A three-step approach for the derivation and validation of high-performing predictive models using an operational dataset: congestive heart failure readmission case study

Table 4 Final status of categorical and accumulated discretized explanatory variables

	P-values	Main values	No (557)	Yes (132)	G/W	Inf	Sym
*Gender*	0.2	Female	235	64
*Gender*		Male	322	68
*Marital status*	0.046	Married	233	74
*Marital status*		Others	334	58
Race	0.073	White Or Caucasian	401	95
Race		Others	156	37
Religion	0.85	Latter Day Saints	201	51
Religion		Others	356	81
Discharge disposition	≤0.001	Home Or Self Care	348	77	2	1	1
Discharge disposition		Others	209	55
Hospital services	≤0.001	Cardiology	366	72	8	2	7
Hospital services	≤0.001	Others	191	60
Insurance/finance class	0.12	Medicare	347	73
Insurance/finance class		Others	210	59
Prior 6-month ED frequency	0.001	1	250	41	13	6	12
Prior 6-month ED frequency		Others	307	91
Age-at-Admission (AGE)	≤0.001	≤ 83.5	475	129	3	8	5
Age-at-Admission (AGE)	≤0.001	>83.5	82	3
Length of Stay (LOS)	≤0.001	≤ 5	330	68
Length of Stay (LOS)	≤0.001	>5	227	64
Home proximity	≤0.001	≤ 44	407	86
Home proximity	≤0.001	>44	150	46
*Mean household income*	≤0.001	≤68958.5	364	86
*Mean household income*	≤0.001	>68958.5	193	46
Prior 6-month mean ED LOS	≤0.001	≤0.235	382	110	11	13	13
Prior 6-month mean ED LOS	≤0.001	>0.235	175	22

Variables included as risk factors in the final model are indicated in non-italic font, and variables excluded from the model are indicated in italics. The dataset used is the 2008–2012 derivation dataset. The count under Yes indicates the frequency of the variable value when readmissions occurred, and the coun under No indicates the frequency of the variable value when readmissions did not occur. G/W, Info and Sym are the relative ranks of the most significant variables (relative weight ≥ 0.001) using the GainRatioAttributeEval/Wrapper, InfoGainAtttributeEval, and SymmetricalUncertAttrbuteEval ranking strategies, respectively. G/W combines two strategies, because the relative ranks were equivalent using both approaches.

ISSN: 1472-6947