Predictors of Testing Positive for Diabetes

This analysis uses the Pima Indians Diabetes dataset to determine which factors are important predictors of testing positive for diabetes.

Data Structure

Let’s get a quick understanding of the variables included in our dataset and how they relate to each other.

str(diabetes)
plot(diabetes, col = "lightblue")

'data.frame':   768 obs. of  9 variables:
 $ pregnant: int  11 6 2 4 1 5 8 5 0 2 ...
 $ glucose : int  143 92 90 111 103 88 176 44 109 109 ...
 $ pressure: int  94 92 68 72 80 66 90 62 88 92 ...
 $ triceps : int  33 0 42 47 11 21 34 0 30 0 ...
 $ insulin : int  146 0 0 207 82 23 300 0 0 0 ...
 $ mass    : num  36.6 19.9 38.2 37.1 19.4 24.4 33.7 25 32.5 42.7 ...
 $ pedigree: num  0.254 0.188 0.503 1.39 0.491 0.342 0.467 0.587 0.855 0.845 ...
 $ age     : int  51 28 27 56 22 30 58 36 38 54 ...
 $ diabetes: Factor w/ 2 levels "neg","pos": 2 1 2 2 1 1 2 1 2 1 ...

Optimal Decision Tree Model

The optimal Decision Tree (hold-out set accuracy of 79.22%) model has the following hyper-parameters:

jsonlite::toJSON(best_tree$control, pretty = TRUE, auto_unbox = TRUE)

{
  "minsplit": 20,
  "minbucket": 7,
  "cp": 0.01,
  "maxcompete": 4,
  "maxsurrogate": 5,
  "usesurrogate": 2,
  "surrogatestyle": 0,
  "maxdepth": 18,
  "xval": 10
} 

We can also plot the tree structure using rpart.plot.

rpart.plot::rpart.plot(best_tree, roundint = FALSE)

Variable Importance

Finally, we can determine the most important predictors of a positive diabetes diagnosis, using the variable importance feature of our Decision Tree.

var_importance <- transform(
  data.frame(
    "Feature" = names(best_tree$variable.importance),
    "Importance" = best_tree$variable.importance
  ),
  Importance = Importance/sum(Importance)
)

ggplot2::ggplot(
  var_importance,
  ggplot2::aes(x = forcats::fct_reorder(Feature, Importance), y = Importance)
) +
  ggplot2::geom_bar(stat = "identity") +
  ggplot2::labs(x = "Feature") +
  ggplot2::theme_light() +
  ggplot2::theme(
    axis.title = ggplot2::element_text(face = "bold"),
    panel.border = ggplot2::element_blank()
  )

Conclusion

Unsurprisingly, glucose levels are by far the most important indicator of diabetes.