Fig. 8

Percent of high and low quality features in the test set remaining after each filtering step. Each step of the proposed data-adaptive filtering pipeline considerably reduces the number of remaining low quality features in the test set. The desired trade-off between removing low and high quality features can be obtained by adjusting the stringency of the cutoffs at each step. For the CRC dataset, 76% of the low quality features and 28% of the high quality features in the test set were removed. For the urine dataset, 74% of the low quality features and 17% of the high quality features in the test set were removed. For the cell line dataset, 76% of the low quality features and 21% of the high quality features in the test set were removed