Table 5.

Performance of ML with targets PANSS, BDI on the ER40 (only SZ group was considered). Gini was used to rank top features. Input variables considered were 120 (cognitive: 40 RT, 40 corr and meta-cognitive: 40 CR). Target variable considered were categorical (0, 1) while all others were numeric. Here threshold refers to the value of target used for categorization. Thresholds for categorization for BDI was taken from (Chemerinski et al., 2008) while for PANSS_pos and PANSS_neg was based on the number closer to the mean of the respective distributions. Performance is shown for the best ranked features and model.

Target	Threshold	Method	Number of ranked featuresa, b	Count of (CR)a, b features	Count of negative featuresa	Count of negative CRa features	F1	AUC	Equivalent Cohen's dc
PANSS1_neg	15	Neural Network (ReLu)	35	6	19	3	0.65	0.57	0.25
PANSS1_pos	15	Tree	24	10	13	6	0.69	0.69	0.70
PANSS Reduced Emotional Experienced	See belowe	Stackf	50	17	33	13	0.86	0.78	0.81
BDI	9	Neural Network (tanh)	25	6	16	4	0.72	0.69	0.70
BDI	15	SVM (linear)	25	7	15	5	0.81	0.77	1.04
BDI	19	Random Forest	20	6	13	3	0.78	0.83	1.35
BDI	29	Neural Network (ReLu)	20	5	9	2	0.91	0.90	1.81

^aNegative features considered for ER40 are S, A and F.

^bPositive features considered for ER40 are N, H.

^cTable 2 from (Salgado, 2018).

^dThe items in the PANSS Reduced Emotional Experience factor are: Emotional Withdrawal (N2), Passive-apathetic Social Withdrawal (N4) and Active social avoidance (G16).

^e1 if any factor is >4 else 0.

^fComprising Naïve Bayes, Neural Network (ReLu), Random Forest, Tree.