Table 3 Comparison of related methods.

Michiels et al, 2005 [2]

(1) Selected genes most correlated with prognosis,

(2) Created nearest centroid classification rule.

Used

Test and training sample splits in entire data set.

(1) Misclassification rate for test samples

(2) Frequencies of genes selected in training sample

Ma et al, 2006 [7]

(1) Split into training-training sample and training-test sample,

(2) Using cross-validation, maximized the binormal area under ROC curve as a linear function of genes; then selected genes with non-zero coefficients.

Used

Training-training and training-test samples (i.e. the cross-validation and evaluation is repeated)

(1) Area under ROC curve for test samples,

(2) Frequencies of genes selected in training sample.

Li et al, 2004 [8]

(1) Split into training-training sample and training-test sample,

(2) Cross-validated classification tree to maximize fit.

Not used

Resampling for training-training samples and training test samples.

(1) Relevancy intensity, which equals frequencies of genes selected in training sample when weights equal 1.

Proposed method

(1) Selected genes with highest individual, classification performance

(2) Created classification rule using nearest centroid and score function.

Used

Test and training samples splits in entire data set.

(1) ROC curve and area under ROC curve for test samples with emphasis on comparing many versus few genes,

(2) Frequencies of genes selected in training sample.