Table 4.

Comparison of different strategies of feature selection on benchmark training datasets.

Strategy		Sensitivity	Specificity	Precision	Accuracy	MCC	F1	AUC	Number of features
Combination of all features	Average stdev p-value	0.707 ± 0.017 6.9e-10	0.716 ± 0.010 5.2e-10	0.713 ± 0.008 5.6e-11	0.711 ± 0.009 1.4e-11	0.423 ± 0.018 1.4e-11	0.710 ± 0.011 2.4e-11	0.765 ± 0.010 4.7e-19	470 N/A N/A
LASSO	Average stdev p-value	0.734 ± 0.016 5.6e-05	0.750 ± 0.004 1.6e-05	0.746 ± 0.006 2.2e-06	0.742 ± 0.009 1.2e-06	0.484 ± 0.017 1.1e-06	0.740 ± 0.011 2.2e-06	0.784 ± 0.006 2.3e-18	74 ± 11 1.8e-20
GA	Average stdev p-value	0.752 ± 0.024 0.16	0.755 ± 0.009 5.2e-04	0.754 ± 0.009 4.8e-04	0.753 ± 0.013 3.3e-03	0.507 ± 0.025 3.2e-03	0.753 ±0.016 0.01	0.813 ± 0.006 9.4e-16	233 ± 11 3.2e-05
DPSO	Average stdev p-value	0.752 ± 0.013 0.0248	0.752 ± 0.005 4.7e-05	0.753 ± 0.006 5.7e-05	0.751 ± 0.008 1.7e-04	0.504 ± 0.016 1.7e-04	0.752 ±0.009 4.7e-04	0.819 ± 0.004 1.4e-16	280 ± 6 5.8e-10
DFA	Average stdev p-value	0.763 ± 0.007 N/A	0.777 ± 0.014 N/A	0.774 ± 0.012 N/A	0.770 ± 0.009 N/A	0.540 ± 0.018 N/A	0.768 ±0.008 N/A	0.876 ± 0.005 N/A	254 ± 4 N/A

The threshold is set where the MCC achieve the maximum value. LASSO, DFA, DPSO, and GA stand for least absolute selection and shrinkage operator, discrete firefly algorithm, discrete particle swam optimization, and genetic algorithm, respectively.