TABLE 5.

Confusion matrix of each model with optimal probability cutoff of training set.

Real predicted		Cluster 0		Cluster 1		Cluster 2		Cluster 3
		CIRKP	CISKP	CIRKP	CISKP	CIRKP	CISKP	CIRKP	CISKP
LR	CIRKP	127	62	166	97	15	6	38	35
	CISKP	41	206	29	167	5	91	24	140
SVM	CIRKP	117	36	156	49	16	2	36	18
	CISKP	51	232	39	215	4	95	26	157
RF	CIRKP	160	160	192	209	17	19	49	69
	CISKP	8	108	3	55	3	78	13	106
XGB	CIRKP	105	37	161	57	14	5	35	22
	CISKP	63	231	34	207	6	92	27	153
		Cluster 4		Cluster 5		Cluster 6		Cluster 7
		CIRKP	CISKP	CIRKP	CISKP	CIRKP	CISKP	CIRKP	CISKP
LR	CIRKP	91	46	159	57	0	3	28	20
	CISKP	22	103	61	413	1	51	8	55
SVM	CIRKP	85	27	156	37	0	2	30	21
	CISKP	28	122	64	433	1	52	6	54
RF	CIRKP	111	129	202	187	0	11	36	49
	CISKP	2	20	18	283	1	43	0	26
XGB	CIRKP	93	36	169	50	0	2	34	20
	CISKP	20	113	51	420	1	52	2	55
		LR		SVM		RF		XGB
REF		CIRKP	CISKP	CIRKP	CISKP	CIRKP	CISKP	CIRKP	CISKP
	CIRKP	624	326	596	192	767	825	611	229
	CISKP	191	1,226	219	1,360	48	727	204	1,323

RF model is severely overfitted to CIRKP group. The in-cluster performance of cluster 6 is acceptable but low AUC value is caused by insufficient positive test samples.