Table 3.

NPH prediction scores using various methods and features. All rows except the first row (Evan’s index) used a linear support vector machine for training and testing. The predictive models are trained and tested for 100 iterations using 5-fold cross-validation with randomized selection at each fold using scikit-learn, as explained in Pedregosa et al. [25].

	Precision (train/test)	Recall (train/test)
Evan’s index, thresholding	86	70
Volumetric features (model 1)	$86\pm 3$/$86\pm 13$	$80\pm 7$/$76\pm 17$
Network properties (model 2)	$80\pm 5$/$78\pm 17$	$85\pm 7$/$75\pm 21$
Volumetric features+network properties (model 3)	$96\pm 3$/$93\pm 12$	$89\pm 4$/$89\pm 13$