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

Table 5.

Studies that applied machine learning models to movement data to diagnose PD (n = 51).

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 HC Diagnosis Collected from participants 103; 71 HC + 32 PD Ensemble method of 8 models (SVM, MLP, logistic regression, random forest, NSVC, decision tree, KNN, QDA) Sensitivity = 96%
Specificity = 97%
AUC = 0.98		 2017 Adams, 2017
Classification of PD, HC and other neurological stance disorders Diagnosis and differential diagnosis Collected from participants 293; 57 HC + 27 PD + 49 AVS + 12 PNP + 48 CA + 16 DN + 25 OT + 59 PPV Ensemble method of 7 models (logistic regression, KNN, shallow and deep ANNs, SVM, random forest, extra-randomized trees) with 90% training and 10% testing data in stratified k-fold cross-validation 8-class classification accuracy = 82.7% 2019 Ahmadi et al., 2019
Classification of PD from HC Diagnosis Collected from participants 137; 38 HC + 99 PD SVM with leave-one-out-cross validation PD vs. HC accuracy = 92.3% 2016 Bernad-Elazari et al., 2016
Mild vs. severe accuracy = 89.8%
Mild vs. HC accuracy = 85.9%
Classification of PD from HC Diagnosis Collected from participants 30; 14 HC + 16 PD SVM (linear, quadratic, cubic, Gaussian kernels), ANN, with 5-fold cross-validation Classification with ANN: 2019 Buongiorno et al., 2019
Accuracy = 89.4%
Sensitivity = 87.0%
Specificity = 91.8%
Severity assessment with ANN:
Accuracy = 95.0%
sensitivity = 90.0%
Specificity = 99.0%
Classification of PD from HC Diagnosis Collected from participants 28; 12 HC + 16 PD NN with a train-validation-test ratio of 70:15:15, SVM with leave-one-out cross-validation, logistic regression with 10-fold cross validation SVM:
Accuracy = 85.71%
Sensitivity = 83.5%
Specificity = 87.5%		 2017 Butt et al., 2017
Classification of PD from HC Diagnosis Collected from participants 28; 12 HC + 16 PD Logistic regression, naïve Bayes, SVM with 10-fold cross validation Naïve Bayes: 2018 Butt et al., 2018
Accuracy = 81.45%
Sensitivity = 76%
Specificity = 86.5%
AUC = 0.811
Classification of PD from HC Diagnosis Collected from participants 54; 27 HC + 27 PD Naïve Bayes, LDA, KNN, decision tree, SVM-linear, SVM-RBF, majority of votes with 5-fold cross validation Majority of votes (weighted) accuracy = 96% 2018 Caramia et al., 2018
Classification of PD, HC and PD, HC, IH Diagnosis Collected from participants 90; 30 PD + 30 HC + 30 IH SVM, random forest, naïve Bayes with 10-fold cross validation Random forest: 2019 Cavallo et al., 2019
HC vs. PD:
Accuracy = 0.950
F-measure = 0.947
HC + IH vs. PD:
Accuracy = 0.917
F-measure = 0.912
HC vs. IH vs. PD:
Accuracy = 0.789
F-measure = 0.796
Classification of PD from HC and classification of HC, MCI, PDNOMCI, and PDMCI Diagnosis, differential diagnosis and subtyping Collected from participants PD vs. HC: Decision tree, naïve Bayes, random forest, SVM, adaptive boosting (with decision tree or random forest) with 10-fold cross validation Adaptive boosting with decision tree: 2015 Cook et al., 2015
75; 50 HC + 25 PD PD vs. HC:
Accuracy = 0.79
Subtyping: AUC = 0.82
52; 18 HC + 16 PDNOMCI + 9 PDMCI + 9 MCI Subtyping (HOA vs. MCI vs. PDNOMCI vs. PDMCI):
Accuracy = 0.85
AUC = 0.96
Classification of PD from HC Diagnosis Collected from participants 580; 424 HC + 156 PD Hidden Markov models with nearest neighbor classifier with cross validation and train-test ratio of 66.6:33.3 Accuracy = 85.51% 2017 Cuzzolin et al., 2017
Classification of PD from HC Diagnosis Collected from participants 80; 40 HC + 40 PD Random forest, SVM with 10-fold cross validation SVM-RBF: 2017 Djurić-Jovičić et al., 2017
Accuracy = 85%
Sensitivity = 85%
Specificity = 82%
PPV = 86%
NPV = 83%
Classification of PD from HC Diagnosis Collected from participants 13; 5 HC + 8 PD SVM-RBF with leave-one-out cross validation 100% HC and PD classified correctly (confusion matrix) 2014 Dror et al., 2014
Classification of PD from HC Diagnosis Collected from participants 75; 38 HC + 37 PD SVM with leave-one-out cross validation Accuracy = 85.61% 2014 Drotár et al., 2014
Sensitivity = 85.95%
Specificity = 85.26%
Classification of PD from ET Differential diagnosis Collected from participants 24; 13 PD + 11 ET SVM-linear, SVM-RBF with leave-one-out cross validation Accuracy = 83% 2016 Ghassemi et al., 2016
Classification of PD from HC Diagnosis Collected from participants 41; 22 HC + 19 PD SVM, decision tree, random forest, linear regression with 10-fold and leave-one-individual out (L1O) cross validation SVM accuracy = 0.89 2018 Klein et al., 2017
Classification of PD from HC Diagnosis Collected from participants 74; 33 young HC + 14 elderly HC + 27 PD SVM with 10-fold cross validation Sensitivity = ~90% 2017 Javed et al., 2018
Classification of PD from HC and assess the severity of PD Diagnosis Collected from participants 55; 20 HC + 35 PD SVM with leave-one-out cross validation PD diagnosis: 2016 Koçer and Oktay, 2016
Accuracy = 89%
Precision = 0.91
Recall = 0.94
Severity assessment:
HYS 1 accuracy = 72%
HYS 2 accuracy = 77%
HYS 3 accuracy = 75%
HYS 4 accuracy = 33%
Classification of PD from HC Diagnosis Collected from participants 45; 20 HC + 25 PD Naïve Bayes, logistic regression, SVM, AdaBoost, C4.5, BagDT with 10-fold stratified cross-validation apart from BagDT BagDT:
Sensitivity = 82%
Specificity = 90%
AUC = 0.94		 2015 Kostikis et al., 2015
Classification of PD from HC Diagnosis Collected from participants 40; 26 HC + 14 PD Random forest with leave-one-subject-out cross-validation Accuracy = 94.6%
Sensitivity = 91.5%
Specificity = 97.2%		 2017 Kuhner et al., 2017
Classification of PD from HC Diagnosis Collected from participants 177; 70 HC + 107 PD ESN with 10-fold cross validation AUC = 0.852 2018 Lacy et al., 2018
Classification of PD from HC Diagnosis Collected from participants 39; 16 young HC + 12 elderly HC + 11 PD LDA with leave-one-out cross validation Multiclass classification (young HC vs. age-matched HC vs. PD): 2018 Martínez et al., 2018
Accuracy = 64.1%
Sensitivity = 47.1%
Specificity = 77.3%
Classification of PD from HC Diagnosis Collected from participants 38; 10 HC + 28 PD SVM-Gaussian with leave-one-out cross validation Training accuracy = 96.9% 2018 Oliveira H. M. et al., 2018
Test accuracy = 76.6%
Classification of PD from HC Diagnosis Collected from participants 30; 15 HC + 15 PD SVM-RBF, PNN with 10-fold cross validation SVM-RBF: 2015 Oung et al., 2015
Accuracy = 88.80%
Sensitivity = 88.70%
Specificity = 88.15%
AUC = 88.48
Classification of PD from HC Diagnosis Collected from participants 45; 14 HC + 31 PD Deep-MIL-CNN with LOSO or RkF With LOSO: 2019 Papadopoulos et al., 2019
Precision = 0.987
Sensitivity = 0.9
specificity = 0.993
F1-score = 0.943
With RkF:
Precision = 0.955
Sensitivity = 0.828
Specificity = 0.979
F1-score = 0.897
Classification of PD, HC and post-stroke Diagnosis and differential diagnosis Collected from participants 11; 3 HC + 5 PD + 3 post-stroke MTFL with 10-fold cross validation PD vs. HC AUC = 0.983 2017 Papavasileiou et al., 2017
Classification of PD from HC Diagnosis Collected from participants 182; 94 HC + 88 PD LSTM, CNN-1D, CNN-LSTM with 5-fold cross-validation and a training-test ratio of 90:10 CNN-LSTM: 2019 Reyes et al., 2019
Accuracy = 83.1%
Precision = 83.5%
Recall = 83.4%
F1-score = 81%
Kappa = 64%
Classification of PD from HC Diagnosis Collected from participants 60; 30 HC + 30 PD Naïve Bayes, KNN, SVM with leave-one-out cross validation SVM: 2019 Ricci et al., 2020
Accuracy = 95%
Precision = 0.951
AUC = 0.950
Classification of PD, HC and IH Diagnosis and differential diagnosis Collected from participants 90; 30 HC + 30 PD + 30 IH SVM-polynomial, random forest, naïve Bayes with 10-fold cross validation HC vs. PD, naïve Bayes or random forest: 2018 Rovini et al., 2018
Precision = 0.967
Recall = 0.967
Specificity = 0.967
Accuracy = 0.967
F-measure = 0.967
HC + IH vs. PD, random forest:
Precision = 1.000
Recall = 0.933
Specificity = 1.000
Accuracy = 0.978
F-measure = 0.966
Multiclass classification, random forest:
Precision = 0.784
Recall = 0.778
Specificity = 0.889
Accuracy = 0.778
F-measure = 0.781
Classification of PD, HC and IH Diagnosis and differential diagnosis Collected from participants 45; 15 HC + 15 PD + 15 IH SVM-polynomial, random forest with 5-fold cross validation HC vs. PD, random forest: 2019 Rovini et al., 2019
Precision = 1.000
Recall = 1.000
Specificity = 1.000
Accuracy = 1.000
F-measure = 1.000
Multiclass classification (HC vs. IH vs. PD), random forest:
Precision = 0.930
Recall = 0.911
Specificity = 0.956
Accuracy = 0.911
F-measure = 0.920
Classification of PD from ET Differential diagnosis Collected from participants 52; 32 PD + 20 ET SVM-linear with 10-fold cross validation Accuracy = 1 2016 Surangsrirat et al., 2016
Sensitivity = 1
Specificity = 1
Classification of PD from HC Diagnosis Collected from participants 12; 10 HC + 2 PD Naive Bayes, LogitBoost, random forest, SVM with 10-fold cross-validation Random forest: 2017 Tahavori et al., 2017
Accuracy = 92.29%
Precision = 0.99
Recall = 0.99
Classification of PD from HC Diagnosis Collected from participants 39; 16 HC + 23 PD SVM-RBF with 10-fold stratified cross validation Sensitivity = 88.9% 2010 Tien et al., 2010
Specificity = 100%
Precision = 100%
FPR = 0.0%
Classification of PD from HC Diagnosis Collected from participants 60; 30 HC + 30 PD Logistic regression, naïve Bayes, random forest, decision tree with 10-fold cross validation Random forest: 2018 Urcuqui et al., 2018
Accuracy = 82%
False negative rate = 23%
False positive rate = 12%
Classification of PD from HC Diagnosis PhysioNet 47; 18 HC + 29 PD SVM, KNN, random forest, decision tree SVM with cubic kernel: 2017 Alam et al., 2017
Accuracy = 93.6%
Sensitivity = 93.1%
Specificity = 94.1%
Classification of PD from HC Diagnosis PhysioNet 34; 17 HC + 17 PD MLP, SVM, decision tree MLP: 2018 Alaskar and Hussain, 2018
Accuracy = 91.18%
Sensitivity = 1
Specificity = 0.83
Error = 0.09
AUC = 0.92
Classification of PD from HC and assess the severity of PD Diagnosis PhysioNet 166; 73 HC + 93 PD 1D-CNN, 2D-CNN, LSTM, decision tree, logistic regression, SVM, MLP 2D-CNN and LSTM accuracy = 96.0% 2019 Alharthi and Ozanyan, 2019
Classification of PD from HC Diagnosis PhysioNet 146; 60 HC + 86 PD SVM-Gaussian with 3- or 5-fold cross validation Accuracy = 100%, 88.88%, and 100% in three test groups 2019 Andrei et al., 2019
Classification of PD from HC Diagnosis PhysioNet 166; 73 HC + 93 PD ANN, SVM, naïve Bayes with cross validation ANN accuracy = 86.75% 2017 Baby et al., 2017
Classification of PD from HC Diagnosis PhysioNet 31; 16 HC + 15 PD SVM-linear, KNN, naïve Bayes, LDA, decision tree with leave-one-out cross validation SVM, KNN and decision tree accuracy = 96.8% 2019 Félix et al., 2019
Classification of PD from HC Diagnosis PhysioNet 31; 16 HC + 15 PD SVM-linear with leave-one-out cross validation Accuracy = 100% 2017 Joshi et al., 2017
Classification of PD from HC Diagnosis PhysioNet 165; 72 HC + 93 PD KNN, CART, decision tree, random forest, naïve Bayes, SVM-polynomial, SVM-linear, K-means, GMM with leave-one-out cross validation SVM:
Accuracy = 90.32%
Precision = 90.55%
Recall = 90.21%
F-measure = 90.38%		 2019 Khoury et al., 2019
Classification of ALS, HD, PD from HC Diagnosis PhysioNet 64; 16 HC + 15 PD + 13 ALS + 20 HD String grammar unsupervised possibilistic fuzzy C-medians with FKNN, with 4-fold cross validation PD vs. HC accuracy = 96.43% 2018 Klomsae et al., 2018
Classification of PD from HC Diagnosis PhysioNet 166; 73 HC + 93 PD Logistic regression, decision trees, random forest, SVM-Linear, SVM-RBF, SVM-Poly, KNN with cross validation KNN: 2018 Mittra and Rustagi, 2018
Accuracy = 93.08%
Precision = 89.58%
Recall = 84.31%
F1-score = 86.86%
Classification of PD from HC Diagnosis PhysioNet 85; 43 HC + 42 PD LS-SVM with leave-one-out, 2- or 10-fold cross validation Leave-one-out cross validation: 2018 Pham, 2018
AUC = 1
Sensitivity = 100%
Specificity = 100%
Accuracy = 100%
10-fold cross validation:
AUC = 0.89
Sensitivity = 85.00%
Specificity = 73.21%
Accuracy = 79.31%
Classification of PD from HC Diagnosis PhysioNet 165; 72 HC + 93 PD LS-SVM with leave-one-out, 2- or 5- or 10-fold cross validation Accuracy = 100% 2018 Pham and Yan, 2018
Sensitivity = 100%
Specificity = 100%
AUC = 1
Classification of PD from HC Diagnosis PhysioNet 166; 73 HC + 93 PD DCALSTM with stratified 5-fold cross validation Sensitivity = 99.10% 2019 Xia et al., 2020
Specificity = 99.01%
Accuracy = 99.07%
Classification of HC, PD, ALS and HD Diagnosis and differential diagnosis PhysioNet 64; 16 HC + 15 PD + 13 ALS + 20 HD SVM-RBF with 10-fold cross validation PD vs. HC: 2009 Yang et al., 2009
Accuracy = 86.43%
AUC = 0.92
Classification of PD, HD, ALS and ND from HC Diagnosis PhysioNet 64; 16 HC + 15 PD + 13 ALS + 20 HD Adaptive neuro-fuzzy inference system with leave-one-out cross validation PD vs. HC: 2018 Ye et al., 2018
Accuracy = 90.32%
Sensitivity = 86.67%
Specificity = 93.75%
Classification of PD from HC and assess the severity of PD Diagnosis mPower database 50; 22 HC + 28 PD Random forest, bagged trees, SVM, KNN with 10-fold cross validation Random forest: 2017 Abujrida et al., 2017
PD vs. HC accuracy = 87.03%
PD severity assessment accuracy = 85.8%
Classification of PD from HC Diagnosis mPower database 1,815; 866 HC + 949 PD CNN with 10-fold cross validation Accuracy = 62.1% 2018 Prince and de Vos, 2018
F1 score = 63.4%
AUC = 63.5%
Classification of PD from HC Diagnosis Dataset from Fernandez et al., 2013 49; 26 HC + 23 PD KFD-RBF, naïve Bayes, KNN, SVM-RBF, random forest with 10-fold cross validation Random forest accuracy = 92.6% 2015 Wahid et al., 2015

ALS, amyotrophic lateral sclerosis; ANN, artificial neural network; AUC, area under the receiver operating characteristic (ROC) curve; AVS, acute unilateral vestibulopathy; BagDT, bootstrap aggregation for a random forest of decision trees; CA, anterior lobe cerebella atrophy; CART, classification and regression trees; DCALSTM, dual-modal with each branch has a convolutional network followed by an attention-enhanced bi-directional LSTM; DN, downbeat nystagmus syndrome; ESN, echo state network; FKNN, fuzzy k-nearest neighbor; GMM, Gaussian mixture model; HC, healthy control; HD, Huntington's disease; IH, idiopathic hyposmia; KFD, kernel Fisher discriminant; KNN, k-nearest neighbors; LDA, linear discriminant analysis; LOSO, leave-one-subject-out; LS-SVM, least-squares support vector machine; LSTM, long short-term memory; MCI, mild cognitive impairment; MIL, multiple-instance learning; MLP, multilayer perceptron; MTFL, multi-task feature learning; NN, neural network; NSVC, nu-support vector classification; OT, primary orthostatic tremor; PD, Parkinson's disease; PDMCI, PD participants who met criteria for mild cognitive impairment; PDNOMCI, PD participants with no indication of mild cognitive impairment; PNN, probabilistic neural network; PNP, sensory polyneuropathy; PPV, phobic postural vertigo; QDA, quadratic discriminant analysis; RkF, repeated k-fold; SVM, support vector machine; SVM-Poly, support vector machine with polynomial kernel; SVM-RBF, support vector machine with radial basis function kernel.