Skip to main content
. 2021 May 6;13:633752. doi: 10.3389/fnagi.2021.633752

Table 6.

Studies that applied machine learning models to MRI data to diagnose PD (n = 36).

Objectives Type of diagnosis Source of data Number of
subjects (n)		 Machine learning method(s), splitting strategy and cross validation Outcomes Year References
Classification of PD from MSA Differential diagnosis Collected from participants 150; 54 HC + 65 PD + 31 MSA SVM with leave-one-out-cross validation MSA vs. PD: 2019 Abos et al., 2019
Accuracy = 0.79
Sensitivity = 0.71
Specificity = 0.86
MSA vs. HC:
Accuracy = 0.79
Sensitivity = 0.84
Specificity = 0.74
MSA vs. subsample of PD:
Accuracy = 0.84
Sensitivity = 0.77
Specificity = 0.90
Classification of PD from MSA Differential diagnosis Collected from participants 151; 59 HC + 62 PD + 30 MSA SVM with leave-one-out-cross validation Accuracy = 77.17% 2019 Baggio et al., 2019
Sensitivity = 83.33%
Specificity = 74.19%
Classification of PD from HC Diagnosis Collected from participants 94; 50 HC + 44 PD CNN with 85 subjects for training and 9 for testing Training accuracy = 95.24% 2019 Banerjee et al., 2019
Testing accuracy = 88.88%
Classification of PD from HC Diagnosis Collected from participants 47; 26 HC + 21 PD SVM-linear with leave-one-out cross validation Accuracy = 93.62% 2015 Chen et al., 2015
Sensitivity = 90.47%
Specificity = 96.15%
Classification of PD from PSP Differential diagnosis Collected from participants 78; 57 PD + 21 PSP SVM with leave-one-out cross validation Accuracy = 100% 2013 Cherubini et al., 2014a
Sensitivity = 1
Specificity = 1
Classification of PD, MSA, PSP and HC Diagnosis and differential diagnosis Collected from participants 106; 36 HC + 35 PD + 16 MSA + 19 PSP Elastic Net regularized logistic regression with nested 10-fold cross validation HC vs. PD/MSA-P/PSP: 2017 Du et al., 2017
AUC = 0.88
Sensitivity = 0.80
Specificity = 0.83
PPV = 0.82
NPV = 0.81
HC vs. PD:
AUC = 0.91
Sensitivity = 0.86
Specificity = 0.80
PPV = 0.82
NPV = 0.89
PD vs. MSA/PSP:
AUC = 0.94
Sensitivity = 0.86
Specificity = 0.87
PPV = 0.88
NPV = 0.84
PD vs. MSA:
AUC = 0.99
Sensitivity = 0.97
Specificity = 1.00
PPV = 1.00
NPV = 0.93
PD vs. PSP:
AUC = 0.99
Sensitivity = 0.97
Specificity = 1.00
PPV = 1.00
NPV = 0.94
MSA vs. PSP:
AUC = 0.98
Sensitivity = 0.94
Specificity = 1.00
PPV = 1.00
NPV = 0.93
Classification of HC, PD, MSA and PSP Diagnosis and differential diagnosis Collected from participants 64; 22 HC + 21 PD + 11 MSA + 10 PSP SVM-linear with leave-one-out cross validation PD vs. HC: 2011 Focke et al., 2011
Accuracy = 41.86%
Sensitivity = 38.10%
Specificity = 45.45%
PD vs. MSA:
Accuracy = 71.87%
Sensitivity = 36.36%
Specificity = 90.48%
PD vs. PSP:
Accuracy = 96.77%
Sensitivity = 90%
Specificity = 100%
MSA vs. PSP:
Accuracy = 76.19%
MSA vs. HC:
Accuracy = 78.78%
Sensitivity = 54.55%
Specificity = 90.91%
PSP vs. HC:
Accuracy = 93.75%
Sensitivity = 90.00%
Specificity = 95.45%
Classification of PD and atypical PD Differential diagnosis Collected from participants 40; 17 PD + 23 atypical PD SVM-RBF with 10-fold cross-validation Accuracy = 97.50% 2012 Haller et al., 2012
TPR = 0.94
FPR = 0.00
TNR = 1.00
FNR = 0.06
Classification of PD and other forms of Parkinsonism Differential diagnosis Collected from participants 36; 16 PD + 20 other Parkinsonism SVM-RBF with 10-fold cross validation Accuracy = 86.92% 2012 Haller et al., 2013
TP = 0.87
FP = 0.14
TN = 0.87
FN = 0.13
Classification of HC, PD, PSP, MSA-C and MSA-P Diagnosis and differential diagnosis Collected from participants 464; 73 HC + 204 PD + 106 PSP + 21 MSA-C + 60 MSA-P SVM-RBF with 10-fold cross validation PD vs. HC: 2016 Huppertz et al., 2016
Sensitivity = 65.2%
Specificity = 67.1%
Accuracy = 65.7%
PD vs. PSP:
Sensitivity = 82.5%
Specificity = 86.8%
Accuracy = 85.3%
PD vs. MSA-C:
Sensitivity = 76.2%
Specificity = 96.1%
Accuracy = 94.2%
PD vs. MSA-P:
Sensitivity = 86.7%
Specificity = 92.2%
Accuracy = 90.5%
Classification of PD from HC Diagnosis Collected from participants 42; 21 HC + 21 PD SVM-linear with stratified 10-fold cross validation Accuracy = 78.33% 2017 Kamagata et al., 2017
Precision = 85.00%
Recall = 81.67%
AUC = 85.28%
Classification of PD, PSP, MSA-P and HC Diagnosis and differential diagnosis Collected from participants 419; 142 HC + 125 PD + 98 PSP + 54 MSA-P CNN with train-validation ratio of 85:15 PD: 2019 Kiryu et al., 2019
Sensitivity = 94.4%
Specificity = 97.8%
Accuracy = 96.8%
AUC = 0.995
PSP:
Sensitivity = 84.6%
Specificity = 96.0%
Accuracy = 93.7%
AUC = 0.982
MSA-P:
Sensitivity = 77.8%
Specificity = 98.1%
Accuracy = 95.2%
AUC = 0.990
HC:
Sensitivity = 100.0%
Specificity = 97.5%
Accuracy = 98.4%
AUC = 1.000
Classification of PD from HC Diagnosis Collected from participants 65; 31 HC + 34 PD FCP with 36 out of the 65 subjects as the training set AUC = 0.997 2016 Liu H. et al., 2016
Classification of PD, PSP, MSA-C and MSA-P Differential diagnosis Collected from participants 85; 47 PD + 22 PSP + 9 MSA-C + 7 MSA-P SVM-linear with leave-one-out cross validation 4-class classification (MSA-C vs. MSA-P vs. PSP vs. PD) accuracy = 88% 2017 Morisi et al., 2018
Classification of PD from HC Diagnosis Collected from participants 89; 47 HC + 42 PD Boosted logistic regression with nested cross-validation Accuracy = 76.2% 2019 Rubbert et al., 2019
Sensitivity = 81%
Specificity = 72.7%
Classification of PD, PSP and HC Diagnosis and differential diagnosis Collected from participants 84; 28 HC + 28 PSP + 28 PD SVM-linear with leave-one-out cross validation PD vs. HC: 2014 Salvatore et al., 2014
Accuracy = 85.8%
Specificity = 86.0%
Sensitivity = 86.0%
PSP vs. HC:
Accuracy = 89.1%
Specificity = 89.1%
Sensitivity = 89.5%
PSP vs. PD:
Accuracy = 88.9%
Specificity = 88.5%
Sensitivity = 89.5%
Classification of PD, APS (MSA, PSP) and HC Diagnosis and differential diagnosis Collected from participants 100; 35 HC + 45 PD + 20 APS CNN-DL, CR-ML, RA-ML with 5-fold cross-validation PD vs. HC with CNN-DL: 2019 Shinde et al., 2019
Test accuracy = 80.0%
Test sensitivity = 0.86
Test specificity = 0.70
Test AUC = 0.913
PD vs. APS with CNN-DL:
Test accuracy = 85.7%
Test sensitivity = 1.00
Test specificity = 0.50
Test AUC = 0.911
Classification of PD from HC Diagnosis Collected from participants 101; 50 HC + 51 PD SVM-RBF with leave-one-out cross validation Sensitivity = 92%
Specificity = 87%		 2017 Tang et al., 2017
Classification of PD from HC Diagnosis Collected from participants 85; 40 HC + 45 PD SVM-linear with leave-one-out, 5-fold, 0.632-fold (1-1/e), 2-fold cross validation Accuracy = 97.7% 2016 Zeng et al., 2017
Classification of PD from HC Diagnosis PPMI database 543; 169 HC + 374 PD RLDA with JFSS with 10-fold cross validation Accuracy = 81.9% 2016 Adeli et al., 2016
Classification of PD from HC Diagnosis PPMI database 543; 169 HC + 374 PD RFS-LDA with 10-fold cross validation Accuracy = 79.8% 2019 Adeli et al., 2019
Classification of PD from HC Diagnosis PPMI database 543; 169 HC + 374 PD Random forest (for feature selection and clinical score); SVM with 10-fold stratified cross validation Accuracy = 0.93 2018 Amoroso et al., 2018
AUC = 0.97
Sensitivity = 0.93
Specificity = 0.92
Classification of PD, HC and prodromal Diagnosis PPMI database 906; 203 HC + 66 prodromal + 637 PD MLP, XgBoost, random forest, SVM with 5-fold cross validation MLP: 2020 Chakraborty et al., 2020
Accuracy = 95.3%
Recall = 95.41%
Precision = 97.28%
F1-score = 94%
Classification of PD from HC Diagnosis PPMI database Dataset 1: 15; 6 HC + 9 PD SVM with leave-one-out cross validation Dataset 1: 2014 Chen et al., 2014
EER = 87%
Dataset 2: 39; 21 HC + 18 PD Accuracy = 80%
AUC = 0.907
Dataset 2:
EER = 73%
Accuracy = 68%
AUC = 0.780
Classification of PD from HC Diagnosis PPMI database 80; 40 HC + 40 PD Naïve Bayes, SVM-RBF with 10-fold cross validation SVM: 2019 Cigdem et al., 2019
Accuracy = 87.50%
Sensitivity = 85.00%
Specificity = 90.00%
AUC = 90.00%
Classification of PD from HC Diagnosis PPMI database 37; 18 HC + 19 PD SVM-linear with leave-one-out cross validation Accuracy = 94.59% 2017 Kazeminejad et al., 2017
Classification of PD, HC and SWEDD Diagnosis and subtyping PPMI database 238; 62 HC + 142 PD + 34 SWEDD Joint learning with 10-fold cross validation HC vs. PD: 2018 Lei et al., 2019
Accuracy = 91.12%
AUC = 94.88%
HC vs. SWEDD:
Accuracy = 94.89%
AUC = 97.80%
PD vs. SWEDD:
accuracy = 92.12%
AUC = 93.82%
Classification of PD and SWEDD from HC Diagnosis PPMI database Baseline: 238; 62 HC + 142 PD + 34 SWEDD12 months: 186; 54 HC + 123 PD + 9 SWEDD
24 months: 127; 7 HC + 88 PD + 22 SWEDD		 SSAE with 10-fold cross validation
 HC vs. PD:
Accuracy = 85.24%, 88.14%, and 96.19% for baseline, 12 m, and 24 mHC vs. SWEDD:
Accuracy = 89.67%, 95.24%, and 93.10% for baseline, 12 m, and 24 m		 2019 Li et al., 2019
Classification of PD from HC Diagnosis PPMI database 112; 56 HC + 56 PD RLDA with 8-fold cross validation Accuracy = 70.5% 2016 Liu L. et al., 2016
AUC = 71.1
Classification of PD from HC Diagnosis PPMI database 60; 30 HC + 30 PD SVM, ELM with train-test ratio of 80:20 ELM: 2016 Pahuja and Nagabhushan, 2016
Training accuracy = 94.87%
Testing accuracy = 90.97%
Sensitivity = 0.9245
Specificity = 0.9730
Classification of PD from HC Diagnosis PPMI database 172; 103 HC + 69 PD Multi-kernel SVM with 10-fold cross validation 2017 Peng et al., 2017
Accuracy = 85.78%
Specificity = 87.79%
Sensitivity = 87.64%
AUC = 0.8363
Classification of PD from HC Diagnosis and subtyping PPMI database 109; 32 HC + 77 PD (55 PD-NC + 22 PD-MCI) SVM with 2-fold cross validation PD vs. HC: 2016 Peng et al., 2016
Accuracy = 92.35%
Sensitivity = 0.9035
Specificity = 0.9431
AUC = 0.9744
PD-MCI vs. HC:
Accuracy = 83.91%
Sensitivity = 0.8355
Specificity = 0.8587
AUC = 0.9184
PD-MCI vs. PD-NC:
Accuracy = 80.84%
Sensitivity = 0.7705
Specificity = 0.8457
AUC = 0.8677
Classification of PD, HC and SWEDD Diagnosis and subtyping PPMI database 831; 245 HC + 518 PD + 68 SWEDD LSSVM-RBF with cross validation Accuracy = 99.9%
Specificity = 100%
Sensitivity = 99.4%		 2015 Singh and Samavedham, 2015
Classification of PD, HC and SWEDD Diagnosis and differential diagnosis PPMI database 741; 262 HC + 408 PD + 71 SWEDD LSSVM-RBF with 10-fold cross validation PD vs. HC accuracy = 95.37% 2018 Singh et al., 2018
PD vs. SWEDD accuracy = 96.04%
SWEDD vs. HC accuracy = 93.03%
Classification of PD from HC Diagnosis PPMI database 408; 204 HC + 204 PD CNN (VGG and ResNet) ResNet50 accuracy = 88.6% 2019 Yagis et al., 2019
Classification of PD from HC Diagnosis PPMI database 754; 158 HC + 596 PD FCN, GCN with 5-fold cross validation AUC = 95.37% 2018 Zhang et al., 2018

APS, atypical parkinsonian syndromes; AUC, area under the receiver operating characteristic (ROC) curve; CNN, convolutional neural network; CNN-DL, convolutional neural network with discriminative localization; CR-ML, contrast ratio classifier; EER, equal error rate; ELM, extreme learning machine; FCN, fully connected network; FCP, folded concave penalized (learning); FN, false negative; FNR, false negative rate; FP, false positive; FPR, false positive rate; GCN, graph convolutional network; HC, healthy control; JFSS, joint feature-sample selection; LSSVM, least-squares support vector machine; MLP, multilayer perceptron; MSA, multiple system atrophy; MSA-C, multiple system atrophy with a cerebellar syndrome; MSA-P, multiple system atrophy with a parkinsonian type; PD, Parkinson's disease; PD-MCI, PD participants who met criteria for mild cognitive impairment; PD-NC, PD participants with no indication of mild cognitive impairment; PSP, progressive supranuclear palsy; RA-ML, radiomics based classifier; ResNet, residual neural network; RFS-LDA, robust feature-sample linear discriminant analysis; RLDA, robust linear discriminant analysis; SSAE, stacked sparse auto-encoder; SVM, support vector machine; SVM-RBF, support vector machine with radial basis function kernel; SWEDD, PD with scans without evidence of dopaminergic deficit; TN, true negative; TNR, true negative rate; TP, true positive; TPR, true positive rate; XgBoost, extreme gradient boosting.