Splitting random forest (SRF) for determining compact sets of genes that distinguish between cancer subtypes

Table 2 Overall performance comparison of 5 gene lists in the GB, BC and OC validation datasets

Cancer	Methods	No. of Genes	Prediction Accuracy	Change in Prediction Accuracy⁺	Multi-Class AUC⁺⁺	Area Covered by Radar Chart⁺⁺⁺
GB	SRF50	36	80.1%		0.87	1.68
	Single RF	88	77.8%	−2.3%	0.86	1.63
	Verhaak et al.	833	86.0%	5.9%	0.92	1.91
	ANOVA	8670	84.1%	4.0%	0.9	1.83
	Top 50 ANOVA	50	77.2%	−2.9%	0.87	1.71
BC	SRF50	48	84.0%		0.91	1.85
	Single RF	46	74.1%	−9.9%	0.87	1.69
	Parker et al.	979	89.0%	5.0%	0.89	1.78
	ANOVA	4976	85.2%	1.2%	0.86	1.65
	Top 50 ANOVA	50	82.7%	−1.3%	0.87	1.72
OC	SRF50	189	89.8%		0.96	2.06
	Single RF	245	88.9%	−0.9%	0.96	2.06
	Tothill et al.	2106	91.7%	1.9%	0.97	2.11
	ANOVA	7144	90.7%	0.9%	0.97	2.09
	Top 50 ANOVA	50	87.0%	−2.8%	0.95	2.01

⁺All changes of prediction accuracy were calculated using SRF50 as the referent; ⁺⁺ Multi-class AUC values were calculated by taking average of the all pairwise AUC values; ⁺⁺⁺ Area covered by radar chart (ACRC) [29].