Effect of training-sample size and classification difficulty on the accuracy of genomic predictors

Breast Cancer Research

Table 2 Congruencies across different endpoints and different feature-selection methods

Same endpoint but different feature selection (FS)
Endpoint	Gene-level	Level of canonic-pathway maps
ER status	0.541	0.573
pCR	0.544	0.572
pCR(ER^-)	0.593	0.532
Same FS but different endpoints
FS	Gene-level	Level of canonic-pathway maps
FS1	0.300	0.290
FS2	0.299	0.274
FS3	0.291	0.278
FS4	0.295	0.291
FS5	0.272	0.282

The table shows that kappa statistics (that is, congruency) are high for different feature-selection methods for the same endpoint but are low for the same feature-ranking method for different endpoints. Both gene-level and pathway-level analyses show similar results.

ISSN: 1465-542X