Fig. 3

C-index computed with different pre-screening thresholds for BRCA and the elastic net penalty. (A) Median C-indices obtained by 10 repetitions of a K-fold cross-validation (K = 5) for different pre-screening thresholds for corrected p-values on the x-axis and for interquartile range (IQR) on the y-axis. Box surrounded by gray: no pre-screening; box surrounded by blue: optimal case (highest median C-index); white numbers: number of genes after the pre-screening step. (B) Number of genes retained by unsupervised (purple) and supervised (lightblue) pre-screening in the optimal case (the blue square in (A)). (C) Boxplot of C-indices obtained with 10 repetitions of 5-fold nested cross-validation without pre-screening (gray box in (A) for elastic net) and in the optimal case for bi-dimensional pre-screening (blue box in (A) for elastic net) for the ridge, the lasso, the elastic net (EN), the adaptive elastic net (AEN). The p-value above each method is calculated with a one-sided Wilcoxon signed-rank test between the C-indices obtained in the optimal case and without pre-screening; blue numbers are the number of genes retained after optimal bi-dimensional pre-screening; black and red numbers are respectively the number of genes and the optimal thresholds retained by supervised (left) and unsupervised (right) pre-screening in the optimal case. ***: \(p \le 0.001\), **: \(p \le 0.01\), *: \(p \le 0.05\), +: \(p \le 0.1\), n.s. : \(p > 0.1\)