Table 4.

Average AUC (area under the ROC curve) performance measures for the various feature selection and classifier combinations in the predictive modeling pipelines for the three different EHR data sets. The average and standard deviation is computed over 100 classifier runs (10 × 10-fold cross-validation).

Data Set	Feature Selection	Classification	Average AUC (std)
Small	Information Gain	K-Nearest Neighbor	0.680 (0.075)
		Naïve Bayesian	0.687 (0.042)
		Logistic Regression	0.690 (0.044)
		Random Forest	0.717 (0.045)
	Fisher Score	K-Nearest Neighbor	0.655 (0.102)
		Naïve Bayesian	0.690 (0.042)
		Logistic Regression	0.689 (0.046)
		Random Forest	0.713 (0.043)
Medium	Information Gain	K-Nearest Neighbor	0.598 (0.013)
		Naïve Bayesian	0.692 (0.013)
		Logistic Regression	0.746 (0.012)
		Random Forest	0.752 (0.012)
	Fisher Score	K-Nearest Neighbor	0.616 (0.013)
		Naïve Bayesian	0.688 (0.014)
		Logistic Regression	0.741 (0.012)
		Random Forest	0.749 (0.012)
Large	Information Gain	K-Nearest Neighbor	0.602 (0.027)
		Naïve Bayesian	0.634 (0.007)
		Logistic Regression	0.706 (0.006)
		Random Forest	0.705 (0.006)
	Fisher Score	K-Nearest Neighbor	0.597 (0.023)
		Naïve Bayesian	0.632 (0.007)
		Logistic Regression	0.705 (0.006)
		Random Forest	0.704 (0.006)