Fig. 3.

Performance comparison with relevant approaches with inherent feature selection. The performance of the proposed multi-objective optimiser was compared with relevant methods with inherent feature selection—i.e., single-objective optimiser, Lasso, guided RRF and penalised SVM. a The accuracy, specificity sensitivity and functional relevance score were averaged over 50 runs of sample partitioning using fivefold cross validation. b Sizes of the identified signatures or the number of features selected by each method over 50 independent runs were shown as box plots. c As a measure of signature stability, Jaccard Index was computed for all pairs of signatures identified by each of compared methods across 50 runs and the average values were reported. In all bar charts, error bars show standard deviations and multi-objective optimiser bars were marked by ‘*’ when the proposed method significantly outperforms others (Wilcoxon test p values < 0.001)