. Author manuscript; available in PMC: 2014 Dec 9.

Published in final edited form as: J Biomed Inform. 2014 Jul 15;52:260–270. doi: 10.1016/j.jbi.2014.07.007

Table 4.

Combined phenotype validation results

	ILP-1¹	ILP+BP²	PART³	J48⁴	JRIP⁵
Accuracy	0.878	0.912	0.886	0.883	0.895
Precision	0.897	0.895	0.893	0.893	0.904
Recall	0.860	0.940	0.880	0.880	0.890
F-Measure	0.876	0.917	0.889	0.886	0.895

ILP-1: Inductive Logic Programming with using POS(after) − (NEG(after) + POS(before))

ILP+BP: Inductive Logic Programming + Borderline Positives using POS(after) − (NEG(after) + POS(before) + FP(after))

PART: Java implementation of a rule based classifier in WEKA

⁴

J48: Java implementation of C4.5 classifier available in WEKA

⁵

JRIP: Java implementation of RIPPER rule-based classifier available in WEKA

Note: The results from a binomial classification (counting # wins for each method by phenotype), then using a two-sided sign test (binomial distribution test) at 95% confidence to determine if there is a difference. There was a significant difference favoring ILP+BP when compared to PART (p=0.039), J48 (p=0.003) and JRIP (p=0.003) when evaluating AUROC. There was no significant difference when testing accuracy, precision, recall, and F-Measure.