PSP-class {ePCR}

Penalized Single Predictor (PSP) S4-class as a member of PEP-ensembles

Description

PSP is a single penalized Cox regression model, where an alpha/lambda grid has been optimized using cross-validation and a chosen prediction metric. PSPs are single entities that will compile together into PEPs, the ensemble objects that will average over multiple PSPs to generate an ensemble prediction. Typically a single PSP models a part of the data, such as a cohort strata.

Slots

description

A general user-provided string describing the PSP

features

A character vector indicating feature names

strata

Information whether data matrix x included substrata (will be used in plotting functions etc)

alphaseq

The sequence of alpha values to test, ranging between [0,1]; alpha = 0 being ridge regression, 0 < alpha < 1 being elastic net and alpha = 1 being LASSO

cvfolds

The number of cross-validation folds to utilize; by default 10

nlambda

The amount of lambda values utilized in each regularization path; by default 100 as in glmnet-package

cvmean

A matrix indicating the mean CV performance in alpha/lambda grid (preferred over median)

cvmedian

A matrix indicating the median CV performance in alpha/lambda grid

cvstdev

A matrix indicating the standard deviation in CV performance over the folds in the alpha/lambda grid

cvmin

A matrix indicating minimum CV performance in alpha/lambda grid

cvmax

A matrix indicating maximum CV performance in alpha/lambda grid

score

The scoring function, user-defined or one provided by ePCR package such as score.cindex or score.iAUC

cvrepeat

Number of cross-validation procedures to run multiple times and then average over, in order to reduce the effect of binning samples

impute

The imputation function used if provided matrix 'x' includes missing values; by default the impute.knn-function from BioConductor package 'impute'

optimum

The optimum in alpha/lambda grid, with optimal alpha and similarly for lambda

seed

The initial random seed used for cross-validation

x

The input data matrix

x.expand

A function that allows expansion of matrix 'x' to include interactions between variables; if no such are desired, this should be an identity function

y

The Surv-object as in survival-package, which serves as the response y

fit

The glmnet coxnet-object obtained with optimal alpha

criterion

The optimizing criterion; by default "min" for minimizing CV-error

dictionary

A list of discriptions for each variable

regAUC

A numeric vector for the AUC under regularization curve as computed by integrateRegCurve-function

Examples

# As an example, illustrate a naive PSP built on the small medication cohort
data(TYKSSIMU)
library(survival)
# Minimal example with much fewer patients and variables
psp_ex <- new("PSP", alphaseq=c(0.2, 0.8), nlambda=20, folds=3,
	x = xMEDISIMU[1:80,c(1:20,40:50)], y = yMEDISIMU[1:80,"surv"],
seeds = 1, score=score.cindex)

plot(psp_ex) # Optimization surface of alpha/lambda

# Illustrate the use of some PSP-methods:
PSP.KM(psp_ex, cutoff = 0.5) # Kaplan-Meier
PSP.PCA(psp_ex) # PCA plot of training data
PSP.BOX(psp_ex) # Boxplots, here for the first training variable
PSP.CSP(psp_ex) # Cumulative survival probabilities for the training data
invisible(PSP.NA(psp_ex)) # Time-to-event Nelson-Aalen heuristic algorithm

## Not run: 
# Computationally intensive novel PSP-fitting is omitted from the test runs
# Functions for readily fitted PSP-objects are illustrated above
data(TYKSSIMU)
library(survival)
psp_meditext <- new("PSP", x = rbind(xMEDISIMU, xTEXTSIMU), 
y = Surv(rbind(yMEDISIMU, yTEXTSIMU)[,"surv"]),
plot = TRUE, alphaseq = seq(0, 1, by=.01), scorefunc = score.cindex, 
seed = 1, folds = 10, nlambda = 100)
plot(psp_meditext)

## End(Not run)