PEP-class {ePCR}

Penalized Ensemble Predictor (PEP) S4-class ensemble consisting of individual PSP-members

Description

This class constructs an ensemble of individual Penalized Ensemble Predictor (PSP-class) members. Each member contributes to the model output equally, and ensemble-level functions wrap up individual predictions into an averaged ensemble prediction. The user may define an arbitrary number of PSPs and tailor them to suit the particular needs, and then provide them as a list to the PEP-constructor. As such, constructing well tailored individual ensemble members (of PSP-class) in order to produce a powerful ensemble (of PEP-class) is important on both levels.

Slots

PSPs

List of PSP-objects that will be treated as equal members of the ensemble

description

A character string describing the structure or purpose of the ensemble

features

A character list of variable/feature names

dictionary

A named list of above variables/features and their more precise description

predens

A function for compiling all predictions from the PSPs into consensus prediction

prednorm

A function for normalizing the predictions e.g. to risk scores in [0,1]

Examples

## Not run: 
# The PEP-construction is wrapped in NOT RUN, because cross-validating multiple PSPs
# is very time consuming especially if a tight grid of alpha/lambda is to be explored.
# The simulated data from Turku University Hospital (TYKS) is used as an example:
data(TYKSSIMU)

# Two cohorts and corresponding data matrices:
head(xMEDISIMU)
head(xTEXTSIMU)
# Two survival responses:
head(yMEDISIMU)
head(xTEXTSIMU)

# Search L1/L2 norm alpha-grid with 10 values between [0,1]
aseq <- seq(from=0, to=1, by=0.1)
# Lambda sequence penalization is of 100 length conditional for each alpha
nlamb <- 100

library(survival)
# Create three ensemble members; one for MEDI cohort, one for TEXT cohort,
# and finally one member that combines both cohorts simultaneously in a coxnet
psp1 <- new("PSP", x = rbind(xMEDISIMU, xTEXTSIMU), 
y = Surv(rbind(yMEDISIMU, yTEXTSIMU)[,"surv"]),
plot = TRUE, alphaseq = aseq, scorefunc = score.cindex, seed = 1,
folds = 10, nlambda = nlamb)
psp2 <- new("PSP", x = xMEDISIMU, 
y = Surv(yMEDISIMU[,"surv"]),
plot = TRUE, alphaseq = aseq, scorefunc = score.cindex, seed = 1,
folds = 10, nlambda = nlamb)
psp3 <- new("PSP", x = xTEXTSIMU, 
	y = Surv(yTEXTSIMU[,"surv"]),
plot = TRUE, alphaseq = aseq, scorefunc = score.cindex, seed = 1,
folds = 10, nlambda = nlamb)
par(mfrow=c(1,3))
plot(psp1); plot(psp2); plot(psp3); # Inspect the alpha/lambda surfaces

# Create an ensemble of the above 3 members
simuens <- new("PEP", PSPs = list(psp1, psp2, psp3))
simuens
# Ready PEP-object can be used for novel predictions etc


## End(Not run)

# Run example predictions from a previously optimized PEP-model
data(ePCRmodels)
data(TYKSSIMU)

# Perform risk predictions from the joint cohort ensemble member as an example
MEDIpred <- predict(TYKS@PSPs[[1]]@fit, s=TYKS@PSPs[[1]]@optimum["Lambda"], 
newx = conforminput(TYKS@PSPs[[1]], xMEDISIMU))[,1]
TEXTpred <- predict(TYKS@PSPs[[1]]@fit, s=TYKS@PSPs[[1]]@optimum["Lambda"], 
newx = conforminput(TYKS@PSPs[[1]], xTEXTSIMU))[,1]

# Risk scores obtained for the new patients (arbitrary unit as per Cox regression)
head(MEDIpred)
head(TEXTpred)