% IMPORTANT: The following is UTF-8 encoded. This means that in the presence
% of non-ASCII characters, it will not work with BibTeX 0.99 or older.
% Instead, you should use an up-to-date BibTeX implementation like “bibtex8” or
% “biber”.
@INPROCEEDINGS{Wischnewski:903158,
author = {Wischnewski, Kevin and Eickhoff, Simon and Popovych,
Oleksandr},
title = {{E}fficient validation of dynamical whole-brain models via
mathematical optimization algorithms},
school = {Heinrich Heine University Düsseldorf},
reportid = {FZJ-2021-04881},
year = {2021},
abstract = {INTRODUCTION: Investigating the resting-state brain
dynamics involves its simulation by dynamical whole-brain
models that have attracted a great interest over the last
years [1]. With a rising number of utilized mathematical
models and their complexity, the challenge of adequate model
fitting to empirical data has become apparent. An intuitive
approach is to optimize model parameters by a grid search.
However, an exploration of the entire parameter space on a
dense grid becomes computationally expensive for models with
many free parameters. In search of alternatives,
mathematical optimization algorithms have received increased
attention [2]. These methods can outperform the grid search
approach regarding computation time and result quality. In
this work, we investigate several such optimization schemes
for the validation of whole-brain models and compare their
outcomes as well as computational costs with each other and
the grid search. We suggest two most efficient algorithms
for the optimization of the correspondence between simulated
and empirical data. METHODS: Neuroimaging data of 105
subjects of the Human Connectome Project [3] were used for
the extraction of structural and resting-state functional
connectivity (SC and FC, respectively). Schaefer’s
functional atlas [4] with 100 cortical regions was chosen as
a brain parcellation. Additionally, an MRtrix-based
probabilistic tractography [5] was applied to compute
atlas-based SC. It was used to determine the coupling
weights and delays between units of the dynamical model of
coupled phase oscillators [6]. The model was deployed to
simulate resting-state brain dynamics and eventually
generate simulated FC (sFC). This in turn was fitted to
empirical FC (eFC) by simultaneously adjusting up to 3 model
parameters: global coupling and delay (2Dim parameter space)
as well as noise intensity (3Dim parameter space). In
addition to the grid search, the following derivative-free
methods were tested: Nelder Mead Algorithm (NMA, [7]),
Particle Swarm Optimization (PSO, [8]), Covariance Matrix
Adaptation Evolution Strategy (CMAES, [9]) and Bayesian
Optimization (BO, [10]). NMA is a deterministic local search
method, the others represent global stochastic approaches.
PSO and CMAES share the feature of being population-based.
RESULTS: For all tested algorithms, the detected
goodness-of-fit values range from $-5\%$ to $+11\%$ around
those found by the grid search. Among the methods, the order
PSO > CMAES > BO > NMA in respect of the goodness-of-fit
(larger is better) can be observed. The values of the best
fit in the 3Dim parameter space are on average around $16\%$
higher than those obtained in the 2Dim case, regardless of
applied algorithm. We found that the algorithms reliably
detect global maxima in the parameter space, where the
spread of solutions was tested for multiple random initial
conditions. This effect is most pronounced for CMAES, while
NMA demonstrates a high susceptibility to local optima.
Comparing the computation time required by the investigated
methods to the grid search $(100\%)$ yielded relative values
from $3500\%$ (PSO) to $87\%$ (NMA) in 2Dim and from $73\%$
(PSO) to $2\%$ (BO) in 3Dim. To evaluate methods, we
analyzed a cost function that included the goodness-of-fit
values, spread of algorithm solutions as well as the
required computation time. We identified CMAES and BO as the
most efficient approaches that may be used for the
validation of dynamical models. CONCLUSIONS: We showed that
some available mathematical optimization algorithms can be
used as an efficient tool to improve and accelerate the
search for the parameters that maximize the similarity
between empirical and simulated data. The tested methods
perform a continuous search and do not rely on a selected
grid granularity in order to detect optimal solutions. Thus,
our findings may contribute to a more efficient validation
of complex models with high-dimensional parameter spaces and
facilitate precise and personalized modeling of brain
dynamics. REFERENCES: [1] doi:10.3389/fnsys.2018.00068, [2]
doi:10.1007/s12021-018-9369-x, [3]
doi:10.1016/j.neuroimage.2013.05.041, [4]
doi:10.1093/cercor/bhx179, [5]
doi:10.1016/j.neuroimage.2019.116137, [6]
doi:10.1016/j.neuroimage.2011.04.010, [7]
doi:10.1093/comjnl/7.4.308, [8]
doi:10.1007/s11721-007-0002-0, [9]
doi:10.1145/2330784.2330919, [10]
doi:10.1287/educ.2018.0188},
month = {Jun},
date = {2021-06-21},
organization = {The 27th Annual Meeting of the
Organization for Human Brain Mapping,
Virtual (Virtual), 21 Jun 2021 - 25 Jun
2021},
subtyp = {After Call},
cin = {INM-7},
cid = {I:(DE-Juel1)INM-7-20090406},
pnm = {5232 - Computational Principles (POF4-523) / HBP SGA2 -
Human Brain Project Specific Grant Agreement 2 (785907) /
HBP SGA3 - Human Brain Project Specific Grant Agreement 3
(945539) / VirtualBrainCloud - Personalized Recommendations
for Neurodegenerative Disease (826421)},
pid = {G:(DE-HGF)POF4-5232 / G:(EU-Grant)785907 /
G:(EU-Grant)945539 / G:(EU-Grant)826421},
typ = {PUB:(DE-HGF)24},
url = {https://juser.fz-juelich.de/record/903158},
}