A comparative analysis of predictive models of morbidity in intensive care unit after cardiac surgery – Part II: an illustrative example

Table 9 Number of selected features and corresponding model performance

Model	No. of features	Testing discrimination AŨC (CI, CI%)	Training discrimination AŨC	Generalization ΔAŨC%	Calibration HL-p
BL	8	0.778 (0.722–0.831, 14.0%)	0.815	4.5%	0.65*
BQ	3	0.785 (0.738–0.832, 12.0%)	0.780	-0.6%	0.19*
k NN	5	0.772 (0.717–0.822, 13.6%)	0.792	2.5%	0.01*
LR	14	0.781 (0.721–0.830, 14.0%)	0.827	5.6%	0.29
HS	14	0.768 (0.714–0.821, 13.9%)	0.828	7.2%	<0.001*
DS	16	0.779 (0.727–0.830, 13.2%)	0.836	6.8%	<0.001*
ANN1	13	0.776 (0.715–0.827, 14.4%)	0.843	7.9%	0.07*
ANN2	10	0.778 (0.726–0.825, 12.7%)	0.837	7.0%	0.01*

*after recalibration
BL, Bayes linear; BQ, Bayes quadratic; kNN, k-nearest neighbour; LR, logistic regression; HS, Higgins score; DS, direct score; ANN1, one-layer artificial neural network; ANN2, two-layer artificial neural network; AŨC, median value of area under receiver operating characteristic curve calculated from 1000 bootstrap samples; ΔAŨC%, difference between AŨC of training and test data; CI and CI%, confidence interval and percentage confidence interval; HL-p, p-value of the Hosmer-Lemeshow goodness-of-fit test.

ISSN: 1472-6947