Fig. 1

Nested cross validation scheme with stable variable selection (StaVarSel). In the inner loop the level of regularisation (lambda) for the regression model was optimised via repeated tenfold cross validation. For the StaVarSel, the miR-ratios derived from applying lasso regression with the optimised lambda to each training set were collated, ranked according to frequency of selection, and then subjected to stepwise selection at percentile cut-offs to determine the optimum model with the least prediction error. The stable miR-ratios thus selected from the inner loop cross validation were then used to build regression models in the cross validation outer loop and make predictions of the held-out samples