Table 3.

Comparative performance (ACC, SPE, SEN %) of our MCI/NC classifier vs. other methods.

Approach	Year	Dataset	Modalities	Validation Method	Metric
					Accuracy (%)	Sensitivity (%)	Specificity (%)
Our Method	2017	OASIS	MRI	semi-supervised method using 25% of the whole data set as training data ★	93.86	94.65	93.22
Hosseini-Asl et al. [10]	2016	ADNI	MRI	10-fold cross-validation	90.8	n/a	n/a
Zu et al. [42]	2016	ADNI	PET+MRI	10-fold cross-validation	80.26	84.95	70.77
Moradi et al. [43]	2015	ADNI	MRI	10-fold cross-validation	82	87	74
Liu et al. [5]	2015	ADNI	MRI	10-fold cross-validation	71.98	49.52	84.31
Suk et al. [3]	2014	ADNI	PET+MRI	10-fold cross-validation	85.7	99.58	53.79
Casanova et al. [44]	2013	ADNI	Only cognitive measures	10-fold cross-validation	65	58	70
Chyzhyk et al. [45]	2012	OASIS	MRI	10-fold cross-validation	74.25	96	52.5
Coupé et al. [46]	2012	ADNI	MRI	Leave-one-out cross-validation	74	73	74
Cho et al. [47]	2012	ADNI	MRI	Independent test set	71	63	76
Cheng et al. [48]	2012	ADNI	MRI	10-fold cross-validation	69.4	64.3	73.5
Savio et al. [49]	2011	OASIS	MRI	10-fold cross-validation	84	90	77
Westman et al. [50]	2011	ADNI	MRI	10-fold cross-validation	59	74	56
Chyzhyk et al. [51]	2011	OASIS	MRI	10-fold cross-validation	69	81	56
Savio et al. [32]	2009	OASIS	MRI	10-fold cross-validation	83	74	92
Chupin et al. [52]	2009	ADNI	MRI	Independent test set	64	60	65
García-Sebastián et al. [33]	2009	OASIS	MRI	Independent test set	80.61	89	75
Savio et al. [34]	2009	OASIS	MRI	10-fold cross-validation	85	78	92

★ All the existing methods use supervised learning while our proposed model utilizes a semi-supervised learning method which can further justify its efficiency. ACC: Accuracy, SPE: Specificity, SEN: Sensitivity, PET: Positron Emission Tomography, n/a: Not Available, MCI: mild cognitive impairment; NC: normal condition.