README.md

In liupeng2117/SubgroupBoost: Value function Guided Subgrouping via Gradient Tree Boosting

SubgroupBoost

This tutorial guides users through using our package SubgroupBoost to identify treatment performing and non-performing subgroups.

Installation

Install the released version of SubgroupBoost from GitHub with:

devtools::install_github("liupeng2117/SubgroupBoost")

Data input and preprocessing

library(SubgroupBoost)

• Load data. In this tutorial, we will use a simulated data that comes with SubgroupBoost package for demonstration.

data("simdata")
names(simdata)
str(simdata$train)
str(simdata$test)

• The data is a list containing 4 elements: training data, testing data, true subgroup label for training data, and true subgroup label for testing data. The training data contains 600 samples, and testing data contains 5000 samples. The data columns has 26 variables: trt01p is treatment indicator. evnt1 and aval1 are for the primary outcome overall survival (OS). evnt1 is the indicator of event (death) for OS, and aval1 is the time value. evnt2 and aval2 are for the secondary outcome time-to-progression (TTP), where evnt2 is the indicator of event (disease progression), and aval2 is the time value. Variable s1 is the predictive variable that interact with treatment, while variables z1 to z20 are not related to treatment effect.

• Preprocess the training and testing data

#----- data manipulation-----#
#training data
dat.train=simdata[[1]][,-c(1:5)]
info.train=simdata[[1]][,1:5]
dtrain <- xgb.DMatrix(as.matrix(dat.train),label = info.train$trt01p)

#testing data
dat.test=simdata[[2]][,-c(1:5)]
info.test=simdata[[2]][,1:5]
dtest <- xgb.DMatrix(as.matrix(dat.test),label = info.test$trt01p)

RMST

• Use RMST based gradient tree boosting method to identify subgroups of patients. Here we choose OS as outcome, as RMST based method can only take one survival outcome. Note you can update xgboost parameters eta and max_depth in SubgroupBoost.RMST. The default is to search among eta = c(.005, .01),max_depth = c(2,4,6) via CV to find the optimal values.

#----- RMST death -----#
dat1=data.frame(dat.train,trt01p=info.train$trt01p, evnt=info.train$evnt1, aval=info.train$aval1)
set.seed(123)
model_RMST<-SubgroupBoost.RMST(dat1)

#check variables selected
importance=xgb.importance(model_RMST$feature_names, model_RMST)$Feature
importance

# [1] "s1"  "z18" "z20" "z9"  "z6"  "z4"  "z17" "z19" "z16" "z7"  "z13" "z11" "z2"  "z1"  "z12" "z15" "z14" "z3"

# Get predicted value for each sample in training data
pred_RMST<-predict(model_RMST, dtrain)
# Calcualate probabilities of belonging to treatment performing subgroup by sigmoid transformation
prob_RMST<-1/(1+exp(-pred_RMST))
#plot the resulting probabilities
hist(prob_RMST)

#Wrap it up to be a function, use probability 0.5 as cutoff for subgroup label
SubgroupBoost.predict<-function(model, data){
  # Get predicted value of model
  pred<-predict(model, data)
  # Calcualate probabilities by sigmoid transformation
  prob<-1/(1+exp(-pred))
  # Use 0.5 cutoff to seperate samples into two subgroups
  assign<-ifelse(prob>0.5,1,0)
  return(group_assign=assign)
}

#2 by 2 table of predicted and true subgroup label for training and testing data
RMST.predict.train<-SubgroupBoost.predict(model_RMST, dtrain)
table(Predicted=RMST.predict.train, True=simdata[[3]])

#         True
#Predicted   0   1
#        0 278  76
#        1  12 234

RMST.predict.test<-SubgroupBoost.predict(model_RMST, dtest)
table(Predicted=RMST.predict.test, True=simdata[[4]])

#         True
#Predicted    0    1
#        0 2319  511
#        1  162 2008

Win-difference

• Use win-difference based gradient tree boosting method to identify subgroups of patients. Here we take both OS and TTP as outcomes for win-ratio, OS surives as the primary endpoint and TTP as the secondary outcome, i.e., surrogate of OS when OS is censored.

#----- win-ratio ----#
#Specify the type of outcome by argument comparison. For two survival outcomes, set it equals "survival survival". 
#Note: Make sure the Primary outcome is defined in variable evnt1 and aval1, and secondary outcome is defined by columns evnt2 and aval2 in info.train object, for two survival outcomes case. You can update xgboost parameters eta and max_depth in SubgroupBoost.wd. The default is to search among eta = c(.005, .01),max_depth = c(2,4,6) via CV to find the optimal values.
comparison="survival survival"
set.seed(123)
model_wd<-SubgroupBoost.wd(dat.train, info.train, comparison)

#check variables selected
importance=xgb.importance(model_wd$feature_names, model_wd)$Feature
importance

# [1] "s1"  "z18" "z20" "z13" "z19"

#2 by 2 table of predicted and true subgroup label for training and testing data
wd.predict.train<-SubgroupBoost.predict(model_wd, dtrain)
table(Predicted=wd.predict.train, True=simdata[[3]])

#         True
#Predicted   0   1
#        0 287  76
#        1   3 234

wd.predict.test<-SubgroupBoost.predict(model_wd, dtest)
table(Predicted=wd.predict.test, True=simdata[[4]])

#         True
#Predicted    0    1
#        0 2436  475
#        1   45 2044

liupeng2117/SubgroupBoost documentation built on Feb. 7, 2022, 1 p.m.