Recreating_Article_Code.Rmd

---
title: "Code_Recreation"
author: "Shaan Malhotra"
date: "2023-03-14"
output: html_document
---

```{r setup, include=FALSE}
knitr::opts_chunk$set(echo = TRUE)
```

## Code Reproduced From the Article

Lets install the necessary packages:
```{r}
library(dplyr)
library(caret)
library(ggplot2)
library(imputeTS)
library(caTools)
library(e1071)
library(nnet)
library(rpart)
library(ggpubr)
library("randomForest")
library(factoextra)
```

Load in the dataset. 
```{r}
input_data <- read.csv("~/Desktop/healthcare-dataset-stroke-data.csv")
```

Omit the rows that contain NA
```{r}
input_data<- na.omit(input_data)
```

Display the dataset
```{r}
colnames(input_data)
```

These columns were removed in the article's code so I did the same
```{r}
input_features = input_data[,2:11]
data_without_id = input_data[,2:12]
data_without_id <- subset(data_without_id, select = -bmi)
data_without_id <- subset(data_without_id, select = -work_type)
data_without_id <- data_without_id[!(data_without_id$gender == "Other"), ]
```

Certain columns were converted to factors so they could be used in the model.
```{r}
input_features$gender <- as.factor(input_features$gender)
input_features$ever_married <- as.factor(input_features$ever_married)
input_features$work_type <- as.factor(input_features$work_type)
input_features$Residence_type <- as.factor(input_features$Residence_type)
input_features$smoking_status <- as.factor(input_features$smoking_status)
```

Store the "cleaned" data into a matrix
```{r}
input_features[] <- data.matrix(input_features)
```

Setting the variables for Random Downsampling 
```{r}
minority_class <- data_without_id[data_without_id$stroke == 1,]
majority_class <- data_without_id[data_without_id$stroke == 0,]
```

Code for the neural network model:
```{r}
no_of_exps <-1000

nnetwork_result = c()

for (i in 1:no_of_exps) {
  cat("Current experiment: ", i)
  
  
  majority_sample <- majority_class[sample(nrow(majority_class), 548), ]
  balanced_dataset <- rbind(minority_class, majority_sample)
  
  split = sample.split(balanced_dataset$stroke, SplitRatio = 0.70)
  train_set = subset(balanced_dataset, split == TRUE)
  test_set = subset(balanced_dataset, split == FALSE)
  
  nnetm <- train(as.factor(stroke) ~., data=train_set, method='nnet')
  print(nnetm)
  plot(nnetm)
  pred_nn <- predict(nnetm, newdata=test_set)
  
  # Output labels
  out_labels<-as.data.frame(test_set[, 9])
  out_labels<-t(out_labels)
  
  cm_nn = table(out_labels, pred_nn)
  
  #accuracy
  n_nn = sum(cm_nn)
  diag_nn = diag(cm_nn)
  accuracy_nn = sum(diag_nn) / n_nn
  accuracy_nn
  
  
  nnetwork_result[length(nnetwork_result)+1] = accuracy_nn
}
```

Display the accuracy of the neural network model
```{r}
accuracy_nn
```

Storing the neural network predictions as a new column on the data set and re-labeling it so it can be used with the Aequitas tool:

```{r}
neural_network_data_raw <- cbind(test_set, pred_nn)

neural_network_data <- neural_network_data_raw
neural_network_data <- subset(neural_network_data, select = -c(hypertension, heart_disease, avg_glucose_level, ever_married, smoking_status))

colnames(neural_network_data)[4] ="score"
colnames(neural_network_data)[5] ="label_value"

#remove the comment from the following line to save the data as a csv so it can be put into Aequitas 
#write.csv(neural_network_data, "~/Desktop//neural_network.csv", row.names=FALSE)
```



Code for the Decision Tree model:
```{r}
dtree_result = c()

for (i in 1:no_of_exps) {
  cat("Current experiment: ", i)
  
  
  majority_sample <- majority_class[sample(nrow(majority_class), 548), ]
  balanced_dataset <- rbind(minority_class, majority_sample)
  
  split = sample.split(balanced_dataset$stroke, SplitRatio = 0.70)
  train_set = subset(balanced_dataset, split == TRUE)
  test_set = subset(balanced_dataset, split == FALSE)
  

  train_set$stroke <- factor(train_set$stroke)
  classifier_dt = rpart(formula =stroke~ .,
                        data = train_set)
  print(classifier_dt)
  # Predicting the Test set results
  y_pred_dt= predict(classifier_dt, newdata = test_set, type = 'class')
  
  # Output labels
  out_labels<-as.data.frame(test_set[, 9])
  out_labels<-t(out_labels)
  
  # Making the Confusion Matrix
  cm_dt = table(out_labels, y_pred_dt)
  
  #accuracy
  n_dt = sum(cm_dt)
  diag_dt = diag(cm_dt)
  accuracy_dt = sum(diag_dt) / n_dt
  accuracy_dt

  dtree_result[length(dtree_result)+1] = accuracy_dt
}
```


Display the accuracy of the Decision Tree model: 
```{r}
accuracy_dt
```


Storing the predictions as a new column on the data set and re-labeling it so it can be used with the Aequitas tool:
```{r}
decision_tree_data_raw <- final3
decision_tree_data <- subset(decision_tree_data_raw, select = -c(hypertension, heart_disease, avg_glucose_level, ever_married, smoking_status))
colnames(decision_tree_data)[4] ="score"
colnames(decision_tree_data)[5] ="label_value"

#remove the comment from the following line to save the data as a csv so it can be put into Aequitas
#write.csv(decision_tree_data, "~/Desktop//decision_tree.csv", row.names=FALSE)
```

Code for the random forest model: 
```{r}
rforest_result = c()

for (i in 1:no_of_exps) {
  cat("Current experiment: ", i)
  
  majority_sample <- majority_class[sample(nrow(majority_class), 548), ]
  balanced_dataset <- rbind(minority_class, majority_sample)
  
  split = sample.split(balanced_dataset$stroke, SplitRatio = 0.70)
  trainData = subset(balanced_dataset, split == TRUE)
  testData = subset(balanced_dataset, split == FALSE)
  
  
  
  trainData$stroke <- as.character(trainData$stroke)
  trainData$stroke <- as.factor(trainData$stroke)
  stroke_rf = randomForest(stroke~., data=trainData, ntree=100, proximity=T)
  strokePred = predict(stroke_rf, newdata=testData)
  CM = table(strokePred, testData$stroke)
  accuracy = (sum(diag(CM)))/sum(CM)
  accuracy

  rforest_result[length(rforest_result)+1] = accuracy
}
```


Display the accuracy for the random forest model
```{r}
accuracy
```


Storing the predictions as a new column on the data set and re-labeling it so it can be used with the Aequitas tool:
```{r}
random_forest_data_raw<- cbind(testData, strokePred)
random_forest_data <- subset(random_forest_data_raw, select = -c(hypertension, heart_disease, avg_glucose_level, ever_married, smoking_status))
colnames(random_forest_data)[4] ="score"
colnames(random_forest_data)[5] ="label_value"

#remove the comment from the following line to save the data as a csv so it can be put into Aequitas
#write.csv(random_forest_data, "~/Desktop//random_forest.csv", row.names=FALSE)
```


Citation:
Code was reproducded this source: 
C. S. Nwosu*, S. Dev*, P. Bhardwaj, B. Veeravalli and D. John, Predicting Stroke from Electronic Health Records, Proc. 41st Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), 2019. (* Authors contributed equally.)