In brian-j-smith/MachineShop: Machine Learning Models and Tools

Diabetes in Pima Indian Women

Characteristics of Pima Indian women tested for diabetes are used in this example to predict their disease statuses.

library(mlbench)
data(PimaIndiansDiabetes2)

stats <- list(
  list("Number of women" = ~ length(diabetes)),
  "diabetes" = list("pos" = ~ n_perc(diabetes == "pos"),
                    "neg" = ~ n_perc(diabetes == "neg")),
  "pregnant" = list("Median (Range)" = ~ median_range(pregnant)),
  "glucose" = list("Median (Range)" = ~ median_range(glucose)),
  "pressure" = list("Median (Range)" = ~ median_range(pressure)),
  "triceps" = list("Median (Range)" = ~ median_range(triceps)),
  "insulin" = list("Median (Range)" = ~ median_range(insulin)),
  "mass" = list("Median (Range)" = ~ median_range(mass)),
  "pedigree" = list("Median (Range)" = ~ median_range(pedigree)),
  "age" = list("Median (Range)" = ~ median_range(age))
)

summary_kbl(stats, na.omit(PimaIndiansDiabetes2))

Training Set Analysis

## Analysis libraries
library(MachineShop)
library(ggplot2)

## Parallel processing
library(doParallel)
registerDoParallel(cores = 6)

## Dataset
data(PimaIndiansDiabetes2, package = "mlbench")
Pima <- na.omit(PimaIndiansDiabetes2)

## Model formula
fo <- diabetes ~ .

## Model selected from tuned models
selected_model <- SelectedModel(TunedModel(KNNModel, grid = 5),
                                TunedModel(NNetModel, grid = 5),
                                TunedModel(RandomForestModel, grid = 5))

## Model fit
model_fit <- fit(fo, data = Pima, model = selected_model)

## Variable importance
vi <- varimp(model_fit)
plot(vi)

Generalization Performance

## Resample estimation with cross-validation
res <- resample(fo, data = Pima, model = selected_model, control = CVControl)

## Estimated performance
summary(performance(res))

## Variable probability cutoff
summary(performance(res, cutoff = 0.25))

ROC Curve

## True positive and false positive rates over all probability cutoffs
roc <- performance_curve(res)

## ROC curve
plot(roc, diagonal = TRUE) + coord_fixed()

## Area under the curve
auc(roc)

Confusion Matrices

(conf <- confusion(res))

summary(conf)

plot(conf)

## Variable probability cutoff
summary(confusion(res, cutoff = 0.25))

Calibration Curve

cal <- calibration(res, breaks = NULL)
plot(cal, se = TRUE)

Partial Dependence Plots

pd <- dependence(model_fit, select = c(glucose, age, insulin))
plot(pd)

brian-j-smith/MachineShop documentation built on Sept. 22, 2023, 10:01 p.m.