In EmilyLMorris/ImagingPred: Code for performing prediction of functional connectivity using brain imaging data

ImagingPred

This package performs three main functions.

1. Reads multiple NIfT1 files and converts them into a dataframe with a summary of each region of interst.

2. Performs network estimation using the R package huge - currently implemented with the Meinshausen-Buhlmann option for undirected graph estimation.

3. Performs variable selection on covariates provided to predict connectivity at each edge of the network.

Installation

In order to install the package, several R packages are needed. Some functions rely on oro.nifti, neurobase, huge, glmnet, e1071, and randomForest.

The following line of code installs the package from GitHub:

library(devtools)
devtools::install_github("EmilyLMorris/ImagingPred")

Example 1

Here is a simple example for each of the functions. First read in a NIfTI file, this example may take a few minutes to run.

library(oro.nifti)
library(neurobase)
set.seed(2020)
# generate an example nifti file
img = array(rnorm(100*100*100*120), dim = c(100,100,100,120))
nifti.img = oro.nifti::nifti(img) # generate an example nifti file
nifti.img
writenii(nifti.img, "example1") 
img1 = read.nifti.files("example1")
network1 = network.estimate(img1)
table(network1)
# repeat this for multiple "subjects"
for(n in 1:10){
  set.seed(2020 + n)
  img = array(rnorm(100*100*100*120), dim = c(100,100,100,120))
  nifti.img = oro.nifti::nifti(img) # generate an example nifti file
  writenii(nifti.img, paste("example", n+1, sep = "")) 
}

# function read.nifti.files reads the .nii files and summarizes into 264 regions of interest
file.names = list.files(getwd())[grep(".nii", list.files(getwd()))]
img.summary = read.nifti.files(file.names)
length(img.summary)
dim(img.summary[[1]])

Function used to generate network estimate for each of the subjects:

networks.list = network.estimate(img.summary)

Finally we demonstrate how to use the function, prediction.analysis, in two ways to perform variable selection and to predict connectivity.

# generate covsariate matrix 
n = length(networks.list)
X <- data.frame(matrix(rnorm(n*10), nrow = n), subj.id = file.names)

results.ex1 = prediction.analysis(networks = networks.list, # network estimates for each subject
                               covariates = X, # covariate matrix
                               method = "SVM", # can choose between SVM and random forest 
                               missing.prop = 0, # can specify a threshold to remove covariates if missing for too many subjects
                               ID = NULL, # specify subject ID to match networks to covariates 
                               newdata = NULL) 
results.ex1$predicted
results.ex1$variables_selected

Example 2

Example using simulated data with some true signal:

load("example_pkg.RData")

results.ex2 = prediction.analysis(networks = networks.list, 
                               covariates = data.frame(X), 
                               method = "SVM", 
                               missing.prop = 0, 
                               ID = names(networks.list), 
                               newdata = NULL) 
dim(results.ex2$predicted)
length(which(!is.na(results.ex2$variables_selected)))
results.ex2$index_edges # use to compare to true connections in simulated data 

EmilyLMorris/ImagingPred documentation built on June 19, 2021, 9:33 p.m.