. 2012 Aug 16;6:59. doi: 10.3389/fnsys.2012.00059

Table 4.

Classifier performance on the test dataset, by site and imaging modality.

Modality	Metric	Site
		KKI	NI	NYU	OHSU	Peking	All
Combined	Accuracy	0.73	0.68	0.37	0.76	0.57	0.59
	Sensitivity	0	0.45	0.34	0	0.25	0.24
	Specificity	1	0.86	0.58	0.93	0.96	0.85
	Subtype	–	1	0.8	–	0.5	0.71
Combined (no IQ)	Accuracy	0.73		0.41	0.76	0.47	0.55
	Sensitivity	0		0.41	0	0.21	0.2
	Specificity	1		0.58	0.93	0.82	0.79
	Subtype	–		0.83	–	0.4	0.59
Global connectivity	Accuracy	0.73				0.49	0.53
	Sensitivity	0				0.33	0.27
	Specificity	1				0.82	0.85
	Subtype	–				0.38	0.38
Regional homogeneity	Accuracy	0.55		0.34	0.71	0.49	0.5
	Sensitivity	0		0.41	0	0.21	0.2
	Specificity	0.75		0.5	0.86	0.85	0.74
	Subtype	–		0.67	–	0.4	0.52
Raw rs-fcMRI connectivity matrices	Accuracy		0.6	0.37		0.49	0.47
	Sensitivity		0.27	0.28		0.33	0.3
	Specificity		0.86	0.58		0.74	0.71
	Subtype		1	1		0.63	0.84
Nodal/global graph metrics	Accuracy		0.52	0.39	0.56		0.48
	Sensitivity		0.45	0.17	0.33		0.3
	Specificity		0.57	0.92	0.68		0.75
	Subtype		1	1	0		0.66
Power spectra	Accuracy		0.52	0.44	0.82	0.53	0.57
	Sensitivity		0.18	0.55	0.17	0.08	0.25
	Specificity		0.79	0.58	0.96	0.96	0.83
	Subtype		1	0.69	1	0.5	0.75
Freesurfer subcortical	Accuracy		0.8				0.8
	Sensitivity		0.55				0.55
	Specificity		1				1
	Subtype		1				1
Freesurfer cortical	Accuracy			0.51	0.62		0.56
	Sensitivity			0.41	0.33		0.38
	Specificity			0.75	0.71		0.73
	Subtype			1	0.5		0.77

Classifier metrics include 3-class accuracy, 2-class sensitivity/specificity, and ADHD subtype accuracy. Data for Pittsburgh and Brown sites were not available. Blank cells indicate modalities that were omitted from a particular final classifier due to feature ranking/selection predicting they were not expected to improve performance over chance.