Table 9.

Summary of classification and prediction methods

				Classification challenge
Modality	Feature selection/dimensionality reduction	Classifier	N	AD versus CN		MCI versus CN		MCI versus AD		MCIc versus MCInc		Other		Unique approach	Reference
MRI	Multi-atlas	SVM	459	ACC	91.6					ACC	72.4			Subjects registered to multiple atlases, features selected from each atlas, then jointly selected	[372]
				SEN	88.6					SEN	72.1
				SPE	93.9					SPE	72.6
				AUC	0.87					AUC	0.67
MRI	Multi-atlas	SVM	459	ACC	90.7					ACC	73.7			Joint learning of optimum representation of features from multiple atlases and classifier using maximum margin approach	[373]
				SEN	87.6					SEN	76.4
				SPE	93.0					SPE	70.8
MRI	Multitemplate	SVM	459	ACC	93.3					ACC	80.9			Inherent structure-based multiview learning (ISML) clusters subjects within a specific class into subclasses based on features learned using multiple templates	[374]
				SEN	92.8					SEN	86.0
				SPE	95.7					SPE	78.4
MRI	Independent component analysis	SVM	818	ACC	86.4	ACC	70.2	ACC	69.8					Uses novel ICA-based feature extraction method and linear SVM for classification	[375]
				SEN	88.3	SEN	72.9	SEN	73.4
				SPE	84.0	SPE	67.5	SPE	66.2
MRI	Spatially weighted principal component analysis	SVM	390	Average misclassification %: 0.216										Incorporates spatial structure by using both global and local spatial weights for feature selection to improve on PCA alone	[376]
MRI	Circular harmonic function + principal component analysis	SVM	NA	ACC	83.8	ACC	69.5	ACC	62.1					Uses circular harmonic functions to extract local features from the hippocampus and posterior cingulate cortex and PCA for dimensionality reduction	[377]
				SEN	88.2	SEN	74.8	SEN	75.1
				SPE	79.1	SPE	62.5	SPE	49.0
MRI	Gaussian process models		415	AUC	0.94							AUC	0.91	Generates quantitative maps of z-scores for WM, GM, and CSF abnormalities for clinical decision support	[378]
												MCIc versus CN
MRI	Manifold learning: relevant variable selection (learned ROIs)	MR1 disease state score	363									ACC	61	Novel MRI disease state score (MRI DSS). Also used to predict MMSE scores and achieved separation between CN, EMCI, MCInc, MCIc, and AD groups	[379]
			ADNI-GO									SEN	50
												SPE	72
												EMCI versus CN
		SVM	511	ACC	86					ACC	71	ACC	82
			ADNI	SEN	86					SEN	75	SEN	81
				SPE	85					SPE	67	SPE	83
												MCIc versus CN
MRI	Manifold learning: self-organizing map	SVM	818	ACC	90			ACC	84					Uses self-organizing map to select maximally discriminative ROIs and computes relative importance measure for each ROI	[380]
				SEN	87			SEN	82
				SPE	92			SPE	87
				AUC	0.92			AUC	0.84
MRI	Manifold learning: local binary map + custom masks	SVM	363	ACC	82.8	ACC	61.5							Uses local binary maps to directly select disease-specific patterns on a voxelwise basis without nonrigid registration. Also uses seven knowledge-based ROIs	[381]
				SEN	80.4	SEN	61.5
				SPE	82.7	SPE	63.5
				AUC	0.87	AUC	0.64
MRI	Multivariate regression: elastic net	Low density separation	834							AUC	0.73			Used a semisupervised learning method, low density separation, which uses unlabeled data to improve prediction of progression over supervised methods	[382]
MRI	Multivariate regression: Ordinal regression	ORCHID score: probabilistic (not binary)	564	ACC	91					ACC	70			Models all clinical groups simultaneously as a continuum of disease, generates a summary score, ORCHID (Ordinal Regression Characteristic Index of Dementia)	[383]
				AUC	0.95					AUC	0.75
MRI	Multivariate regression: knockout strategy	Random Forests	245	SEN	98.0									Iterative multivariate knockout algorithm that uses Random Forests to construct set of relevant features	[384]
MRI	Tree-structured sparse learning	SVM	830	ACC	90.2					ACC	70.7	ACC	87.2	Uses new tree construction method to cluster similar discriminative voxels, then a tree-structured sparse learning step and SVM classifier	[385]
				SEN	85.3					SEN	56.2	SEN	80.1
				SPE	94.3					SPE	80.9	SPE	92.2
												MCIc versus CN
MRI	FreeSurfer to select ROIs, cortical thickness measures	v-one class SVM	814	ACC	84.3	ACC	54.4							Focuses on detection of outliers who have abnormal brain patterns in a cognitively normal group. Trained on CN. Produces a brain abnormality index.	[386]
				Labeled as outliers. In control group, ACC = 67.5%
MRI	Recursive feature elimination significance (p) maps from SVM weights		509	AUC	0.92	AUC	0.74	AUC	0.69	AUC	0.61			Uses p-maps of SVM weights as selection features in a wrapper method	[387]
MRI	Recursive feature elimination—SVM		370	ACC	95.1									Wrapper method based on recursive feature elimination and SVM selected GM and WM ROIs for classification with no need for dimensionality reduction step	[388]
MRI (VBM)	Probability distribution function based	SVM	260	ACC	89.7									Extracts statistical patterns at multiple levels including use of probability distribution function of VOIs to represent statistical patterns	[389]
				SEN	87.7
				SPE	91.6
				AUC	0.95
MRI	Hierarchical fusion of multilevel classifiers		652	ACC	92.0	ACC	85.3							Divides classification of high dimensional features into multiple low-dimensional classification problems and integrates features hierarchically	[390]
				SEN	90.9	SEN	82.3
				SPE	93.0	SPE	88.2
MRI	Predefined ROIs and routine feature ranking (Pearson Correlation Coefficient)	Distance informed metric learning	321	ACC	83.1	ACC	71.6							Semisupervised metric learning framework that uses pairwise constraints that specify the relative distance between a pair of patients, according to their classification group	[391]
				SEN	75.0	SEN	77.6
				SPE	91.2	SPE	65.6
MRI	Graph-based multiple instance learning	SVM	834	ACC	89.2					ACC	69.3			Extracts local intensity patches as features and uses a novel graph-based multiple instance learning approach to assign disease labels to patches	[392]
				SEN	85.1					SEN	66.7
				SPE	92.6					SPE	71.2
MRI	t-tests of predefined cortical thickness measures	C4.5 decision tree	364			SEN	80.0							Encodes selected features into an ontology and uses C4.5 algorithm for clinical decision support	[393]
						SPE	80.0
MRI	Pseudo-Zernike moments	Neural network	500	ACC	97.3	ACC	95.6	ACC	94.9					Uses Pseudo-Zernike moments (30) for feature selection and neural network scaled conjugate gradient back propagation algorithm as classifier	[394]
				SEN	96.6	SEN	95.9	SEN	94.2
				SPE	97.8	SPE	95.3	SPE	95.6
fMRI	Sparse inverse covariance estimation matrix + Kernel-based PCA	SVM	82									ACC	73.2	Uses a sparse inverse covariance estimation technique to model brain connectivity	[395]
												Note: EMCI
fMRI	Sparse learning using dictionary learning algorithm	SVM	210	ACC	94.1	ACC	92.0					ACC	92.3	Functional features are represented in a weighted dictionary matrix	[396]
												Note: SMC
FDG PET	Spatially weighted principal component regression		196	Average misclassification 0.117%										Spatially weighted selection of both individual features and neighboring patterns	[397]
FDG PET	Voxel-based longitudinal	SVM	233	ACC	92.6	ACC	70.2							Extension of [398] using whole-brain voxel approach for longitudinal analysis of FDG PET	[399]
FDG PET	Multi-channel pattern analysis + 3D Gabor wavelets feature extraction	SVM	369	MAP	56.3	MAP	76.2							Multi-channel approach integrates multiple patterns of hypometabolism from different patient groups using different analysis tools	[400]
FDG PET	Gaussian mixture model	SVM	84	ACC	89.1	ACC	63.2	ACC	80.2					Semi-data-driven approach to feature selection defines clusters of ROIs from an NC image and uses these to extract features from scans of MCI and AD patients	[401]
				SEN	92.0	SEN	65.0	SEN	80.0
				SPE	86.0	SPE	61.0	SPE	80.0
FDG PET	Optimization of preprocessing steps for SPM8	t-sum score	108							ACC	68.0			Modified preprocessing steps using motion correction, custom FDG template, different regions for intensity scaling and varied smoothing.	[402]
										SEN	70.0
										SPE	68.0
										AUC	0.83
MRI	Recursive feature elimination	Random Forests	575	ACC	89.6					SEN	78.0			Compares efficacy of Random Forest classifiers in several datasets including ADNI. RF model outperformed SVM, best model combined volumetric and cortical thickness measures with APOE	[403]
APOE4 demographics				SEN	90.7					Note: 12-month progression
				SPE	82.9
MRI	Regularized logistic regression	Low density separation (SVM)	825							AUC	0.90			Uses semisupervised learning to construct an aggregate biomarker by leading a separate MRI biomarker first and subsequently combining it with age and cognitive measures	[404]
Cognitive
MRI	Multivariate regression: LASSO based + novel loss function	SVM	202	ACC	95.9	ACC	82.0			ACC	72.6			Uses a novel loss function combined with a group lasso method for joint feature selection. Multimodal classification outperformed single modalities	[405]
PET				SEN	95.7	SEN	98.0			SEN	48.5
CSF				SPE	98.6	SPE	60.1			SPE	94.4
				AUC	0.99	AUC	87.0			AUC	0.79
MRI	Multivariate regression: fused LASSO + LDDMM	SVM	103							SEN	84.0			Comparison of spatial regularization techniques for detecting longitudinal hippocampal shape changes in MCI progression	[406]
HC shape longitudinal										SPE	48.0
										NPV	0.93
										PPV	0.27
MRI	Multivariate regression: weak hierarchical LASSO	Random Forest	293							ACC	74.8			Uses hierarchical constraints and sparsity regularization to capture underlying interactions between biosignatures	[407]
APOE4										SEN	66.7
Cognitive										SPE	81.4
Demographics
MRI	Multitask learning:: K means clustering + l 2,1 regression	SVM	202	ACC	95.1	ACC	79.5							Uses a novel clustering approach to discover mulitpeak data distributions. These are accounted for in a subsequent a multitask learning step in a l 2,1 regression framework	[408]
FDG PET				SEN	94.0	SEN	88.9
				SPE	96.3	SPE	62.0
				AUC	0.96	AUC	0.78
MRI	Multitask learning: K means clustering + l 2,1 regression	SVM	202	ACC	94.3	ACC	80.1	ACC	74.6					Similar approach to [408]. NOTE: MCI versus AD classification uses ONLY MRI + PET data	[409]
FDG PET				SEN	94.0	SEN	86.8	SEN	46.7
CSF				SPE	94.3	SPE	67.3	SPE	89.0
				AUC	0.96	AUC	0.82	AUC	0.72
MRI	Multitask learning: relational function + l 2,1 regression	SVM	202	ACC	95.7	ACC	79.9	ACC	72.4					Introduces a function to conserve information about feature-feature relations, response-response relations, and sample-sample relations while jointly solving both classification and prediction of continuous variables	[410]
FDG PET				SEN	96.6	SEN	97.0	SEN	49.1
				SPE	98.2	SPE	59.2	SPE	94.6
				AUC	0.98	AUC	0.85	AUC	0.83
MRI	Multitask learning: label-aligned regularization	SVM	202	ACC	95.6	ACC	80.3			ACC	69.8			Uses a label-aligned regularization term in multitask learning step to use relationships across subjects, as well as across modalities in the feature selection step	[411]
FDG PET				SEN	95.1	SEN	85.0			SEN	66.7
				SPE	96.5	SPE	70.8			SPE	71.4
				AUC	0.97	AUC	0.81			AUC	0.69
MRI	Multitask learning: graph-guided + latent		199	ACC	93.0	ACC	80.0							Uses new latent group LASSO penalty combined with an undirected graph approach to select correlated features that can jointly predict class label and clinical scores	[412]
FDG PET	LASSO
CSF	group penalty
MRI, demographics, clinical, APOE4, PICALM	Statistical learning embedded in SVM-based multiple kernel framework		213							ACC	91.0			Longitudinal, multimodal data are leveraged by an embedded novel statistical learning approach in a multikernel framework.	[413]
										SEN	95.0
										SPE	80.0
MRI	Domain transfer feature and sample selection	SVM	202							ACC	76.5			Uses both target domain (MCI) and auxiliary domains (CN and AD) to extract features from imaging and CSF modalities in domain transfer method. Fuses selected features with a domain transfer SVM	[414]
FDG PET										SEN	81.2
CSF										SPE	71.9
										AUC	0.84
MRI	Multimodal manifold-regularized transfer learning		202							ACC	80.1			Uses auxiliary domain information (as in [414]). Integrates kernel-based maximum mean discrepancy criterion and a manifold regularization function into single learning algorithm for both feature selection and classification	[415]
FDG PET										SEN	85.3
CSF										SPE	73.3
										AUC	0.85
MRI	Multimodal canonical correlation analysis		103	ACC	95.1									Used canonical correlation analysis to fuse data in multimodal classification as it preserves intermodal relationships.	[416]
FDG PET
CSF
MRI	Deep architecture + multitask learning	SVM	202	ACC	95.1	ACC	80.193.9			ACC	74.1			Uses novel deep learning architecture that discards uninformative features in a hierarchical manner during multitask learning	[417]
PET				SEN	92	SEN	53.7			SEN	50.5
CSF				SPE	98.3	SPE				SPE	92.7
				MRI/PET/CSF		MRI/PET				MRI/PET
MRI	Deep learning with Deep Boltzman Machine	Hierarchical SVM	398	ACC	95.4	ACC	85.7			ACC	75.9			Uses deep learning with a Boltzman Machine to find hierarchical feature representation from MRI features and then fuses these with complementary information from PET	[418]
PET				SEN	94.7	SEN	95.4			SEN	48
				SPE	95.2	SPE	65.9			SPE	95.2
				AUC	0.99	AUC	0.88			AUC	0.75
MRI	PCA/LASSO + multitask deep learning with dropout technique	SVM	202	ACC	91.4	ACC	77.4			ACC	57.4			Uses dropout technique as method of regularization for deep learning with small data	[419]
PET
CSF
Demographics
MRI	Independent component analysis	Discriminant classification analysis	320	ACC	94.3	ACC	83.3	ACC	84.1	ACC	80.0			Derived GM covariates patterns using independent component analysis and used for diagnostic classification in combination with cognitive performance	[304]
Demographics				SEN	94.9	SEN	76.7	SEN	86.1	SEN	78.3
				SPE	94.0	SPE	89.1	SPE	73.0	SPE	81.5
MRI	Stacked auto-encoder + multitask deep learning with zero-mask strategy	Softmax logistic regression	331	ACC	91.4	ACC	82.1							Uses stacked autoencoders to extract high-level features and a zero-masking strategy to extract synergy between modalities	[420]
PET				SEN	92.3	SEN	60.0
				SPE	90.4	SPE	92.3

Abbreviations: AD, Alzheimer’s disease; CN, cognitively normal; MCI, mild cognitive impairment; MCIc, mild cognitive impairment converters (progressive MCI); MCInc, mild cognitive impairment non-converters (stable MCI); MRI, magnetic resonance imaging; SVM, support vector machine; ACC, accuracy; SEN, sensitivity; SPE, specificity; AUC, area under the receiver operating curve; ICA, independent component analysis; PCA, principal component analysis; WM, white matter; GM, gray matter; CSF, cerebrospinal fluid; ADNI, Alzheimer’s Disease Neuroimaging Initiative; EMCI, early mild cognitive impairment; ROI, regions of interest; VBM, voxel-based morphometry; fMRI, functional magnetic resonance imaging; SMC, subjective memory concern; FDG PET, ¹⁸F-flurodeoxyglucose positron emission tomography; MAP, mean average precision; APOE4, apolipoprotein ε4 allele; PET, positron emission tomography; HC, hippocampus; LASSO, least absolute shrinkage and selection operator; LDDMM, large deformations via diffeomorphisms.