Compensation of feature selection biases accompanied with improved predictive performance for binary classification by using a novel ensemble feature selection approach

Table 3 Results on datasets

Dataset	All [CI]	AUC-FS [CI]	EFS [CI]	AUC-FS vs. EFS*	all vs. EFS**
MI-Mortality	0.758 [0.700, 0.800]	0.757 [0.704, 0.811]	0.776 [0.725, 0.826]	0.228	0.201
Fibrosis	0.493 [0.300, 0.600]	0.681 [0.537, 0.824]	0.746 [0.617, 0.874]	0.273	0.018
FLIP	0.759 [0.600, 0.900]	0.723 [0.582, 0.863]	0.761 [0.633, 0.890]	0.254	0.971
SPECTF	0.807 [0.700, 0.900]	0.856 [0.811, 0.901]	0.865 [0.821, 0.910]	0.444	4.68e-4
Sonar	0.792 [0.700, 0.900]	0.840 [0.787, 0.894]	0.862 [0.813, 0.911]	0.200	0.009
WBC	0.611 [0.600, 0.700]	0.987 [0.977, 0.998]	0.991 [0.981, 1.000]	0.020	1.21e-41

Column 1 to 3 are AUCs values of all features, selected by AUC-FS and by the EFS with confidential intervalls in brackets. The last two columns show the p-values of the comparison by the method of [28]. The function compares the AUC of the ROC curves of (*) the AUC-FS and EFS method and (**) of all parameters and EFS outcome. Statistical significant p-values are printed in bold

ISSN: 1756-0381