A comparison of feature selection methodologies and learning algorithms in the development of a DNA methylation-based telomere length estimator

Table 4 Features remaining in models after both feature-selection stages for the nested CV analysis (Stage 1)

Models developed using nested CV	Average features remaining after initial feature- selection stage	Average features remaining after elastic net stage (Std. Error)
Baseline (Elastic net)	–	30 (16.3)
F-test (FDR: 0.01)/Elastic net	23,685	4637 (283.5)
F-test (FDR: 0.05)/Elastic net	55,750	6398 (672.2)
Gradient Boosting/Elastic net	584	264 (8.7)
Pearson Correlation/Elastic net	750	68 (22.1)
Mutual Information/Elastic net	6500	394 (109.6)
Linear SVR/Elastic net	8500	4453 (12.6)
Random Forest Regression/Elastic net	3250	406 (10.6)
PCA/elastic net	1072*	41(9.5)*

As described in “Feature-selection methods” Section, the optimal number of features for the four ranking feature-selection methods (Pearson correlation, mutual information, linear SVR and random forest) were selected from analysis of Figs. 2 and 3. For example, the minimum MAE corresponded to passing 3250 features from the random forest feature ranking to the elastic net regression stage, resulting in an average of 406 features being selected. The figures denote the average (rounded) from the five training sets of the nested CV process described in “Modelling overview”. Values in parentheses denote the standard error of the mean. *denotes principal components

ISSN: 1471-2105