do_feature_selection: Perform feature selection based on different methods

Description Usage Arguments Value Author(s)

View source: R/stats_functions.R

Description

This function implements different methods to perform feature selection of radiomic datasets.

Usage

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
do_feature_selection(
  rdr = NULL,
  n_features = NULL,
  select_by = NULL,
  method = "hcl",
  surv_obj = NULL,
  which_data = "scaled",
  corr_measure = "pearson",
  min_features_per_group = 5,
  thr_pca_cum_prop = 0.8,
  response = NULL,
  lambda = "min",
  alpha = NULL
)

Arguments

rdr

A RadAR object (class SummarizedExperiment).

n_features

(numeric) Number of features to be selected. Required.

select_by

(character) Which criteria use to select informative radiomic features within clusters of similar (i.e., redundant) features. It can be one of the following: "variability", "random", "concordance".

method

(character) Which method use to identify redundant features. It can be one of the following: "mRMR" (minimum-redundancy-maximum-relevance),"hcl" (hierarchical clustering of correlation matrix), "pca" (K-means applied to Principal Component Analysis), "glmnet-cox" (generalized linear model via penalized maximum likelihood (glmnet) fitting cox regression model), "glmnet-binonial" (glmnet fitting binomial regression model), Using mRMR, this function works as a wrapper to mRMR package. Using glmnet-*, this function works as a wrapper to glmnet package.

surv_obj

An object of class Surv. Required if select_by is "concordance".

which_data

(character) Which data use to compute concordance index. It can be one of the following: "normal", "scaled", "normalized".

corr_measure

(character) Which method use to calculate correlation. It can be one of the following: "pearson", "kendall", "spearman".

min_features_per_group

(numeric) Minimum number of features for each cluster.

thr_pca_cum_prop

(numeric) Threshold to select number of components based on cumulative proportion of explained variance criterion.

response

(numeric) A response variable, required if any of mRMR or glmnet-binomial or methods are used.

lambda

(character) In glmnet, it controls the overall strength of the penalty. Possible values are "min" or "1se" (1 standard deviation). For more details see glmnet

alpha

(numeric) In glmnet, it controls elastic-net penalty. Typical values are 0 (ridge) or 1 (lasso). For more details see glmnet.

Value

A list including two elements: 'rdr': the updated (reduced) rdr (a RadAR object) 'signature': the radiomic features included in the signature

Author(s)

Matteo Benelli (matteo.benelli@uslcentro.toscana.it)


cgplab/RadAR documentation built on Nov. 10, 2021, 1:32 a.m.