Mixed-effects multilevel analysis (MEMA) & Canonical correlation analysis (CCA) in fMRI study

• Guide for an analytical framework with MEMA followed by CCA in leave-one-subject-out cross-validation (LOOCV)
• To find the evidence of association between behavioral data and neuronal activations from fMRI data based on heterogeniety across subjects

Sample data:

for MEMA:

• Blood-oxygenation-level-dependent(BOLD) fMRI volumes were acquired while the subjects were smoking with the MR-competible e-cigarette apparatus (ECIG; with nicotine) in MRI scanner
• Preprocessed BOLD fMRI data were analyzed using a general linear model (GLM) for an individual-level analysis (using “3dREMLfit” for least squares restricted maximum likelihood estimation in AFNI)

for CCA:

• brain_response : The z-scored coefficient (beta) of the condition (ECIG) from the first-level analysis (GLM)
  • The dataset consists of 18 subjects x 7 voxels (beta values of the center voxel and its six neighboring voxels within the right insula identified from MEMA)
• behavior_data : Three behavioral data (i.e., Similarity, Urge-to-smoke, and Smoking duration) normalized between zero to one.
  • All of subjects scored two behavioral measurements (Similarity to their own e-cigarette in terms of nicotine absorption and the level of Urge-to-smoke) after smoking each type of MR-competible e-cigarette apparatus and the Smoking duration was recorded
  • The dataset consists of 18 subjects x 3 behavior data

MEMA:

• fMRI group analysis that incorporates both the variability across subjects and the precision estimate of each effect of interest from individual subject anlyses
  • Chen, Gang, et al. "FMRI group analysis combining effect estimates and their variances." Neuroimage 60.1 (2012): 747-765. https://doi.org/10.1016/j.neuroimage.2011.12.060

Requirement:

• Install AFNI and R software on your own desktop/laptop: https://afni.nimh.nih.gov/pub/dist/doc/htmldoc/background_install/install_instructs/index.html

Download sample data:

• Please download the sample data from this link: http://bspl.korea.ac.kr/Research_data/MEMA-CCA/MEMA_dir.zip

Data structure:

The sample dataset consists of two folders:

• beta: NIfTI files (S_E_beta.nii) of 18 subjects' individual beta value in whole brain
• tscore: NIfTI files (S_E_tscore.nii) of 18 subjects' individual Student's t-test in whole brain The preprocessed BOLD fMRI were analyzed with General Linear Model (GLM) for an individual level analysis

Shell script:

• Open a terminal, navigate to the MEMA_dir directory,

cd [MEMA_dir] 

• Edit the directory and set the prefix for your data in MEMA_script.tcsh
• Run the shell script

tcsh MEMA_script_bspl.tcsh 

• The results saved in results folder as MEMA_ecig: MEMA_ecig+tlrc, MEMA_ecig_ICC+tlrc, MEMA_ecig_resZ+tlrc

Note : please change the prefix if necessary.

• Open AFNI software in order to check the results

afni MEMA_ecig+tlrc.HEAD

CCA in leave-one-subject-out cross-validation:

• To test the statistical significance of the relationship between neuronal activations from fMRI and behavioral data
• In the validation set (n=1), canonical variates and their corresponding canonical correlations were created using coefficients from the CCA performed in the training set (n=17)
  • Dinga, Richard, et al. "Evaluating the evidence for biotypes of depression: Methodological replication and extension of." NeuroImage: Clinical 22 (2019): 101796. https://doi.org/10.1016/j.nicl.2019.101796

Requirement:

• Install MATLAB (>2014a) on your own PC

Download sample data:

• Please download the sample dataset from this link: http://bspl.korea.ac.kr/Research_data/MEMA-CCA/cca_loocv_dataset.mat

Data structure:

The input '.mat' file includes:

• 'brain_response' is the z-scored coefficients of seven voxels from subjects' beta values from the GLM within the ROI (right insula) identified from high heterogeniety (chi-square < 10^-8) (18x7)
• 'behavior_data' is the 0 to 1 normalized scores of three behavior data (18x3)
  [Similarity, Urge-to-smoke, Smoking duration]

MATLAB code:

• 'cca_loocv_bspl.m' is the code for CCA in LOOCV
• 'linear_reg_bspl.m' is the code for the linear regression to investigate the relationship between two variables
• Run the 'test_cca_loocv.m'
• Then, three figures will be appeared to show scatter plots as well as the predicted regression line for each of the following scenarios:
  1. Between the pair of canonical variates for the two variables
  2. Between the canonical variate of the behavioral data and each of behavioral data
     [Similarity, Urge-to-smoke, Smoking duration]
  3. Between the canonical variate of the brain response and each of behavioral data
     [Similarity, Urge-to-smoke, Smoking duration]