Motion Effects in Procrustes Aligned Individual-Level Gradients

1724 

Abstract Submission 

Leonard Sasse¹, Casey Paquola², Simon Eickhoff¹, Kaustubh Patil²

¹Institute of Systems Neuroscience, Medical Faculty and University Hospital Düsseldorf, Düsseldorf, Germany, ²Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich, Jülich, Germany

Leonard Sasse  
Institute of Systems Neuroscience, Medical Faculty and University Hospital Düsseldorf
Düsseldorf, Germany

Casey Paquola, Dr.  
Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich
Jülich, Germany

Simon Eickhoff  
Institute of Systems Neuroscience, Medical Faculty and University Hospital Düsseldorf
Düsseldorf, Germany

Kaustubh Patil  
Institute of Neuroscience and Medicine, Brain and Behaviour (INM-7), Research Centre Jülich
Jülich, Germany

Functional connectivity (FC) is a cornerstone of fMRI research for unraveling human brain organization (Biswal et al. 1995). One widely employed technique involves the extraction of FC gradients from FC matrices using dimensionality reduction (Margulies et al. 2016; Vos de Wael et al. 2020). However, as the sign and the ordering of gradients can be different between subjects, they need to be aligned to make them comparable. Procrustes alignment is often employed for this purpose (Vos de Wael et al. 2020). This study aims to investigate the impact of varying the number of gradients used in Procrustes alignment on the aligned principal gradient, and its use in subject-level downstream analyses.

We used rs-fMRI data from the Human Connectome Project (HCP) for 395 unrelated subjects (192 female, 203 male), aged 22-37 (M=28.71, SD=3.82) (Van Essen et al. 2013). Specifically, we used data that has undergone the HCP's ICA-FIX procedure, which also included removal of Friston 24 motion parameters (Glasser et al. 2013; Salimi-Khorshidi et al. 2014). In addition, we regressed out white matter (WM), cerebro-spinal fluid (CSF), and global (GS) signals, their squared terms, and temporal derivatives. The data was bandpass filtered at 0.01 - 0.08 Hz and aggregated using the Schaefer 400 parcellation. FC was calculated as Pearson's correlation. Gradients were extracted using the BrainSpace toolbox, which also provides Procrustes alignment to find the optimal transformation matrix to minimise the sum of squared errors of a source matrix to a reference matrix (Fig. 1a). We split subjects into an analysis and a holdout reference dataset for alignment (Fig. 1b). We performed identification (Finn et al. 2015) and calculated differential identifiability (Amico and Goñi 2018) to test the impact of the alignment on downstream subject-level analyses. To quantify similarity of the aligned principal gradient with the unaligned gradients, we calculated the ratio of the maximum to the sum of absolute values in the first column of the transformation matrix ("Correspondence"; Fig. 1c). Motion was characterized using average framewise displacement (FD) for each subject.

Identification accuracy substantially increased (Fig. 2a), whereas differential identifiability decreased (Fig. 2b) when using more gradients for alignment. In addition, correspondence of the aligned gradient decreased with increasing number of gradients, indicating that the principal gradient takes on an increasingly mixed character, including more information from lower gradients (Fig. 2c). This result suggests that higher distinctiveness of individuals is not due to simply re-ordering of the gradients, but rather due to global subject-specific information being introduced into the aligned principal gradient. One candidate for such subject-specific global information is motion. We found that average FD was correlated with the magnitude of the transformation (Fig. 2d-e), and this correlation increased with increasing number of gradients used in alignment. Additionally, motion signal as indicated by typicality of the FC matrices (Kopal et al., 2020) was correlated to the total transformation (Fig. 2f). Finally, using more gradients in alignment improved prediction of FD based on the aligned principal gradient, indicating an increased presence of motion signal after alignment (Fig. 2g-i).

We observed that the number of components used for Procrustes alignment plays a substantial role in downstream analyses. Specifically, we found that as more gradients are utilized in alignment, identification accuracy increases while differential identifiability decreases. Additionally, our results indicate that Procrustes alignment, especially when using many gradients, introduces information from lower-order information into the principal gradient. Notably, we identified that motion effects are introduced in the alignment procedure.

Classification and Predictive Modeling ²

Connectivity (eg. functional, effective, structural)

Exploratory Modeling and Artifact Removal

fMRI Connectivity and Network Modeling ¹

Methods Development

Data analysis

Design and Analysis

FUNCTIONAL MRI

Machine Learning

Statistical Methods