. Author manuscript; available in PMC: 2017 Jun 1.

Published in final edited form as: J Biomed Inform. 2016 Mar 16;61:119–131. doi: 10.1016/j.jbi.2016.03.009

Table 3.

Net reclassification (cNRI) comparisons for IPC weighted versions of the machine learning techniques described in Section 3 evaluated on the hold-out test set.

	cNRI
	Events	Non-Events	cNRI Overall	Overall Weighted
Tree
vs. k-NN	−0.003	0.048	0.045	0.045
vs. Bayes	−0.064	0.058	−0.006	0.050
vs. Logistic	−0.065	0.045	−0.020	0.038
vs. GAM	−0.056	0.030	−0.026	0.024

k-NN
vs. Tree	0.003	−0.048	−0.045	−0.045
vs. Bayes	−0.065	0.015	−0.050	0.009
vs. Logistic	−0.108	0.009	−0.099	0.001
vs. GAM	−0.069	−0.013	−0.082	−0.016

Bayes
vs. Tree	0.064	−0.058	0.006	−0.050
vs. k-NN	0.065	−0.015	0.050	−0.009
vs. Logistic	−0.013	−0.017	−0.030	−0.017
vs. GAM	0.028	−0.040	−0.012	−0.035

Logistic
vs. Tree	0.065	−0.045	0.020	−0.038
vs. k-NN	0.108	−0.009	0.099	−0.001
vs. Bayes	0.013	0.017	0.030	0.017
vs. GAM	0.037	−0.022	0.015	−0.018

GAM
vs. Tree	0.056	−0.030	0.026	−0.024
vs. k-NN	0.069	0.013	0.082	0.016
vs. Bayes	−0.028	0.040	0.012	0.035
vs. Logistic	−0.037	0.022	−0.015	0.018

Positive numbers indicate that the bolded technique correctly reclassifies subjects more frequently than the technique preceded by “vs”. cNRI (Events) and cNRI (Non-Events) give the reclassification improvement among those who did and did not experience events, and cNRI (Overall) is their sum. cNRI (Overall Weighted) is a weighted sum where the reclassification performance among Events and Non-Events is weighted according to the event and non-event probabilities, respectively. Tree: Classification trees; k-NN: k-nearest neighbors; Bayes: Bayesian network models; Logistic: Logistic regression; GAM: Generalized additive models.