Amount and Localisation of Non-Linearity in rs-fMRI Functional Connectivity
Poster No:
1544
Submission Type:
Abstract Submission
Authors:
Giulio Tani Raffaelli1, Jaroslav Hlinka1, Jakub Kopal2
Institutions:
1Institute of Computer Science of the Czech Academy of Sciences, Prague, Czech Republic, 2McGill University, Montreal, Quebec
First Author:
Giulio Tani Raffaelli
Institute of Computer Science of the Czech Academy of Sciences
Prague, Czech Republic
Institute of Computer Science of the Czech Academy of Sciences
Prague, Czech Republic
Co-Author(s):
Jaroslav Hlinka
Institute of Computer Science of the Czech Academy of Sciences
Prague, Czech Republic
Institute of Computer Science of the Czech Academy of Sciences
Prague, Czech Republic
Introduction:
The last 20 years have shown a growth in interest towards research and clinical applications for Functional Connectivity (FC). Along with the interest in FC, there is increasing attention to alternatives to Pearson's correlation in the connectivity estimate. While the correlation accounts only for linear relationships between regions or electrodes, a common alternative is Mutual Information (MI), which accounts for non-linearities. As the application of MI increases, despite a previous study (Hlinka, 2011) pointing at a marginal role of non-linearity, new evidence shows there might be more (Motlaghian, 2022). We follow up on Hlinka's study and leverage a twenty times larger cohort to quantify and localise the non-linearity in rs-fMRI BOLD signal and assess the dependency on the atlas.
Methods:
In this study, we use rs-fMRI BOLD data from 242 healthy subjects from the publicly available dataset associated with (Kopal, 2020). From this, we selected data subject to stringent preprocessing. We chose parcellation with the AAL atlas with 90 regions and the Craddock atlas with 10 to 950 regions. To ensure that the observed non-linearity comes from the relationship between time series and not from their marginal distribution, we applied a monotonic transformation to ensure normally distributed marginals. We computed the total MI via the equiquantal binning method and corrected for its bias. To evaluate the amount of non-linearity, we generated 99 Fourier-transform multivariate surrogates, preserving the covariance matrix of the data. For each subject, we evaluate the amount of MI not justified by correlation as the relative difference between all regions' total MI and the average MI in surrogates. We then assess non-linearities' localisation by checking for regions whose total MI, across all subjects, was significantly higher than surrogate MI. We controlled for bias in the whole non-linearity estimation by repeating the entire analysis on a shadow dataset generated by surrogation of the original one.
Results:
We obtain a consistent presence of non-linearity for the different region sizes from the Craddock atlas. The amount of MI not explained by the covariance (Fig. 1) sits stable at around 4% for all region sizes except for very small (single voxel) and large regions. At the same time, the non-linearity observed in the shadow dataset remains lower than 1%.
We observe that regions with relationships with MI consistently higher than surrogates are grouped primarily on the occipital lobe (Fig. 2). For comparison, the number of such connections in the shadow dataset is lower by two-thirds, and the non-linearities localisation is more sparse.
We observe that regions with relationships with MI consistently higher than surrogates are grouped primarily on the occipital lobe (Fig. 2). For comparison, the number of such connections in the shadow dataset is lower by two-thirds, and the non-linearities localisation is more sparse.
·Fig. 1, Relative amount of non-linearity in the brain varying the number of regions in the Craddock atlas. For single voxel results, only 100 voxels are considered.
·Fig. 2, For each region we show the number of other regions with whom the mutual information significantly exceeds what explained by a linear relationship.
Conclusions:
This study confirms that the non-linearity in the rs-fMRI BOLD signal is measurable above the bias in the estimation. Furthermore, we show that non-linearities are consistent, independent of the atlas of choice. This study also shows the localisation of such non-linearities and agrees with the independent findings from (Motlaghian, 2022). At the same time, the non-linearity is minor, even in the most non-linear regions. Observing it reliably at the individual level will present a future challenge due to the effect of noise.
Modeling and Analysis Methods:
Connectivity (eg. functional, effective, structural) 1
Novel Imaging Acquisition Methods:
BOLD fMRI 2
Keywords:
Data analysis
FUNCTIONAL MRI
Other - Mutual Information; Functional Connectivity; Non-Linearity; BOLD
1|2Indicates the priority used for review
Provide references using author date format
Kopal, J. (2020), ‘Typicality of functional connectivity robustly captures motion artifacts in rs‐fMRI across datasets, atlases, and preprocessing pipelines’, Human Brain Mapping, vol. 41, no. 18, pp. 5325-5340.
Motlaghian, S. M. (2022), ‘Nonlinear functional network connectivity in resting functional magnetic resonance imaging data’, Human brain mapping, vol. 43, no. 15, pp. 4556-4566.