Skip to main content

Table 2 Methods for predictor selection before and after modelling and hyperparameter tuning for 152 developed clinical prediction models, by modelling type

From: Methodological conduct of prognostic prediction models developed using machine learning in oncology: a systematic review

  All (n = 152) Regression-based models (n = 42) Non-regression-based models (n = 71) Ensemble models (n = 39)
n (%) n (%) n (%) n (%)
Predictor selection (before modelling) reported 52 (34) 20 (48) 23 (32) 9 (23)
 A-priori 5 3 1 1
 No selection before modelling 3 1 2
 Univariable 24 12 8 4
 Clinically relevant and available data 1 1
 Dropout technique at input layer 1 1
 Random forest with RPA 9 1 6 2
 Other modelling approacha 9 3 4 2
Predictor selection (during modelling) reported 63 (41) 25 (59) 27 (38) 11 (28)
 Stepwise 6 4 2
 Forward selection 6 5 1
 Backward elimination 5 3 2
 Full model approach (no selection) 11 4 5 2
 Feed forward/backpropagation 5 5
 Recursive partitioning analysis 7 7
 LASSO 5 5
 Gini index (minimised) 7 1 4 2
 Cross validation 4 2 2
 Otherb 7 1 2 4
Hyperparameter tuning methods reported 31 (21) 4 (10) 15 (23) 12 (31)
 Cross validation 19 4 7 8
 Grid search (no further details provided) 6 4 2
 Max tree depth 2 1 1
 Adadelta method 2 2
 Default software values 2 1 1
  1. RPA Recursive partitioning analysis, LASSO Least Absolute Shrinkage and Selection Operator
  2. aModelling approaches include support vector machine, logistic regression, Cox regression, best subset linear regression, decision tree, meta-transformer (base algorithm of extra trees)
  3. bOther includes change in unspecified performance measure, stochastic gradient descent, function, aggregation of bootstrapped decision trees and Waikato Environment for Knowledge Analysis for development-only studies, and hyperbolic tangent function, greedy algorithm for all models and using final chosen predictors from comparator model