Table 4.

Variable selection for prostate cancer data (prostate cancer vs. BPH)

	GDFS	CFS	rpart	Predominant glycans (GDFS method)
Peak 1	✗	✗	✗
Peak 2	✗	✗	✗
Peak 3	✗	✗	✗
Peak 4	✗	✗	✗
Peak 5	✗	✗	✗
Peak 6*	✗	✗	✗	FA2[3]G1, FA2[6]BG1
Peak 7	✗	✗	✗
Peak 8	✗	✗	✗
Peak 9	✗	✗	✗
Peak 10	✓	✗	✗	FA2G2, FA2[6]G1S1, FA2[6]BG1S1
Peak 11	✗	✗	✗
Peak 12	✗	✗	✗
Peak 13**	✗	✗	✗	A2BG2S1
Peaks 14 - 24	✗	✗	✗

Features selected from the prostate cancer dataset (prostate cancer vs. BPH cases) by the proposed GDFS method (GDFS), correlation-based feature selection (CFS), and recursive partitioning (rpart). Features that were selected in 90% more of the cross-validation models are marked with $✓$. Also listed are the predominant glycan structures corresponding to each selected peak. Detailed N-glycan composition of human serum was described in Royle et al. [9], and peak 10 was also assigned in Saldova et al. [24]. *Peak 6 was the most commonly identified feature by the rpart method, although it was selected less than 90% of the time. **Peak 13 was selected more than 60% of the time by the GDFS method.