A, Drenckhahn R, Buchner H
(1998) Improving source reconstructions by combining bioelectric and biomagnetic data.
Clinical Neurophysiology 107(2):93–111
Götz T, Huonker R, Witte OW, Haueisen J (2014) Thalamocortical impulse propagation and
information transfer in eeg and meg. Journal Of Clinical Neurophysiology 31(3):253–260
Hämäläinen M, Hari R, Ilmoniemi RJ, Knuutila J, Lounasmaa OV (1993) Magnetoen-
cephalography — theory, instrumentation, and applications to invasive studies of the
working human brain. Reviews of Modern Physics 65:413–498
18 A. Rezaei et al.

Hari R, Puce A (2017) MEG-EEG Primer. Oxford University Press

Hari R, Baillet S, Barnes G, Burgess R, Forss N, Gross J, Hämäläinen M, Jensen O, Kakigi
R, Mauguière F, et al. (2018) Ifcn-endorsed practical guidelines for clinical magnetoen-
cephalography (MEG). Clinical Neurophysiology
Haueisen J, Tuch D, Ramon C, Schimpf P, Wedeen V, George J, Belliveau J (2002) The
influence of brain tissue anisotropy on human EEG and MEG. NeuroImage 15:159–166
Haueisen J, Leistritz L, Süsse T, Curio G, Witte H (2007) Identifying mutual information
transfer in the brain with differential-algebraic modeling: evidence for fast oscillatory
coupling between cortical somatosensory areas 3b and 1. NeuroImage 37(1):130–136
He Q, Rezaei A, Pursiainen S (2018) Zeffiro user interface for electromagnetic brain imag-
ing: a gpu accelerated fem tool for forward and inverse computations in matlab. arXiv
preprint arXiv:181107717
Kaipio JP, Somersalo E (2004) Statistical and Computational Methods for Inverse Problems.
Springer, Berlin
Lucka F, Pursiainen S, Burger M, Wolters CH (2012) Hierarchical bayesian inference
for the EEG inverse problem using realistic FE head models: Depth localization
and source separation for focal primary currents. NeuroImage 61(4):1364–1382, DOI
10.1016/j.neuroimage.2012.04.017
Miinalainen T, Rezaei A, Us D, Nüßing A, Engwer C, Wolters CH, Pursiainen S (2019)
A realistic, accurate and fast source modeling approach for the eeg forward problem.
NeuroImage 184:56–67
de Munck J, Wolters CH, Clerc M (2012) EEG & MEG forward modeling. In: Brette R,
Destexhe A (eds) Handbook of Neural Activity Measurement, Cambridge University
Press, New York, DOI 10.1017/CBO9780511979958.006
Nagarajan SS, Portniaguine O, Hwang D, Johnson C, Sekihara K (2006) Controlled support
MEG imaging. NeuroImage 33(3):878–885
Niedermeyer E, da Silva FL (2004) Electroencephalography: Basic Principles, Clinical Appli-
cations, and Related Fields, Fifth Edition. Lippincott Williams & Wilkins, Philadelphia
O’Hagan A, Forster JJ (2004) Kendall’s advanced theory of statistics, volume 2B: Bayesian
inference, vol 2. Arnold
Pursiainen S (2012) Raviart–Thomas-type sources adapted to applied EEG and MEG: im-
plementation and results. Inverse Problems 28(6):065,013
Pursiainen S, Lucka F, Wolters CH (2012) Complete electrode model in EEG: relationship
and differences to the point electrode model. Phys Med Biol 57(4):999–1017
Pursiainen S, Vorwerk J, Wolters C (2016) Electroencephalography (EEG) forward model-
ing via H(div) finite element sources with focal interpolation. Physics in Medicine and
Biology 61(24):8502–8520, DOI 10.1088/0031-9155/61/24/8502
Ramon C, Schimpf P, Haueisen J (2006) Influence of head models on EEG simulations
and inverse source localizations. BioMedical Engineering OnLine 5(10), DOI 10.1186/
1475-925X-5-10
Rullmann M, Anwander A, Dannhauer M, Warfield SK, Duffy FH, Wolters CH (2009)
Eeg source analysis of epileptiform activity using a 1 mm anisotropic hexahedra finite
element head model. NeuroImage 44(2):399–410
Sato Ma, Yoshioka T, Kajihara S, Toyama K, Goda N, Doya K, Kawato M (2004) Hierar-
chical bayesian estimation for MEG inverse problem. NeuroImage 23(3):806–826
Schmidt DM, George JS, Wood CC (1999) Bayesian inference applied to the electromagnetic
inverse problem. Human Brain Mapping 7(3):195–212
Tarkiainen A, Liljeström M, Seppä M, Salmelin R (2003) The 3d topography of MEG source
localization accuracy: effects of conductor model and noise. Clinical Neurophysiology
114(10):1977–1992
Uutela K, Hämäläinen M, Somersalo E (1999) Visualization of magnetoencephalographic
data using minimum current estimates. NeuroImage 10(2):173–180
Vorwerk J, Cho JH, Rampp S, Hamer H, Knösche TR, Wolters CH (2014) A guideline for
head volume conductor modeling in EEG and MEG. NeuroImage 100:590–607
Wang G, Yang L, Worrell G, He B (2009) The relationship between conductivity uncer-
tainties and eeg source localization accuracy. In: Engineering in Medicine and Biology
Society, 2009. EMBC 2009. Annual International Conference of the IEEE, IEEE, pp
4799–4802