Table 4.

The performance of the ADOS diagnostic cut-off (= ADOS), performance of the RF models on the test set (= test) and the previously unseen validation data set (= val) for nonverbal children and young adolescents as well as older adolescents and adults

Module	No. of features	AUC ADOS	Sens. ADOS	Spec. ADOS	AUC test	ACC test	Sens. test	Spec. test	J	p McNe	AUC val	ACC val	Sens. val	Spec. val
Children and young adolescents	Female
	ADOS algorithm	0.83	0.81	0.84
	All 28 features				0.91	0.94	1	0.88	0.40		0.86	0.72	0.63	0.81
	5 features (optimal model)				0.92	0.93	0.97	0.91	0.43	.60	0.83	0.84	0.81	0.87
	Male
	ADOS algorithm	0.87	0.88	0.81
	All 28 features				0.93	0.89	0.93	0.86	0.44		0.79	0.86	0.85	0.88
	7 features (optimal model)				0.92	0.88	0.91	0.86	0.40	.14	0.81	0.85	0.85	0.85
Older adolescents and adults	Female
	ADOS algorithm	0.89	0.82	0.88
	All 31 features				0.83	0.88	0.91	0.82	0.46		0.92	0.83	0.93	0.72
	5 features (optimal model)				0.88	0.88	0.92	0.85	0.42	.18	0.86	0.78	0.84	0.72
	Male
	ADOS algorithm	0.85	0.85	0.72
	all 31 features				0.82	0.82	0.83	0.81	0.55		0.87	0.79	0.80	0.77
	8 features (optimal model)				0.82	0.80	0.84	0.76	0.48	.43	0.82	0.76	0.81	0.71

AUC area under the curve, ACC accuracy, Sens. sensitivity; Spec. specificity, J Youden’s index, McN McNemar level of significance—each model tested against the full-feature sets of available features