Table 4.

Summary of recent and representative studies aiming to distinguish individuals with ASD from TD individuals using multivariate analysis of structural MRI. Reported sample sizes are the numbers used for classification and do not necessarily reflect the study’s total sample size.

Reference	Study Participants	Experimental Methods	Key Features	Multivariate Technique	Key Results
Ecker et al. (2010)192	22 adults with ASD and 22 TD adults	Used structural MRI to obtain images of grey and white matter regions	Voxels from grey matter images	SVM	Classified with 77% sensitivity and 86% specificity using leave-two-out cross-validation
Jiao et al. (2010)111	22 children with ASD and 16 TD children	Measured regional thickness and volumetric morphometry of 66 brain structures via MRI	7 thickness-based features and, separately, 16 volume-based features	Logistic model tree, among others	Attained 95%/77% sensitivity and 75%/69% specificity for best thickness-/volume-based classification with ten-fold cross-validation
Ingalhalikar et al. (2011)193	45 children and adolescents with ASD and 30 TD controls	Computed region-based fractional anisotropy and mean diffusivity maps for diffusion tensor imaging data	18 out of 352 fractional anisotropy/mean diffusivity features	Radial basis function kernel SVM	Achieved 80% accuracy, 74% sensitivity, and 84% specificity with leave-one-out cross-validation
Ingalhalikar et al. (2014)109	75 children with ASD and 37 TD children	Evaluated two functional tasks using MEG and 74 structural white matter features using diffusion tensor imaging	Two MEG features and 12 diffusion tensor imaging features	Ensemble of classifiers fused with weighted aggregation	Averaged 73% sensitivity and 86% specificity with five-fold cross-validation; 87% accuracy on testing set
Wee et al. (2014)194	58 children and adolescents with ASD and 59 TD controls	Used structural MRI to evaluate cortical-related morphology (regional and interregional features)	Combination of regional and interregional features	Multi-kernel SVM	Achieved an average of 96% sensitivity and 97% specificity with two-fold cross-validation
Gori et al. (2015)195	21 children with ASD and 20 TD children	Calculated brain features and global volumes of brain compartments from structural MRI data	314 region of interest features from the grey matter sub-region	SVM	Averaged 0.74 AUROC with leave-pair-out cross-validation
Jin et al. (2015)196	40 infants at high risk for ASD and 40 low-risk infants	Derived connectivity features from multiscale connectivity networks measured through MRI; compared high- and low-risk participants	Multiscale regions of interest and diffusion statistics	Multi-kernel SVM	Used nested five-fold cross-validation to obtain averages of 76% accuracy and 0.80 AUROC
Libero et al. (2015)110	19 adults with ASD and 18 TD adults	Analyzed brain morphometry from structural MRI, diffusion tensor imaging, and proton magnetic resonance spectroscopy data	Fractional anisotropy, radial diffusivity, and cortical thickness	Decision tree	Classified participants with 92% accuracy after leave-one-out cross validation
Hazlett et al. (2017)197	34 (145) infants at high risk for ASD with (without) a later diagnosis of ASD	Evaluated brain volume and surface area metrics from MRI at 6 and 12 months to predict ASD at 24 months	Regional surface area, intracranial volume, cortical thickness, and sex	Three-stage deep neural network	With ten-fold cross-validation, predicted ASD with 88% sensitivity and 95% specificity
Shen et al. (2017)198	47 (174) infants at high risk for ASD with (without) a later diagnosis of ASD	Quantified cerebrospinal fluid and lateral ventricle volume from MRI data collected at 6, 12, and 24 months to predict ASD diagnosis at 24 months	Extra-axial cerebrospinal fluid volume	Balance-boosted trees ensemble algorithm	Predicted ASD with 66% sensitivity and 68% specificity after 25-fold cross-validation; similar results on a validation set