MIS581

Pima Indians Diabetes Data Set Classification Tree

Sys.time()

Load data mining packages

library(rpart) library(rpart.plot) library(caret)

Required to get data but I don't know why it's needed to load the data

if(!require('mlbench')) { install.packages('mlbench') library('mlbench') }

Get cleaned data

PimaIndiansDiabetes2 removes zeros that were not possible

Replaced with NA

data(PimaIndiansDiabetes2) #Cleaned data set from R-Project.org summary(PimaIndiansDiabetes2) #Summary of data set

Create a data frame

Diabetes.df <- PimaIndiansDiabetes2 Diabetesdf2 <- na.omit(Diabetes.df) #Remove NAs to fix issues with models head(Diabetesdf2) summary(Diabetesdf2) summary(Diabetesdf2$diabetes) #Number of participants without/with diabetes Sys.Date()

Create Correlation Matrix

install.packages("corrplot") library(corrplot) install.packages("Hmisc") library("Hmisc") Sys.Date() Diabetes.cor <- round(cor(Diabetesdf2[1:8]),1) Diabetes.cor

Partition Preparation

set.seed(1) #To reproduce the same output of simulation studies

Partition Data Set

train.index <- sample(c(1:dim(Diabetesdf2)[1]), dim(Diabetesdf2)[1]0.6) #Sample data frame test.index <- sample(c(1:dim(Diabetesdf2)[1]), dim(Diabetes.df)[1]0.4) #Sample data frame

New Partition

require(caTools) set.seed(3) sample = sample.split(Diabetesdf2$diabetes, SplitRatio = 0.75) train= subset(Diabetesdf2, sample==TRUE) test= subset(Diabetesdf2, sample== FALSE)

Training Partition

train.df <- Diabetesdf2[train.index,]

Testing Partition

test.df <- Diabetesdf2[test.index,]

Create Logistic Model for Diabetes with Training Data

Train.Model <- glm(train$diabetes ~ ., family = binomial(link = "logit"), data = train) Sys.Date() summary(Train.Model) plot(Train.Model)

Create Logistic Model for Diabetes with Test.df

Sys.Date() Test.model <- glm(test$diabetes ~ ., family = binomial(link = "logit"), data = test) summary(Test.model) View(Test.model)

Graph of Logistic Model using Test Partition

library(ggplot2) plot(Test.Model)

Prediction statistics

Train.pred <- predict(Train.Model, type = "link") Sys.Date() summary(Train.pred) Test.pred <- predict(Test.Model, type = "link") summary(Test.pred)

ROC Curve###WOULDN'T WORK

install.packages("ROCR") library("ROCR") Train.ROCR = prediction(train$diabetes, Diabetesdf2) Train.perf = performance(Train.ROCR, "tpr", "fpr")

Confusion matrix

table(test$diabetes, Test.pred > 0.5)

Classification Tree

library(rpart) library(rpart.plot)

Tree using Train Data

Diabetes.Tree.Train <- rpart(train$diabetes ~., data = train, method = "class", minbucket = 10) prp(Diabetes.Tree.Train)##view tree

Tree using Test Data

Diabetes.Tree <- rpart(test$diabetes ~ ., data = test, method = "class", minbucket = 5) Diabetes.Tree prp(Diabetes.Tree)## View tree

Evaluate Tree

Train.Tree.pred <- predict(Diabetes.Tree.Train, newdata = train, type = "class") table(Train.Tree.pred)

Sys.Date() Test.Tree.pred <- predict(Diabetes.Tree, newdata = test, type = "class")

table(Test.Tree.pred) Sys.Date() table(train$diabetes, Train.Tree.pred) sensitivity.Train <- round(72/(26+72),2) specificity.Train <- round(178/(178+18),2) sprintf("Sensitivity at 0.7 threshold: %s", sensitivity.Train)
sprintf("Specificity at 0.7 threshold: %s", specificity.Train)

table(test$diabetes, Test.Tree.pred) sensitivity.Test <- round(24/(24+8),2) specificity.Test <- round(59/(59+7),2) sprintf("Sensitivity at 0.7 threshold: %s", sensitivity.Test)
sprintf("Specificity at 0.7 threshold: %s", specificity.Test)