TABLE 2.

Tenfold cross-validation performance comparison of four models based on three feature extraction methods on three benchmark data sets.

Data sets	Models	Accuracy (%)	Sensitivity (%)	Specificity (%)	MCC	AUC
NH_990	CNN	67.96	68.09	67.86	0.36	0.737
	CNN + Capsule	66.02	63.83	67.86	0.32	0.742
	CNN + Attention	66.02	46.81	82.14	0.31	0.745
	PseUdeep (CNN+	66.99	74.47	60.71	0.35	0.746
	+Capsule + Attention)
NS_627	CNN	69.71	70.59	68.75	0.39	0.728
	CNN + Capsule	68.18	61.76	75.00	0.37	0.735
	CNN + Attention	69.71	76.47	68.75	0.40	0.734
	PseUdeep (CNN	72.73	61.75	78.13	0.45	0.737
	+Capsule + Attention)
NM_944	CNN	70.41	57.78	86.79	0.41	0.741
	CNN + Capsule	69.39	73.34	66.04	0.39	0.750
	CNN + Attention	70.41	57.78	81.13	0.41	0.751
	PseUdeep (CNN	72.45	66.70	77.36	0.44	0.756
	+Capsule + Attention)

The bold value is the value with the best effect in the corresponding evaluation index.