RTGP

Code for Bayesian scalar-on-image regression with a relaxed-thresholded Gaussian process prior.

Data

Abstract

• Simulation study:

  • Volumetric-based representation: 2D images generated via data generating process described in Section C.3 of appendix in thesis.
  • Surface-based representation: Subsampled surface images generated via data generating process described in Section 4.3.1.1 in thesis.

• ABCD study: Surface-based representation of test statistic values from first-level fMRI analysis of emotional n-back task with 2 vs 0 back contrast.

• X: NxM matrix: vectorised images of test statistics of task fMRI

• X_confounds: NxP matrix: confounding variables

• y: Nx1 vector: intelligence scores

Availability

• Simulation study: Data can be generated via simulation study script '.../simulation_study/volume/data_circle.R' for the volumetric-based simulation study and '.../simulation_study/surface/data_generation.R' for the surface-based simulation study. We additionally provide an example simulation study containing 10 datasets (= 2D images with circle as effect of interest in the volume-based setting and sub-sampled surface maps in the surface-based setting) generated with the setting of sample size N=500 and SNR = 1/0.2 to showcase the reproducibility of our method.
• ABCD study: Healthcare data requires access permission and cannot be attached here (https://abcdstudy.org/).

Description

Simulated 2D volumetric images provided in this repositry under .../simulation_study/volume/N500_poly10_signal1_noise0.2_effect124/, simulated surface images provided in this repositry under .../simulation_study/surface/N500_L100_phi0.077_nu2/, and surface cifti files for COPE and standard deviation of COPE from first-level analysis provided from the ABCD study (not uploaded due to access restrictions).

Code

Abstract

Code to run all baseline models, such as the frequentist approaches, Ridge regression and LASSO regression, the Bayesian regression models that take all the spatial locations as inputs, BR + Normal and BR + Horseshoe, and the Gaussian process regressions, GPR + Normal and GPR + Horseshoe, and our models RTGP-VI and RTGP-Gibbs. Additionally, we provide all code to generate figures & tables.

Description

Simualtion study (volume):

• data_circle.R : Generates simulation study data.
• simulation_study_volume.R : Performs parameter estimation and inference for models, such as RTGP-VI, RTGP-Gibbs, STGP, and GPR + Normal / Horeshoe.

Simulation study (surface):

• data_generation.R :
• distance_matrix_left.py :
• distance_matrix_right.py :
• eigendecomposition_left_sphere.R :
• eigendecomposition_right_sphere.R :
• evaluate_datasets.R :
• evaluate_simulation_study_surface_ELBO.R :
• evaluate_simulation_study_surface_hyperparameters.R :
• evaluation_data_removal.R :
• simstudy_surface.R :
• simstudy_surface_data_removal.R :
• simstudy_surface_sensitivity_variance_hyperparameters.R :

ABCD study:

• BR_Normal_BR_Horseshoe.R :
• GPR_Normal_GPR_Horseshoe.R :
• RTGP_VI_left_hemisphere.R :
• RTGP_VI_right_hemisphere.R :
• RTGP_VI_left_hemisphere_CV.R :
• RTGP_VI_right_hemisphere_CV.R :
• Ridge_LASSO.R :
• analysis_eigendecomposition.R :
• brainsmash_distance_matrix_left_hemisphere.py :
• brainsmash_distance_matrix_right_hemisphere.py :
• data_preprocessing_step1.R :
• data_preprocessing_step2.R :
• eigendecomposition_left_cortex_L800_CV.R :
• eigendecomposition_right_cortex_L800_CV.R :
• evaluate_CV.R :
• evaluation_GPR.R :
• evaluation_RTGP.R :
• massunivariate_analysis.R :

Optional Information

R libraries:

The following R packages are necessary to successfully use the codes:

• MASS
• BayesGPfit
• parallel
• fields
• ggplot2
• reshape2
• gridExtra
• truncnorm
• fastBayesReg
• STGP
• ggpointdensity
• xtable
• data.table
• ciftiTools
• bayestestR

Additional:

• Software package: Connectome Workbench (https://www.humanconnectome.org/software/connectome-workbench)

Instructions for Use

Reproducibility

Code:

• Simulation studies: example case provided for N=500 and SNR = 1/0.2 (all other studies can also be reproduced by generating additional datasets with 'data_circle.R' or 'data_generation.R')
• UK Biobank application: Code to run ABCD study application where input data is concatenated into a NxM matrix from cifti files, where N represents sample size and M is the number of vertices, and save the output as a csv files and then in cifti files.

Creation of figures and tables in the main paper and supplementary material:

Paper:

• Fig. 1: gp_stgp_htgp_rtgp.R
• Fig. 2: data_generation.R
• Fig. 3: simstudy_surface.R
• Fig. 4: simstudy_surface.R
• Fig. 5: Created via Powerpoint.
• Fig. 6: (a) - (b) evaluation_RTGP.R, (c) - (d) GPR_Normal_GPR_Horseshoe.R, (e) - (f) BR_Normal_BR_Horseshoe.R, (g) - (h) Ridge_LASSO.R
• Fig. 7: (a) - (b) evaluation_RTGP.R, (c) - (d) GPR_Normal_GPR_Horseshoe.R
• Tab. 1,2,3: simstudy_surface.R
• Tab. 5: evaluation_RTGP.R

Replication (Optional)