AutoScore_parsimony {AutoScore}

AutoScore STEP(ii): Select the best model with parsimony plot (AutoScore Modules 2+3+4)

Description

AutoScore STEP(ii): Select the best model with parsimony plot (AutoScore Modules 2+3+4)

Usage

AutoScore_parsimony(
  train_set,
  validation_set,
  rank,
  max_score = 100,
  n_min = 1,
  n_max = 20,
  cross_validation = FALSE,
  fold = 10,
  categorize = "quantile",
  quantiles = c(0, 0.05, 0.2, 0.8, 0.95, 1),
  max_cluster = 5,
  do_trace = FALSE,
  auc_lim_min = 0.5,
  auc_lim_max = "adaptive"
)

Arguments

train_set

A processed data.frame that contains data to be analyzed, for training.

validation_set

A processed data.frame that contains data for validation purpose.

rank

the raking result generated from AutoScore STEP(i) AutoScore_rank

max_score

Maximum total score (Default: 100).

n_min

Minimum number of selected variables (Default: 1).

n_max

Maximum number of selected variables (Default: 20).

cross_validation

If set to TRUE, cross-validation would be used for generating parsimony plot, which is suitable for small-size data. Default to FALSE

fold

The number of folds used in cross validation (Default: 10). Available if cross_validation = TRUE.

categorize

Methods for categorize continuous variables. Options include "quantile" or "kmeans" (Default: "quantile").

quantiles

Predefined quantiles to convert continuous variables to categorical ones. (Default: c(0, 0.05, 0.2, 0.8, 0.95, 1)) Available if categorize = "quantile".

max_cluster

The max number of cluster (Default: 5). Available if categorize = "kmeans".

do_trace

If set to TRUE, all results based on each fold of cross-validation would be printed out and plotted (Default: FALSE). Available if cross_validation = TRUE.

auc_lim_min

Min y_axis limit in the parsimony plot (Default: 0.5).

auc_lim_max

Max y_axis limit in the parsimony plot (Default: "adaptive").

Details

This is the second step of the general AutoScore workflow, to generate the parsimony plot to help select a parsimonious model. In this step, it goes through AutoScore Module 2,3 and 4 multiple times and to evaluate the performance under different variable list. The generated parsimony plot would give researcher an intuitive figure to choose the best models. If data size is small (ie, <5000), an independent validation set may not be a wise choice. Then, we suggest using cross-validation to maximize the utility of data. Set cross_validation=TRUE. Run vignette("Guide_book", package = "AutoScore") to see the guidebook or vignette.

Value

List of AUC value for different number of variables

References

• Xie F, Chakraborty B, Ong MEH, Goldstein BA, Liu N, AutoScore: A Machine Learning-Based Automatic Clinical Score Generator and Its Application to Mortality Prediction Using Electronic Health Records, JMIR Med Inform 2020;8(10):e21798, doi: 10.2196/21798

See Also

AutoScore_rank, AutoScore_weighting, AutoScore_fine_tuning, AutoScore_testing, Run vignette("Guide_book", package = "AutoScore") to see the guidebook or vignette.

Examples

# see AutoScore Guidebook for the whole 5-step workflow
data("sample_data")
names(sample_data)[names(sample_data) == "Mortality_inpatient"] <- "label"
out_split <- split_data(data = sample_data, ratio = c(0.7, 0.1, 0.2))
train_set <- out_split$train_set
validation_set <- out_split$validation_set
ranking <- AutoScore_rank(train_set, ntree=100)
AUC <- AutoScore_parsimony(
train_set,
validation_set,
rank = ranking,
max_score = 100,
n_min = 1,
n_max = 20,
categorize = "quantile",
quantiles = c(0, 0.05, 0.2, 0.8, 0.95, 1)
)