Simulating structural connectivity changes based on functional
connectivity patterns in schizophrenia
Thesis MSc Computational Science
November 1, 2019
Author:
Lotte van der Wilt
Daily supervisor:
Yongbin Wei
Assessor:
Dr. Martijn P. van den Heuvel
Examinator:
Dr. Rick Quax
Abstract
Schizophrenia is a severe mental health disorder that is hypothesised to originate from
altered neural connectivity patterns. Much is still unknown about the mechanisms
un-derlying schizophrenia, and the link between structural connectivity (SC) and functional
connectivity (FC) changes.
In this study, the schizophrenic brain is simulated using
computational modelling techniques. Structural and functional networks of 74 healthy
individuals and 52 schizophrenia patients are reconstructed based on DTI and fMRI data.
The Spatial Auto-Regression (SAR) model is implemented to predict FC based on SC,
after which an Evolutionary Algorithm (EA) is implemented to evolve the SC of healthy
control individuals to a simulated SC, of which the SAR-predicted FC resembles the real
FC of schizophrenia patients. The resulting simulated structural networks of patients
showed an increase in strength, clustering coefficient and local efficiency, and a decrease
in mean shortest path length compared to that of controls. The effects were most
promi-nent in the frontal and temporal regions. Furthermore, the rich club nodes showed a
relative decrease in strength, betweenness centrality, clustering coefficient and local
effi-ciency and a decrease in mean path length in simulated patient SC, suggesting a decrease
in the central role of rich club regions in schizophrenia. Finally, the results suggested a
decrease in SC-FC coupling. EAs may provide a valuable method to gain more insight
into the structural and functional connectivity alterations underlying the disease, and in
understanding how the disease progresses over time.
Table of contents
1
Introduction
7
Box 1: Functional and Structural connectivity . . . .
9
Box 2: Graph Theory . . . .
10
2
Methods
11
2.1
Subjects . . . .
11
2.2
Imaging data . . . .
11
2.2.1
Structural connectivity
. . . .
11
2.2.2
Functional connectivity . . . .
11
2.3
Experimental setup . . . .
12
2.4
Spatial Auto-Regression (SAR) model . . . .
13
2.5
Evolutionary algorithm
. . . .
13
2.5.1
Initial population
. . . .
14
2.5.2
Parent selection . . . .
14
2.5.3
Offspring creation . . . .
15
2.5.4
Termination conditions . . . .
15
2.6
Parameter tuning . . . .
16
2.6.1
Simulated Annealing (SA)
. . . .
16
2.7
Experimental design . . . .
16
2.8
Data analysis
. . . .
17
2.8.1
Real structural connectivity . . . .
17
2.8.2
Simulated structural connectivity . . . .
17
3
Results
19
3.1
Real structural connectivity . . . .
19
3.2
Simulated structural connectivity . . . .
19
3.2.1
Fitness . . . .
19
3.2.2
Simulated SC . . . .
19
3.2.3
Rich club
. . . .
21
3.2.4
Atrophy . . . .
22
4
Discussion
23
Acknowledgements
25
References
27
Appendix A: Parameter tuning
31
Appendix B: Simulated SC thresholds
33
Appendix C: Data tables
37
1
Introduction
Schizophrenia is a chronic, severe mental health disorder, characterised by abnormal behaviour and
thought patterns, loss of affect, hallucinations and delusions. The disease affects roughly one in 200
people with significant impact on patients’ quality of life. Schizophrenia has been argued to be a
dis-ease of dysconnectivity since Carl Wernicke (1848-1905) first proposed his ‘sejunction hypothesis’ that
the loosening of association fibres in the brain, which he called sejunction, was the underlying cause
of all psychotic disorders [1][2]. Eugen Bleuler (1857-1939), who named the disease “schizophrenia”
(referring to the splitting (skhizein) of the mind (phren) [3]), also spoke of a ‘loosening of associations’
and hypothesised that the symptoms were caused by disturbances in integration within the brain [4][5].
Their ideas not only strongly influenced the way we currently view schizophrenia, but also impacted the
field of psychiatry as a whole [6]. It was not until the 1970s that modern brain imaging techniques
started to arise.
Brain imaging studies have indeed shown altered brain connectivity patterns in schizophrenia patients.
Brain connectivity can be defined as functional connectivity (FC), showing the synchronisation in brain
activity, or structural connectivity (SC), which represents the anatomical connections in the brain. For
a detailed explanation of FC and SC, see Box 1. Both functional and structural connectivity have been
found to be affected in schizophrenia patients [4]. The region that shows the most consistent decrease in
functional connectivity is the frontal cortex, specifically, the prefrontal cortex [7][8]. The frontal-temporal
connections were most affected, followed by the frontoparietal links [8]. Nonetheless, a few studies have
shown increases in functional connectivity [9][10][11]. The exact cause of such contradictory results is
unknown, but methodological differences and symptomatic variations may play a role [4]. Similar to
the functional alterations, the frontal and temporal lobe show the most stable decrease in structural
connectivity when the fractional anisotropy is measured [12][13]. A study using white matter
tractogra-phy shows the most significant alterations in the white matter tracts attached to frontal cortex as well,
including areas with both increased and decreased SC [14]. Further evidence of the dysconnectivity in
schizophrenia were given by studies in the field of connectomics, where the brain is studied as a network
[15], and network organisation and integration are focused on rather than the function of specific brain
areas or single interareal connectivity.
A brain network consists of nodes and edges. Depending on the scale of the network, network nodes
can represent single neurons or entire brain areas, whereas the edges can stand for connections, for
example, the synapses connecting two neurons or white matter tracts connecting different brain areas.
More information on how these networks are analysed using graph-theoretical analysis can be found
in Box 2. Connectome studies have presented evidence of decreased network efficiency [16] and less
centralised frontal and parietal hubs in schizophrenia patients [17]. The impaired rich club organization,
which describes a group of densely interconnected hub regions with many long-range connections that
play an important role in whole-brain communication and integration, was also observed in
schizophre-nia [18][7][19]. A recent research involving medication-na¨ıve first-episode schizophreschizophre-nia patients already
showed altered rich-club connectivity, suggesting this pattern is present at an early stage and is not
mediated by medication [20]. Furthermore, the medication-naive patients showed a decrease in SC-FC
coupling, whereas both increased [7] and decreased [21] SC-FC coupling have been found in chronic
patients.
Much is still unclear about the mechanisms underlying the whole course of schizophrenia, and the
coherence between structural and functional connectivity changes. One of the methods that allows us
to gain new insights into disease-related connectivity changes, as well as the connectivity pattern of the
healthy brain, is computational modelling. Various computational models have been applied in attempts
to predict FC based on the SC. A model that stands out in its simplicity, as well as its accuracy, is
the Spatial Auto-Regression (SAR) model. Although a multitude of more complex models has been
developed, the SAR model tends to (nearly) parallel them in accuracy while the computational costs
are significantly lower [22][23]. It is important to note that no computational model has been able to
explain all the variance contained in real FC matrices, but the models can explain more of the variation
than SC alone. A different computational model that has not been used for the prediction of FC, but
has shown promising results in a wide range of scientific fields, is the Evolutionary Algorithm (EA) [24].
EAs are optimisation algorithms using techniques inspired by genetics and evolution.
In this study, we apply the SAR model in an attempt to simulate the schizophrenic brain. By
do-ing so, we aim to gain new insights into the development of brain network changes in schizophrenia
patients and the coupling between structural and functional alterations. First, structural and functional
networks of 74 healthy individuals and 52 schizophrenia patients are reconstructed based on DTI and
fMRI data. Then, the SAR model is implemented to predict the functional connectivity matrix based
on structure. An Evolutionary Algorithm (EA) is applied to evolve the structural connectivity - starting
with real structural network of healthy participants - towards a simulated SC that results in a predicted
functional connectivity matrix matching the average FC of schizophrenia patients. Finally, the changes
in structural connectivity over time as well as in the final population will be analysed using statistical
methods and graph-theoretical measures. Our study may provide novel insights into neural mechanisms
underlying schizophrenia and prevention and intervention methods.
1: Box 1: Functional and Structural connectivity
Brain connectivity can be estimated via multi-modal neuroimaging techniques. Although multiple measurements
are available, here we focus on the MRI methods that are most often used: Diffusion Tensor Imaging (DTI) to
define the structural connectivity, and functional MRI (fMRI) to define the functional connectivity.
Diffusion Tensor Imaging. DTI is a form of MRI that measures the diffusion of water molecules. In an open
space, the diffusion of water molecules is random, and therefore uniform in all directions.
Some structures
prevent this random movement of water molecules, allowing us to visualise their shape. Axons, the neuronal
structures forming the white matter tracts that connect the different brain areas, are covered in myelin sheaths.
The myelin provides a sort of insulation, playing an essential part in the transmission of signals to other cells
[25]. Myelin sheaths prevent water molecules from freely diffusing through the brain. The stronger the myelin
around the axons, the more restricted particles are in diffusion.
The DTI scan measures the diffusion of water molecules in three dimensions, resulting in a diffusion ellipsoid,
or tensor, for each voxel. Then, the fibres are traced to see which regions are connected, how many streamlines
connect the areas (number of streamlines: NOS), as well as how elongated the ellipsoids are on average (fractional
anisotropy: FA). This elongation of the diffusion ellipsoid, known as fractional anisotropy, shows how strongly
myelinated the fibres are, and is often used as a measure of white matter integrity. The measures mentioned above
can be used as the connection weight in the structural connectivity graph. One disadvantage of this technique is
that we cannot trace the direction of the axon, and therefore, the data is undirectional.
Figure 1: Displays the principles of Diffusion Tensor Imaging (DTI) [26]. A. When diffusion is entirely unrestrained,
diffusion in all directions will be equal: isotropic diffusion. B. In white matter tracts, diffusion is restricted, leading
to anisotropic diffusion tensors. C. By tracing the diffusion tensors in three-dimensional space, the white matter
structure can be derived.
Functional Magnetic Resonance Imaging. A popular imaging method to study the functional connectivity in
the brain is fMRI. In fMRI studies, blood-oxygen-level-dependent (BOLD) imaging is used. The BOLD signal
represents the relative percentages of oxyhemoglobin and deoxyhemoglobin in the blood, showing the ratio of
oxygenated blood compared to deoxygenated blood. As such, BOLD imaging is used as a measure of brain
activity. When a specific area is active, more oxygen is extracted from the blood, resulting in changes in the
local oxygenation level over time.
When the term functional connectivity is used, this usually refers to the
synchronisation in the BOLD signal between different areas in resting-state. The correlation coefficient between
BOLD signals is used as the connectivity weight.
SC-FC coupling. Nearly all studies that investigated the relationship between SC and FC show a significant
correlation between these two ends of connectivity [27]. However, functional connectivity differs over time, as
well as between tasks [28]. Structural connectivity, on the other hand, is much more stationary. Also, functional
connectivity frequently presents itself in the absence of a direct anatomical connection [29][30], which can be at
least partially the result of indirect structural connectivity [28]. In addition, current DTI resolutions might lead to
an underestimation of SC in areas where many fibres cross [22]. A study that investigated the relationship between
structural connectivity and variability in eight functional connectivity measurements taken over the course of one
day, found that the SC matrix limits the same-day variability as well as the strength in FC [31]. Furthermore,
characteristics of individual FC patterns have been shown to be unique and stable over multiple years [32]. Hence,
although the FC is variable and there is no one-to-one SC-FC coupling, both SC and FC are thought to contain
highly valuable information.
2: Box 2: Graph Theory
Graph theory is the mathematical study of network graphs, often used to gain insight into large, complex systems
and their organisational properties. Graph theory is often involved in studying social networks, transportation and
road networks, and even molecular systems. Networks can be classified into different types, namely undirected
and directed, and binary and weighted networks. In an undirected network, edges are bidirectional. When a
network includes one-way connections, the network is directional. Binary networks only differentiate between
connected and unconnected nodes, whereas weighted networks allow for the discrimination between strong and
weak connections.
Figure 2: Displays some of the different network types (A-C), and illustrates various graph-theoretical metrics
(D-F, in blue). A. Binary undirected matrix, B. Binary directed matrix, C. Weighted undirected matrix, D. Shows
a hub with high degree, E. Maximum shortest path between two nodes; eccentricity, F. High clustering.
Graph theory provides us with some general measures to describe networks. The network density represents the
number of connections that are present, relative to the number of connections the graph would have if it were
fully connected. The clustering coefficient is the likelihood that any nodes’ neighbours are interconnected. We
can also calculate the shortest path between any two nodes in the network. This metric is often used as a measure
of network efficiency. The maximum shortest path length for a specific node is called its eccentricity. Highly
related to the path length, are the local- and global efficiency measures, representing the inverse mean shortest
path length of a node’s nearest neighbours and of the whole network, respectively. The betweenness centrality of
a node signifies how often a node is present on the shortest path between any two other nodes. Nodes that have a
high degree - meaning they have many connections – and also have a central position in the network are known as
hubs. When these hubs are more densely interconnected than would be expected based on chance, we speak of a
rich club organisation [33]. In a weighted network, the sum of all the weights of a node is called the node strength.
Graph-theoretical approaches are often used in connectomics research to analyse brain networks. Network measures
suggest that brain networks are shaped by the trade-off between wiring costs and network efficiency [34]. Brain
wiring is expensive. In an area with significant spatial constraints, connections require physical space as well
as energy. The longer the fibres, and the higher the density, the higher the wiring costs are [35]. Therefore,
efficient use of connections is critical to the functioning of the network. Some costly, inter-modular, long-range
connections are required to pass down information efficiently, to increase conduction speed and enable efficient
integration. This efficiency is achieved by a rich club organisation. The rich club provides a backbone for efficient
functional integration [36].
2
Methods
2.1
Subjects
In this study, imaging data from 52 schizophrenia patients (11 females, 41 males, age µ ± SD =
38.17 ± 14.13) and 74 healthy control individuals (20 females, 54 males, age µ ± SD = 37.88 ± 11.91)
between the ages of 18 and 65 were included. The study was approved by the Institutional Review Board
of the University of New Mexico, and all participants provided written informed consent. Individuals
with a history of neurological disorder or severe head trauma or with a history of substance abuse within
the past 12 months were excluded.
2.2
Imaging data
The imaging data that was used in this study originates from the Center for Biomedical Research
Excel-lence (COBRE) [37], and was downloaded from the SchizConnect database (http//schizconnect.org).
The investigators within SchizConnect contributed to the design and implementation of SchizConnect
and/or provided data but did not participate in analysis or writing of this report. All data was recorded
using 3 Tesla Siemens Magnetom Trio scanner (Siemens, Erlangen, Germany).
2.2.1
Structural connectivity
Data aqcuisition
A Diffusion Weighted Imaging protocol was used to record structural connectivity data. A five-echo
MPRAGE T1-weighted image was used, with the following parameters: TR = 2539, TE = [1.64, 3.5,
5.36, 7.22, 9.08ms], TI = 900ms, flip angle = 7
◦
, 1mm isotropic voxel size, FOV = 256x256mm, slab
thickness = 176mm. Additionally, a DWI set was measured including 30 diffusion weighted volumes
(b-value = 800 1000 s/mm2) and 5 diffusion-unweighted volumes equally inter-spread between the 30
diffusion weighted volumes (parameters: TE = 84 ms, 2 mm isotropic voxel size). FLIRT (FMRIB’s
Linear Image Registration Tool) was used for all registration steps [38].
Data preprocessing
The T1-weighted images were processed using FreeSurfer (http://surfer.nmr.mgh.harvard.edu/) [39].
Images were corrected for eddy-current, realigned to the b=0 image, and automatically segmented into
68 cortical regions based on a subdivision of the Desikan-Killiany atlas (DK-68) [40]. Affine
transfor-mation mapping was used to map the individual T1-weighted image to the DWI image, in order to
co-register the parcellation maps. Then, a robust tensor fitting algorithm [41] was applied to fit tensors
to the diffusion signals measured in each voxel, after which FA was calculated [42]. Fiber Assignment
by Continuous Tracking (FACT) [43] was used to reconstruct DTI white matter tracts. Streamline
reconstruction started from eight seeds in every white matter voxel. Streamlines continued unless they
showed high curvature (> 45
◦
), entered a voxel with an FA below 0.1, or exited the brain mask.
Network reconstruction
The streamline reconstructions and parcellation of the cortical structures were integrated for each subject,
resulting in an individual weighted network. If one or more streamlines were reconstructed touching both
regions, the regions were considered to be connected. The number of streamlines between the regions
was used as the edge weight. This process resulted in a 68x68 connectivity matrix for each subject,
showing the connection weight between each combination of nodes.
2.2.2
Functional connectivity
Data aqcuisition
Resting state fMRI data was collected using the following parameters: 33 axial slices, TR = 2000 ms,
TE = 29 ms, flip angle = 75
◦
, slice thickness = 3.5 mm, slice gap = 1.05 mm, acquisition matrix = 64
150 volumes of functional images were obtained. Subjects were asked to keep their eyes open and stare
at a fixation cross.
Data preprocessing
Like the aforementioned preprocessing of the structural imaging data, T1-weighted images were
pro-cessed using FreeSurfer (http://surfer.nmr.mgh.harvard.edu/) [39], and automatically segmented into
68 cortical regions based on a subdivision of the Desikan-Killiany atlas (DK-68) [40]. The resting-state
fMRI time-series were then realigned with the T1 image, and parcellation maps were co-registered.
Lin-ear trends and first order drifts were removed from the time-series, and time-series were corrected for
global effects (regressing out the white matter, ventricle, and global mean signals, as well as 6 motion
parameters) and band-pass filtered (0.01 - 0.1 Hz).
Network reconstruction
For each node, the fMRI voxels overlapping with the region were selected and the time-series of the
selected voxels were averaged. Then, for each combination of nodes, the average regional time-series
were correlated. The Pearson’s correlation coefficient was used as the network weight. Again, this
resulted in a 68x68 connectivity matrix for each subject, showing the connection weight between each
combination of nodes.
2.3
Experimental setup
The experimental design incorporated two main parts, as shown in Figure 3. First, we implemented
the SAR model to predict functional connectivity based on structural connectivity. The model allows
us to compare simulated functional connectivity (FCsim) to the real, measured FC. Then, we applied
an evolutionary algorithm to evolve the input of the SAR model - the real, individual control SC –
towards an SC matrix that creates a simulated FC matching the real FC. Both steps involved parameters
that required tuning (Figure 3). Once the optimal parameter values were found, the model was run
repeatedly, and the resulting simulated structural connectivity matrices were analysed.
2.4
Spatial Auto-Regression (SAR) model
In the SAR model, the activity of a node can be calculated from a linear combination of its SC matrix,
the neighbours’ activity, an auto-regression parameter (k) that determines the synchronisation strength,
and random noise (Equation 1). Although in some situations it can be valuable to simulate activity over
time, in this study we are merely interested in the FC matrix, which can be calculated directly using
Equation 2 and 3. In the equations below, k represents an auto-regression parameter, determining the
synchronisation strength, σ is the noise level, v
i
represents uncorrelated white Gaussian noise with mean
0 and unit variance. D is the structural connectivity matrix, and I is the identity matrix. Finally,
t
is
the matrix transposition. The σ parameter was set to 1, based on earlier research [22][23]. The input
of the SAR model needs to be normalised. In this study row normalisation was applied, meaning that
each value in the matrix was divided by the sum of its row.
y
i
= k
X
j6=i
D
i,j
y
j
+ σv
i
(1)
cov = σ
2
(I − kD)
−1
(I − kD)
−t
(2)
cor
ij
=
cov
ij
√
cov
ii
cov
jj
(3)
2.5
Evolutionary algorithm
The goal of the Evolutionary Algorithm in this study is to evolve the structural connectivity matrix,
the input of the SAR model, towards a simulated structural connectivity matrix that leads to simulated
functional connectivity resembling the real FC of the schizophrenia group.
The idea behind EAs originates from the biological and genetical basis of evolution. The algorithm aims
to find the optimal solution to a problem. Usually, this concerns either the input or the parameters of a
model, while the desired output is known. In this case, the input of the SAR model will be evolved. A
possible set of inputs for the model is known as a genotype, and the resulting model output is called its
phenotype. Each separate input is known as an allele. The model iterates through generations (Figure
4). In each generation, the fitness of all individuals is tested, and parents are selected based on their
fitness values. Offspring are created through crossover, or recombination, and mutation of genotypes.
The new individuals replace the previous population, and this process repeats itself until the termination
conditions are met.
Figure 4: Main components of an evolutionary algorithm.
In this report, the genotypes were the individual SC matrices, and the resulting simulated FC matrices
were their genotypes. Since the connectivity matrices are symmetrical, only the upper triangular matrix
was used, which was represented as an array of floating-point numbers. As a measure of fitness, the
Pearson correlation coefficient between the individual simulated FC (F C
i
) and the population mean
(F C
real
) was calculated. The functional connectivity matrices were normalised before calculating the
mean F C
real
, using the z-score. Since it is customary to minimise the fitness in an EA, where the optimal
value is close to zero, this correlation coefficient was subtracted from 1 to represent the individual fitness
score (Equation 4).
F itness
i
= 1 −
Cov(F C
i
, F C
real
)
σ
F C
iσ
F C
real(4)
2.5.1
Initial population
The real individual structural connectivity matrices were used to create the initial population. This
population was rounded out to the intended number by adding recombinations of random individuals.
2.5.2
Parent selection
The parents were selected using rank-based stochastic uniform sampling (SUS). First, the individuals
were ranked based on fitness, and the rank was used to calculate a scaled fitness value. The scaled fitness
then depended on rank rather than the exact score (Equation 6). This method ensures that selection
is stronger among high-fitness individuals. The difference in scaled fitness is much more significant
between the top individuals than for the lower-ranked individuals. From this scaled fitness, a selection
probability can be calculated using Equation 6.
F itness
scaled
i=
1
√
rank
i
(5)
P (i) =
F itness
scaled
iP
N
j=1
F itness
scaled
j(6)
In SUS, the parents are chosen based on their respective probabilities. All parents are chosen at once,
using only one random number. This method ensures that, although there is a random effect, the
parents composing the next generation are a good representation of the scaled fitness distribution.
Once all selection probabilities are calculated, a parent sample can be picked from the population. A
cumulative probability distribution is created. The parents are selected at equal distances (d =
N
1
p
) from
the cumulative probability distribution, where N
p
is the number of parents. The distance d is multiplied
by a uniform random number U [0, 1] to choose the first parent. Figure 5 and Table 1 illustrate this
principle for an example population with six individuals, with probabilities P
i
, as shown in the table.
Even with these small sample sizes, the samples roughly correspond to the scaled fitness distribution in
the population.
Figure 5: Stochastic Universal Sampling. On top, the cumulative
proba-bility distribution is shown. Based on a random uniform number, the first
position is chosen. Consequently, 12 random samples are drawn.
Table 1: Shows the probabilities used
in Figure 5, as well as the resulting
sam-ple rates.
Rank
F
scaledP
iSample
1
1
∼0.27
0.3
2
∼0.71
∼0.19
0.2
3
∼0.58
∼0.16
0.2
4
0.5
∼0.14
0.2
5
∼0.45
∼0.12
0.1
6
∼0.41
∼0.11
0.2
2.5.3
Offspring creation
Offspring were created through three different methods: elitism, mutation and crossover.
First, the elitism parameter determined the amount of best individuals that were directly copied to the
next generation. Elitism is meant to ensure that the fit-test individual in the population is always kept.
The remaining population was created through either recombination or mutation. The percentage of the
population that was created by a crossover in proportion to mutation was determined by a parameter
called crossoverF raction.
Secondly, a uniform mutation algorithm was used. A random sample of edges was chosen that were
mutated. This sample was created based on the mutationRate: the probability that each of the alleles
is mutated. Each of the chosen edges was replaced by a random uniform number between 0 and the
individual’s maximum value.
Finally, intermediate crossover was implemented as a recombination method (Equation 7). An essential
parameter in this formula is ratio. The ratio determines the range around the parent values that a child
can adopt. When the ratio is 1, the child’s connection weight always lies between the parents’ weights.
When the ratio is 1.5, the probabilities of the child’s weight falling within or outside the parents’ range
are equal(Figure 6). A ratio within this range ([1,1.5]) is frequently used since this allows exploration
of new possible solutions.
child = p
1
+ U [0, 1] ∗ ratio ∗ (p
2
− p
1
)
(7)
Figure 6: Intermediate crossover.
2.5.4
Termination conditions
Evolutionary Algorithm runs were terminated when there had been no improvement for 100 generations,
or the maximum amount of generations (maxGenerations) was reached.
Table 2: Evolutionary Algorithm parameter tuning settings. Shows the initial values as well as the bounds of the search
space.
Initial value
Lower bound
Upper bound
grid search
populationSize
-
75
150
maxGenerations
-
100
600
eliteCount
-
1
5
SA
mutationRate
0.01
0.0005
0.02
crossoverF raction
0.8
0.6
1.0
2.6
Parameter tuning
As previously discussed, two algorithms used in this study required parameter tuning: the SAR model
and the Evolutionary Algorithm. First, the SAR parameter was optimised to maximise the correlation
between the average individual SC and FC. As there was only one parameter, a simple grid search was
performed. Initial broad tuning revealed the optimal value should be between 0.9 and 0.99. Within
that range, all values were checked with a step-size of 0.005. The Evolutionary Algorithm, on the other
hand, provided six parameters to be tuned: the population size, the maximum amount of generations
and the number of elite individuals, and also the crossover fraction, mutation rate and ratio. The first
three, all categorical, were explored using a grid search. The last three parameters were tuned using
a Simulated Annealing algorithm. Table 2 shows the different parameters and their respective ranges.
Final parameter settings and SA results are shown in Appendix A.
2.6.1
Simulated Annealing (SA)
A simulated annealing algorithm was implemented to find the optimal parameter settings for the
evolu-tionary algorithm. Simulated annealing is a method inspired by the process of annealing: the controlled
heating and cooling of metal to create the optimal density and strength. The algorithm is mainly used
to find approximate global optima in vast, multidimensional search spaces. Considering the running time
of the EA and the number of continuous parameters to be tuned, this specific algorithm was chosen.
Simulated annealing is an iterative algorithm, starting with an initial set of parameter values. At each
iteration, the fitness value is calculated. The fitness, in this case, is the best fitness reached in an
Evolu-tionary Algorithm run with that set of parameters. Then, a new set of parameter values is chosen from
a ‘neighbouring’ range. The size of this range is determined by the temperature (T ). At each iteration,
a step of length T is taken in a random uniform direction. If this new set of parameters has a better
fitness than the previous solution, it is selected as the current solution. If the fitness is low, there is
still a small probability the parameter set is accepted (Equation 8). The lower the temperature, and the
more substantial the decrease in fitness, the smaller the likelihood of acceptance. The selection of lower
fitness values, albeit with a lower probability, prevents the model from getting stuck in local optima.
The temperature gradually decreases for a set amount of generations (reannealInterval), after which
it ‘re-anneals’: the temperature is reset to a higher value. The temperature, aside from re-annealing
steps, is determined by Equation 9. In this equation, g is the number of generations since (re-)annealing.
The initial parameter settings, as well as the possible ranges that were used, are shown in Table 2.
P
acceptance
=
1
1 − exp(
max(T )
∆
)
(8)
T = T
0
∗ 0.95
g
(9)
2.7
Experimental design
The Evolutionary Algorithm was used in 4 different conditions (Table 3). This study aimed to evolve
the controls’ individual SC matrices towards an SC that allowed the predicted FC to match the real
patients’ FC. This was the experimental condition. Moreover, three control conditions were used. Most
importantly, the individual control SC was evolved towards the control functional connectivity, to be able
to differentiate between changes that emerged from optimally fitting to the SAR model and changes
that directly stemmed from the schizophrenia related alterations in patient FC. Then, both of these
experiments were repeated starting with the individual patient SC instead of the control SC.
For each of these conditions, the EA was executed 100 times. The simulated SC and fitness of all
individuals were collected for each generation in every run.
Table 3: Shows the different test conditions that were used to run the Evolutionary Algorithm.
Initial population
Reference FC
Experimental condition
Control SC
Patient FC
Control condition 1
Control SC
Control FC
Control condition 2
Patient SC
Patient FC
Control condition 3
Patient SC
Control FC
2.8
Data analysis
2.8.1
Real structural connectivity
The real structural connectivity matrices of the control and patient group were compared, based on
a set of graph theoretical metrics: the strength, mean shortest path length, betweenness centrality,
clustering coefficient, and local- and global efficiency. Differences between SC of patients and controls
were compared using t-tests, with False Discovery Rate (FDR) correction.
Graph theoretical measures
For each node i, the strength was calculated: the sum of weights of all edges connecting to node i.
Furthermore, the shortest path between node i and every other node was measured. The betweenness
centrality represents the amount of times node i is present on the shortest path between any two other
nodes in the network. The clustering coefficient, which describes the probability that any two neighbours
of node i are interconnected, was calculated by dividing the number of edges in the neighbourhood of
node i by the maximum possible amount of edges (
d
1
i