% 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{Domhof:907813,
author = {Domhof, Justin and Eickhoff, Simon and Popovych, Oleksandr},
title = {{R}eliability and subject specificity of personalized
dynamical whole-brain models},
reportid = {FZJ-2022-02230},
year = {2022},
abstract = {- Introduction - Dynamical whole-brain models originally
provided a biophysically-inspired approach to investigate
the relationship between the structural (SC) and functional
(FC) connectivity (Honey 2009). Nowadays, they are also used
to study the dynamical regimes of the brain and how these
relate to various subject traits (Bansal 2018).
Nevertheless, it is unclear how the modeling results perform
in terms of test-retest reliability and subject specificity
as measured through, e.g., identification accuracies. Here,
we systematically assess these aspects of the modeling
results, and examine how they relate to the reliability and
subject specificity of empirical data.- Methods - We used
the empirical SC and FC matrices of 200 healthy unrelated
subjects (96 males, aged 28.5 ± 3.5 years) from the Human
Connectome Project (Van Essen, 2012, 2013). The models were
built on the basis of the empirical SC matrices of
individuals and were either a network of neural mass models
(Wilson, 1972) or a system of coupled phase oscillators
(Kuramoto, 1984). In order to estimate how model
personalization could contribute to the subject specificity,
the latter model used region-specific natural frequencies
extracted from empirical data that were either subject
specific or the same for all subjects. The models were
simulated for a broad range of parameter settings to yield
the simulated FC matrices exhibiting the highest correlation
with the empirical FCs. Four empirical FC matrices (2
phase-encoding directions scanned on 2 days) and,
correspondingly, four simulated FC matrices were available
per subject for further analyses. We evaluated the
within-subject correlations of both types of FCs (empirical
and simulated), which served as proxies for their
reliability. Additionally, we calculated the between-subject
correlations, and used the difference between the intra- and
inter-subject correlations (specificity index) as a
characterization of the subject specificity. Finally, we
adapted the fingerprinting analysis from Finn et al. (2015)
to provide an additional measure for the subject specificity
of the FCs.- Results - The results show that the reliability
of the simulated FC can exceed that of the empirical one
(Fig. 1A), especially, for the structural atlases and for
the phase oscillator model regardless of the strategy with
respect to the natural frequencies (Fig. 1A, orange and
red), Also, the subject specificity of the simulated FC may
outperform that of the empirical one (Fig. 1B-C). Here
again, the phase oscillator model with subject-specific
frequencies generated FCs with a much higher subject
specificity than the other, less personalized modeling
paradigms (Fig. 1B-C, red).In addition, the atlas has a
larger influence on the reliability of the simulated FC than
on that of the empirical FC (Fig. 1A). There seems to be a
clear distinction between structurally- and
functionally-derived atlases, which result in more and less
reliable simulated FCs, respectively (Fig. 1A, left vs.
right block). Analogously, a change of parcellation affected
the subject specificities of the simulated FC much more than
those of the empirical FC (Fig. 1B-C). In particular, the
atlas may determine whether the subject specificities of the
empirical and simulated FC are at about the same or
different levels (Fig. 1B-C).- Conclusions - Our results
showed that the reliability and the subject specificity can
be higher for the simulated than for the empirical FC. We
also demonstrated the critical roles that the parcellation
and model implementation have on the findings. Taken
together, our results indicate that whole-brain dynamical
models can generate simulated connectomes with high
reliability and (subject) specificity and may outperform the
empirical data in this respect. In turn, this suggests that
these models potentially reduce the variance in the
empirical FC across different realizations for a single
subject by providing a reliable model fit for further
analyses.},
month = {Jun},
date = {2022-06-07},
organization = {The 28th Annual Meeting of the
Organization for Human Brain Mapping,
Virtual (Virtual), 7 Jun 2022 - 8 Jun
2022},
subtyp = {After Call},
cin = {INM-7},
cid = {I:(DE-Juel1)INM-7-20090406},
pnm = {5232 - Computational Principles (POF4-523)},
pid = {G:(DE-HGF)POF4-5232},
typ = {PUB:(DE-HGF)24},
url = {https://juser.fz-juelich.de/record/907813},
}