Table 4.

Performance of classifiers applied to independent validation set of prostate cancer data set.

	Unweighted	Weighted	Random forest	SVM
Accuracy
LOOCV
Top 1%	68.5a	76.5b	92.7g	93.5g
Top 5%	74.5c	81.9c	91.6h	92.4h
60:40 Partitions
Top 1%	86.3d	88.2c	91.6i	89.7i
Top 5%	86.7e	89.9f	90.5j	86.6j
Sensitivity
LOOCV
Top 1%	74.4	76.9	87.2	84.6
Top 5%	89.7	79.5	87.2	74.4
60:40 Partitions
Top 1%	74.4	76.9	82.0	65.0
Top 5%	69.2	74.4	84.6	64.1
Positive predictive value
LOOCV
Top 1%	43.9	53.6	70.8	75.0
Top 5%	50.7	62.0	66.7	74.4
60:40 Partitions
Top 1%	52.7	57.7	68.1	66.7
Top 5%	54	61.7	63.5	54.3

Notes: Accuracy, sensitivity, and positive predictive value of voting classifiers (unweighted and weighted), random forest and SVM applied to independent data sets from the prostate cancer data set. Features to include in the classifiers were derived using the top 1% or 5% of features based on t-statistics through a jackknife procedure using training sets in leave-one-out cross validation (LOOCV) or multiple random validation (60:40 partitions).

^aHighest accuracy achieved with 37 features in classifier;

^bHighest accuracy achieved with 43 features in classifier;

^cHighest accuracy achieved with 45 features in classifier,

^dHighest accuracy achieved with 49 features in classifier;

^eHighest accuracy achieved with 47 features in classifier;

^fHighest accuracy achieved with 27 features in classifier. The number of features used in developing SVM and random forest classifiers were:

^g685 features;

^h2,553 features;

ⁱ9,890 features;

^j14,843 features.