Evaluation of data discretization methods to derive platform independent isoform expression signatures for multi-class tumor subtyping

Table 2 Comparison of classification accuracy using top 100 genes using data from same platform.

Feature selection	CV				SVM-RFE				RF_based_FS
Classifier	FC	Equal-W	Equal-F	k-means	FC	Equal-W	Equal-F	k-means	FC	Equal-W	Equal-F	k-means
SVM	84.7	85.9	85.9	90.6	77.6	81.2	92.9	87.1	96.5	94.1	95.3	96.5
RF	85.9	85.9	84.7	85.9	81.2	78.8	88.2	87.1	91.7	92.9	91.7	90.6
NB	82.3	81.2	80.0	80.0	75.3	69.4	81.2	78.8	90.6	92.9	85.9	84.7
PAM	85.9	87.1	85.9	84.7	71.7	70.6	84.7	80.0	91.7	91.7	87.1	85.9

The classification models were trained (257 samples) and tested (85 samples) on exon-array data. Highest accuracy for each classification method is marked in bold. While SVM in combination with RF_based_FS performed best whit the highest accuracy for both FC data (without discretization) and k-means discretised data, the other three classifiers (RF, NB and PAM) in combination with RF_based_FS achieved comparable classification accuracies on Eaual-W discretized data.

ISSN: 1471-2164