TABLE 1.
Summary of ASD classification studies based on fMRI data.
| Study | Participants | Data | Features | Feature selection | Machine learning method | Accuracy |
| Anderson et al. (2011) | 40 ASD 40 TC |
rs-fMRI data | Whole brain FCs between 7,266 ROIs | A two-tailed t-test for p < 0.001 | Not detailed | 79% |
| Wang et al. (2012) | 29 ASD 29 TC |
fMRI data with a cognitive control task | FCs between 106 ROIs (the AAL atlas) |
None | Logistic regression | 82.8% |
| Murdaugh et al. (2012) | 13 ASD 14 TC |
fMRI data with three stimuli experiments | Seed-based FCs FCs between 102 regions (the AAL atlas) |
None | Logistic regression | 96.3% |
| Nielsen et al. (2013) | 447 ASD 517 TC |
rs-fMRI data (the ABIDE dataset) |
Whole brain FCs between 7,266 ROIs | None | Not detailed | 60% |
| Uddin et al. (2013a) | 20 ASD 20 TC |
rs-fMRI data | Independent components | None | Logistic regression | 83% |
| Deshpande et al. (2013) | 15 ASD 15 TC |
fMRI data with ToM task | FCs between 18 ROIs | Recursive cluster elimination | Linear SVM | 95.9% |
| Just et al. (2014) | 17 ASD 17 TC |
fMRI data with a thinking task | Features obtained by factor analyses proposed by the author | The FA procedure | GNB | 97% |
| Price et al. (2014) | 30 ASD 30 TC |
rs-fMRI data (the ABIDE dataset) |
Dynamic FCs from multi-network | Self-proposed methods | Multi-kernel SVM | 90% |
| Zhou et al. (2014) | 127 ASD 153 TC |
rs-fMRI data (the ABIDE dataset) |
Integrated features | PCA and MRMR | SVM and Bayesian network | 70% |
| Plitt et al. (2015) | 59 ASD 59 TC |
rs-fMRI data | FCs between ROIs from three atlases (the Destrieux atlas, the DiMartino atlas, and the Power atlas) |
RFE | The scikit-learn library | 76.67%(peak) |
| Dodero et al. (2015a) | 42 ASD 37 TC |
rs-fMRI data (the UCLA data) |
FCs between 264 ROIs (the Power atlas) |
None | Grass–Kernel based Manifold Laplacian |
63.29% |
| Iidaka (2015) | 312 ASD 328 TC |
rs-fMRI data | FCs between 90 ROIs (the AAL atlas) |
Threshold | Probabilistic neural network | 89.4% |
| Chen et al. (2015) | 126 ASD 126 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 220 functionally defined ROIs | PSO RFE |
SVM Random forest |
66% 90.8% |
| Chanel et al. (2016) | 15 ASD 14 TC |
fMRI data with emotional stimuli | The beta maps | RFE | SVM | 92.3% |
| Ghiassian et al. (2016) | 538 ASD 573 TC |
rs-fMRI data (the ABIDE dataset) |
Proposed HOG features | MRMR | MHPC learning algorithm | 65% |
| Kassraian-Fard et al. (2016) | 77 ASD 77 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 200 ROIs (the CC200 atlas) |
None | SVM | 63% |
| Odriozola et al. (2016) | 23 ASD 22 TC |
fMRI data with two visual oddball detection tasks | Multivariate activation patterns in the dorsal part of the anterior insula | None | SVM | 85% |
| Abraham et al. (2017) | 871 participants | rs-fMRI data (the ABIDE dataset) |
FCs between ROIs from three atlases (the HO atlas, the Yeo atlas, and the CC200 atlas) |
ICA and MSDL | The scikit-learn library | 67% (peak) |
| Yahata et al. (2016) | 74 ASD 107 TC |
rs-fMRI data | FCs between 140 ROIs (the sulci-based anatomical atlas) |
L1-SCCA | SLR classifier | 85% |
| Dvornek et al. (2017) | 539 ASD 573 TC |
rs-fMRI data (the ABIDE dataset) |
The resting-state fMRI time series | None | LSTM modela | 68.5% |
| Rane et al. (2017) | 539 ASD 573 TC |
rs-fMRI data (the ABIDE dataset) |
All voxels within the GM mask | None | the scikit-learn library | 62% |
| Guo et al. (2017) | 55 ASD 55 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 116 ROIs (the AAL atlas) |
SAEs | DNN classifiera | 86.36% |
| Bi et al. (2018) | 45 ASD 39 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 90 ROIs (the AAL atlas) |
random SVM cluster | RBF-SVM | 96.15% |
| Heinsfeld et al. (2018) | 505 ASD 530 TC |
rs-fMRI data (the ABIDE dataset) |
Whole brain FC between 7,266 ROIs | SAEs | DNN classifiera | 70% |
| Aghdam et al. (2018) | 116 ASD 69 TC |
rs-fMRI and sMRI data (the ABIDE dataset) |
Means of ROIs respectively for rs-fMRI, GM and WM (the AAL atlas) |
None | DBN classifiera | 65.56% |
| Zhao et al. (2018) | 54 ASD 46 TC |
rs-fMRI data (the ABIDE dataset) |
Multi-level, high-order FCs | LASSO | multiple linear SVMs | 81% |
| Soussia and Rekik (2018) | 155 ASD 186 TC |
rs-fMRI data (the ABIDE dataset) |
High-Order Morphological Network | None | SIMLR based pairing + SVM | 61.7% |
| Dekhil et al. (2018) | 123 ASD 160 TC |
rs-fMRI data | PSD | PSD with highest correlation with the 34 rs-fMRI atlases | RBF-SVM | 91% |
| Bernas et al. (2018) | 24 ASD 30 TC |
rs-fMRI data | 7 resting-state networks | Group-ICA | poly-SVM | 86.7% |
| Bhaumik et al. (2018) | 167 ASD 205 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between Brodmann’s areas ROIs | Filter-based test and embedded Elastic Nets | Partial least square regression combined with SVM | 70% |
| Kong et al. (2018) | 78 ASD 104 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 148 ROIs (the Destrieux atlas) |
F-score | DNN classifiera | 90.39% |
| Li et al. (2018) | 38 ASD 23 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 90 ROIs (the AAL atlas) |
SSAE | DTL-NN classifiera | 70.4% |
| Kazeminejad and Sotero (2019) | 109 participants 342 participants 190 participants 137 participants 51 participants |
rs-fMRI data (the ABIDE dataset) |
FCs between 116 ROIs (the AAL atlas) |
A sequential forward floating algorithm | Gaussian SVM | 86% 69% 78% 80% 95% |
| Eslami et al. (2019) | 505 ASD 530 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 200 ROIs (the CC200 atlas) |
AE | A single layer perceptrona | 80% |
| Fredo et al. (2019) | 306 ASD 350 TC (400 participants for each sample) |
rs-fMRI (the ABIDE dataset) |
FCs between 237 ROIs (the Gordon’s cortical atlas the HO atlas) |
Conditional random forest | Random forest | 62.5% 65% 70% 73.75% |
| Niu et al. (2020) | 408 ASD 401 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between ROIs from three atlases separately (the AAL atlas, the HO atlas and the CC200 atlas) |
None | The proposed multichannel DANN | 73.2% |
| Liu Y. et al. (2020) | 506 ASD 548 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 200 ROIs (the CC200 atlas) |
Extra-tree | Linear-SVM | 72.2% |
| Sherkatghanad et al. (2020) | 505 ASD 530 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 392 ROIs (the CC400 atlas) |
None | CNN classifiera | 70.22% |
| Thomas et al. (2020) | 620 ASD 542 TC |
rs-fMRI data (the ABIDE dataset) |
Nine summary measures | None | 3D CNN classifiera | 64% |
| Tang et al. (2020) | 505 ASD 530 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 116 ROIs fMRI × ROI connectivity (the AAL atlas) |
None | DNN classifiera | 74% |
| Zhao et al. (2020) | 45 ASD 47 TC |
rs-fMRI data (the ABIDE dataset) |
FCs, Lo-D-FCs and Ho-D-FCs between 116 ROIs (the AAL atlas) |
A two-sample t-test and LASSO | Linear-SVM | 83% |
| Huang et al. (2020) | 505 ASD 530 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 200 ROIs (the CC200 atlas) |
Graph-based feature-selection method | DBN classifiera | 76.4% |
| Liu Y. et al. (2020) | 403 ASD 468 TC |
rs-fMRI data (the ABIDE dataset) |
D-FCs between ROIs (the AAL atlas) |
MTFS-EM | Multi-kernel SVM | 76.8% |
| Kazeminejad and Sotero (2020) | 493 ASD 530 TC |
rs-fMRI data (the ABIDE dataset) |
FCs between 200 ROIs (the CC200 atlas) |
PCA | A multilayer perceptrona | 64.4% |
| Yin et al. (2021) | 403 ASD 468 TC |
rs-fMRI data (the ABIDE dataset) |
FC between 264 ROIs (the Power atlas) |
An AE-based feature selection method | DNN classifiera | 79.2% |
| Yang et al. (2021) | 79 ASD 105 TC |
rs-fMRI data (the ABIDE dataset) |
8 brain functional networks from group-ICA | Dual regression | 3D CNN classifiera | 77.74% |
| Reiter et al. (2021) | 306 ASD 350 TC (400 participants for each sample) |
rs-fMRI data (the ABIDE dataset and data sample from SDSU) |
FC between 237 ROIs (the Gordon atlas the HO atlas) |
Conditional random forest | Random Forest | 62.5% 65% 70% 73.75% |
aMachine learning methods that are deep learning methods.
Abbreviations: ASD, autism spectrum disorder; fMRI, functional magnetic resonance imaging; rs-fMRI, resting state fMRI; TC, typical control; FC, functional connectivity; ROI, region of interest; AAL, Automated Anatomical Labeling; ABIDE, Autism Brain Imaging Data Exchange; ToM, Theory of Mind; SVM, support vector machine; FA, factor analysis; GNB, Gaussian naïve Bayes; PCA, principal component analysis; RFE, recursive feature elimination; MRMR, maximal relevance and minimal redundancy; UCLA, University of California at Los Angeles; PSO, particle swarm optimization; HOG, histogram of oriented gradients; ICA, independent component analysis; MSDL, multi-subject dictionary learning; L1-SCCA, the L1-norm regularized sparse canonical correlation analysis; SAEs, sparse auto-encoders; SLR, Structured Logistic Regression; LSTM, long short-term memory; DNN, deep neural network; DBN, deep belief network; RBF-SVM, radial basis function-support vector machine; GM, gray matter; WM, white matter; LASSO, least absolute shrinkage and selector operation; SSAE, a stacked sparse auto-encoder; SIMLR, Single-cell Interpretation via Multi-kernel LeaRning; PSD, Power spectral densities; DTL-NN, deep transfer learning neural network; AE, autoencoders; HO, Harvard Oxford; DANN, deep attention neural network; CNN, convolutional neural network; Lo-D-FCs, low-order dynamic functional connectivity networks; Ho-D-FCs, high-order dynamic functional connectivity networks; MTFS-EM, an improved multi-task feature selection method integrating elastic net and manifold regularization; SDSU, San Diego State University.