TPLSr (version 1.0.3)

TPLS_cv: Fit a TPLS model to data with cross validation

Description

Fit a TPLS model to data with cross validation

Usage

TPLS_cv(X, Y, foldid, NComp = 50, W = 0)

Arguments

X

n-by-v data matrix of real numbers. Rows correspond to observations (trials) and columns to variables (e.g., fMRI voxels).

Y

n-by-1 Vector of real numbers. Can be binary (0/1) for classification model, or can be continuous.

foldid

A vector of values between 1 and number of folds identifying what fold each observation is in.

NComp

Maximum number of partial least squares component you want to use. Default is 50, and this is on the safe side for fMRI.

W

n-by-1 vector of positive observation weights.

Value

A TPLS_cv object that contains the following attributes. Most of the time, you won't need to access the attributes.

• NComp: The number of components you specified in the input

• numfold: Total number of cross-validation folds

• testfold: A vector indice that should be the same as foldid, if it was provided accurately.

• cvMdls : A vector of TPLS models, one for each fold.

Examples

# NOT RUN {
# Fit example TPLS data with a TPLS model using cross-validation
# Load example data (included with package).
X = TPLSdat$X # single trial brain image of subjects pressing left/right buttons
Y = TPLSdat$Y # binary variable that is 1 if right button is pushed, 0 if left button is pushed
subj = TPLSdat$subj # 1, 2, or 3, depending on who the subject is
run = TPLSdat$run # 1, 2, ..., 8, depending on the scan run of each subject

# Fit the model, using 3-fold cross-validation at the subject level
# (i.e., train on two subjects, test on 1, repeat three times)
TPLScvmdl <- TPLS_cv(X,Y,subj)

# Evaluate the tuning parameters via cross-validation.
# We'll test 1~50 components and thresholding from 0 to 1 in 0.05 increments.
# Also include subfold information.
# This allows for calculation of correlation at the run-level instead of at the subject level.
cvstats <- evalTuningParam(TPLScvmdl,"pearson",X,Y,1:50,seq(0,1,0.05),subfold=run)

# plot the tuning parameter surface.
# It'll show the point of best performance (and also point of 1SE performance).
# The plot is interactive, so spin it around
plotTuningSurface(cvstats)

# These are the tuning parameters of best performance
cvstats$compval_best # 8 components
cvstats$threshval_best # 0.1 thresholding (leave only 10% of all voxels)

# Now build a new TPLS model, using all the data, using the best tuning parameters
TPLSmdl <- TPLS(X,Y,NComp=cvstats$compval_best)

# Extract the prediction betamap that gave the best CV performance
betamap <- makePredictor(TPLSmdl,cvstats$compval_best,cvstats$threshval_best)

# This is the intercept
betamap$bias

# These are the coefficients for the original variables
betamap$betamap

# Project the betamap into brain-space so that we can look at it.
mask = TPLSdat$mask # mask 3D image of the brain from which X was extracted from
brainimg = mask*1 # make a copy
brainimg[mask] = betamap$betamap # put the betamap into the brain image
fig1 <- plot_ly(z = brainimg[,15,], type = "heatmap") # looking at a slice of the brain image
fig2 <- plot_ly(z = 1*mask[,15,], type = "heatmap") # a slice of the brain mask for reference
fig <- subplot(fig1, fig2)
fig

# Figures show a coronal section of the brain (but flipped right 90 degrees).
# on the left, you should see the bilateral motor cortex coefficients with opposing signs.
# This is just a simple visual demonstration. You should use other packages to output
# coefficients into a nifti file and view them in a separate viewer.
# }