Diabetes Prediction using Tidymodels

library(tidyverse)
library(tidymodels)
library(data.table)
library(gtsummary)
library(mTools)

Diabetes data

diabetes_df_all <- fread("data/diabetes_prediction_dataset.csv")

Data processing

diabetes_df_all[, diabetes_char := factor(diabetes, 
                                      levels = c(0, 1),
                                      labels = c("Non diabetic", "Diabetic"))]

re_balance_class <- function(df, outcome_col = "diabetes_char", pos_class = "Diabetic", pos_class_perc = .4){
    
    pos_class_df = df[get(outcome_col) == pos_class]
    neg_class = df[get(outcome_col) != pos_class]
    pos_perc = nrow(pos_class_df)/nrow(df)
    N = round(nrow(pos_class_df)/pos_class_perc)
    Nneg = N - nrow(pos_class_df)
    neg_class_df = neg_class[sample(1:.N, Nneg)]
    rbind(pos_class_df,neg_class_df )
    
    
}

diabetes_df = re_balance_class(df = diabetes_df_all)

Summary Stats

library(ggiraph)
db_perc <- diabetes_df[, .(freq = .N),
                       by = diabetes_char][
                           ,perc := round(freq/sum(freq) * 100, 1)]


ggplot(db_perc, aes(diabetes_char, freq, fill = diabetes_char))+
    geom_bar_interactive(width = 0.5, stat = "identity")+
    geom_text(aes(label = paste0(freq, "(", perc, "%)")),
              position = position_dodge(width = 0.5),
              vjust = 0.05)+
    scale_fill_brewer(name = "", type = "qual", palette = "Dark2")+
    theme_minimal()+
    theme(
        legend.position = "bottom"
    )

tab2 <- diabetes_df %>%
    tbl_summary(
        by = diabetes_char,
        type = all_continuous() ~ "continuous2",
        statistic = all_continuous() ~ c(
            "{mean} ({sd})",
            "{median} ({p25}, {p75})",
            "[{min}, {max}]"
        ),
        missing = "ifany"
    ) %>%
    add_p(pvalue_fun = ~ style_pvalue(.x, digits = 2))

tab_df = as.data.frame(tab2)
nms <- names(tab_df)
nms <- gsub("\\*", "", nms)
names(tab_df) <- nms
data_table(tab_df)

Model Fitting

set.seed(100)
diabetes_df[, diabetes:= as.factor(diabetes)]
diabetes_df_split <- initial_split(diabetes_df[,.SD, .SDcols = !"diabetes_char"], 
                                   strata = diabetes)

diabetes_df_train <- training(diabetes_df_split)

diabetes_df_test <- testing(diabetes_df_split)

# Specify a logistic regression model
logistic_model <- logistic_reg() %>% 
  # Set the engine
  set_engine('glm') %>% 
  # Set the mode
  set_mode('classification')

# Fit to training data
logistic_fit <- logistic_model %>% 
  fit(diabetes ~ .,
      data = diabetes_df_train)

# Print model fit object
logistic_fit %>% 
    DT_tidy_model()

xgb_spec <- boost_tree(
    trees = 2000,
    tree_depth = tune(), 
    min_n = tune(),
    loss_reduction = tune(),                     ## first three: model complexity
    sample_size = tune(), 
    mtry = tune(),         ## randomness
    learn_rate = tune()                          ## step size
) %>%
    set_engine("xgboost") %>%
    set_mode("classification")

xgb_spec

Boosted Tree Model Specification (classification)

Main Arguments:
  mtry = tune()
  trees = 2000
  min_n = tune()
  tree_depth = tune()
  learn_rate = tune()
  loss_reduction = tune()
  sample_size = tune()

Computational engine: xgboost