Table 4.
Machine-learning-based fMRI studies for PD diagnosis and early detection.
| Study | Sample | Data | Methods | Main findings | Other findings |
|---|---|---|---|---|---|
| Long et al., 201230 | 19 PD, 27 HC | fMRI: ALFF, ReHo, RFCS; MRI: volumes of GM, WM, CSF | Feature selection: 2-sample t-test; Classifier: SVM Validation: leave-one-out | PD showed decreased ALFF in ROL_L, decreased ReHo in bilateral ORBmid; increased RFCS in PHG_L, ANG_L, MTG_R; Classification accuracy: All modal: 87% (sensitivity: 79%; specificity: 93%) | PD showed increased GM in PCG, decreased GM in PCL_L and increased WM in regions such as PreCG_R; Classification accuracy: ReHo+ALFF + RFCS: 74%; ALFF + RFCS:67%; GM + WM + CSF: 80%; |
| Zhang et al., 201479 | 25 PD (15-tremor, 10 non-tremor), 20 HC | Regional network efficiencies (i.e., the local and global efficiencies) | Feature selection: nonparametric permutation tests and t-test; Classifier: Maximum uncertainty linear discriminant analysis; Validation: leave-one-out | Regions distinguishing between PD and HC: the limbic system (e.g., bilateral hippocampus and thalamus), basal ganglia (e.g., bilateral caudate and left putamen), cerebellum, insula and cingular cortex; Classification accuracy (PD vs. HC): 89% (sensitivity 100%, specificity 80%) | Classification accuracy: (tremor-PD vs. HC) 97%; (non-tremor-PD vs. HC) 90%; (tremor-PD vs. non-tremor-PD) 92% |
| Chen et al., 201564 | 21 PD, 26 HC | Network-based whole-brain FC | Feature selection: Kendall tau rank correlation coefficient comparison; Classifier: SVM; Validation: leave-one-out | The most discriminative FCs in: DMN, CO and FP networks and the cerebellum; Classification accuracy: 93.6% (sensitivity of 90.5% and a specificity of 96.2%) | Whole-brain functional connectivity might provide more information for discrimination than do any other characteristics (GM, WM, CSF, ALFF, ReHo and RFCS) |
| Gu et al., 201680 | 52 PD (19 PIGD, 25 TD, 8 mixed subtype), 45 HC | GM, WM, CSF volumes from MRI; FA, MD, RD, AD from DTI; ReHo and ALFF from resting-state fMRI | Feature selection: Recursive feature elimination; Classification: SVM; Validation: Leave-one-out | Classification accuracy: PIGD vs. non-PIGD 92.3% | The diagnostic agreement evaluated by the Kappa test showed Kappa value = 0.83 for agreement with the existing clinical categorization |
| Herz et al., 201681 | 12 PD with LID, 12 PD without LID | Seed-based FC in cortico-striatal network (between putamen and SMA, PSMC, and R IFG) | Feature selection: none; Classifier: SVM; ROC analysis; Linear regression analysis used to test whether FC could predict dyskinesia severity; Validation: leave-one-out | FC between putamen and PSMC increased after levodopa intake in No-LID pts and decreased in LID pts; Classification accuracy (LID vs. no-LID): 95.8% (91.7% Sensitivity;100% Specificity) | FC between putamen and PSMC predicted LID severity (R(2) = 0.627, P = .004); Volumes of putamen, PSMC or SMA did not distinguish LID from no-LID |
| Badea et al., 201762 | (1) NEUROCON: 27 PD, 16 HC; (2) PPMI: 91 PD, 18 HC; (3) Wu: 20 PD, 20 HC | FC obtained from ROI pairs (using parcellations such as Power 264 regions, Gordon 333 regions and Talairach 695 regions) | Feature selection: t-test; Classifier: SVM, Gaussian Naive Bayes; Validation: 50-fold cross-validation | Reproducibility of PD-related FC changes was low across the 3 datasets; Classification accuracy: 50~60% (trained and tested on the same dataset); <50% (trained on one dataset, tested on another) | Different parcellations revealed different FC decrease (between different ROI pairs) in PD |
| Pläschke et al., 201763 | 80 PD, 95 HC(old), 93 HC(young), 86 SCZ | FC from 12 networks such as motor network | Feature selection: Log-likelihood ratios Classification: SVM; ROC analysis; Log-likelihood ratios; Validation: tenfold cross-validation | FC in motor network had the best discrimination power between PD and HC (followed by memory and cognition networks); Classification accuracy: 70% (AUC: 0.77) | FC in all 12 networks performed better in young-old classification than other classifications (highest single network AUC: 0.93); FC in emotion processing, empathy and cognitive action control networks differentiate SCZ from HC (highest single network AUC: 0.79) |
| Tang et al., 201756 | 51 PD, 50 HC | ALFF, fALFF | Feature selection: t-test Classifier: SVM; Validation: leave-one-out | Altered ALFFs in the bilateral lingual gyrus and left putamen and an altered fALFF in the right posterior cerebellum; Classification accuracy: 84.2%; sensitivity 88.2%; specificity 80% | With un-optimized SVM classifier, the poorest classification performance was > 80%; Optimization of the classifier improved classification performance |
R right, L left, ALFF amplitude of low-frequency fluctuations, ANG_L left-angular gyrus, fALFF functional ALFF, LID levodopa-induced dyskinesias, MTG_R right middle-temporal gyrus, RFCS regional FC strength, ROL rolandic operculum, SMA supplementary motor area, mPFC mesial prefrontal cortex, PHG_L left parahippocampal gyrus, R MFC right middle-frontal gyrus, ROC receiver-operating characteristic analysis, PDRP PD-related pattern, DMN default mode network, CO cingulo-opercular, FP frontal-parietal, PPMI Parkinson’s progression markers initiative, PSMC primary sensorimotor cortex, IFG inferior frontal gyrus.