TABLE 7. ECNN and Random Forest Reduced Feature Selection Results.

Metric	ECNN Features		RF Features				Random Selection Features
	NN	RF	NN	P-Value	RF	P-Value	NN	RF
TPR	0.758±0.014	0.721±0.0008	0.775±0.039	0.077	0.720±0.003	0.125	0.998±0.002	0.688±0.540
TNR	0.759±0.025	0.780±0.0012	0.659±0.056	2.76E-11	0.743±0.002	2.00E-79	0.002±0.002	0.323±0.559
PPV	0.766±0.016	0.773±0.0008	0.704±0.024	0.06E-19	0.745±0.001	1.21E-90	0.510±0.001	0.566±0.098
NPV	0.751±0.005	0.728±0.0004	0.739±0.017	0.002	0.719±0.002	1.42E-36	0.308±0.314	0.233±0.251
FPR	0.241±0.025	0.220±0.0011	0.341±0.056	2.10E-12	0.257±0.002	1.50E-78	0.998±0.002	0.677±0.559
FNR	0.242±0.014	0.280±0.0008	0.225±0.039	0.062	0.280±0.003	0.490	0.002±0.002	0.312±0.540
Accuracy	0.759±0.006	0.750±0.0003	0.718±0.007	8.71E-36	0.732±0.001	1.70E-79	0.510±0.001	0.509±0.001
F1 score	0.762±0.002	0.746±0.0003	0.737±0.005	2.15E-38	0.732±0.001	3.82E-73	0.675±0.001	0.489±0.322
AUROC	0.854±0.015	0.785±0.0008	0.734±0.031	4.54E-30	0.735±0.002	6.32E-91	0.591±0.002	0.604±0.532

Shown is a comparison of feature selection effectiveness between ECNN, random forest - a similar traditional feature selection methodology, and random feature selection. Two types of basic classification models: neural network and random forest, trained on the various feature selection subsets, are used in determining selection effectiveness via predictive performance. A baseline random subset of features was also evaluated P-value analysis is a comparison of ECNN features against RF features of the corresponding underlying predictive model.