Automated Parkinson's Disease Detection: A Review of Techniques, Datasets, Modalities, and Open Challenges

Zadoo, Sheerin; Singh, Yashwant; Singh, Pradeep Kumar

doi:10.2478/ijssis-2024-0008

Figures & Tables

Parkinson's disease symptoms [4, 33, 34].

Comparison of the highest accuracies achieved in various modalities in PsD detection.

Process flow diagram of PsD detection using EI.

Taxonomy for detection of Parkinson's disease according to different modalities [42, 48].

Comparative analysis of techniques used to detect PsD using MRI images_

Author (s)	Year	Purpose	Algorithm	Dataset	Output	Merits	Demerits
Sangeetha et al. [62]	2023	PsD Diagnosis	CNN	PPMI dataset PsD: 229 HC: 229	Accuracy = 96%	Preprocessing done and various performance metrics considered	-
Cui et al. [61]	2022	Classification of PsD and HC	Resnet18 + Support Vector Machine	PPMI dataset	Accuracy = 98.66%	Feature fusion improves classification performance	-
Monte-Rubio et al. [63]	2022	Parameters from multiple sites to harmonize MRI clinical data for classification of PsD	Weighted HARMonization Parameters (WHARMPA)	PPMI dataset PsD: 216 HC: 87	Balanced accuracy = 78.60%; AUC = 0.90	WHARMPA encodes global site-effects quantitatively with simple implementation	Unbalanced dataset used,
Hathaliya et al. [64]	2021	To classify PsD and HC patients	Stacked ML model	PPMI dataset	Accuracy = 92.5%, Precision = 98%, F1_score = 98%, Recall = 97%	Detects PsD and measures the disease progression in PsD patients	-
Vyas et al. [5]	2021	Detection of PsD using 3D CNN and compare with the 2D CNN	3D and 2D CNN	PPMI database	Accuracy = 88.9%	Voxel-based morphometry usage is very accurate	Manual feature extraction
Sivaranjini and Sujatha [57]	2021	Detection of PsD	Transfer learning (TL), AlexNet pre-trained model	PPMI MRI images	Accuracy = 88.9%	Improved efficiency of the model using transfer learning	Only one slice of brain image is used
Mangesius et al. [56]	2020	Classification of PsD, MSA and PSP	Decision Tree algorithm	-	Accuracy of classification PsD = 83.7%	Distinguishes PsD, MSA, and PSP with very high accuracy	Does not classify HC and PsD
Chakraborty et al. [60]	2020	Detecting PsD	3D CNN model	MRI scans of 406 subjects from PPMI database	Accuracy = 95.29%, Average recall = 0.943, Precision = 0.927, F1-score = 0.936	The maximum focus was on the substantia nigra region to predict PsD as it is the most-affected part of brain	Specific subcortical structures should be considered

Comparative analysis of PsD public datasets_

Dataset Name	References	Instances and subjects	Attributes	Datatype	Merit	Demerit	Used Device /Sensor
Parkinson's Disease dataset	[132, 133, 134]	I-197, PsD:23, HC:8	23	Speech signal	No missing values	Fewer instances, unbalanced and only suitable for binary classification	Recorder
PhysioNet/Vertical Ground Reaction (VGRF)	[55, 135, 136, 137, 138]	PsD:93, HC:72	-	Multi-channel recordings from force sensors	Disease severity is also monitored	-	8 sensors under each foot
PPMI database	[5, 108, 139, 140]	PsD:600, HC:400	-	MRI Images, DaTscan, PET, etc.	Most widely used database of PsD datasets	Imbalanced dataset	Tesla scanner
Parkinson's telemonitoring/Oxford PsD Telemonitoring Dataset	[141, 142]	I-5875, PsD: 42	26	Voice recordings	No missing values	Suitable only for binary classification	Telemonitoring device
Daphnet Freezing of Gait Data Set	[134, 143, 144]	I-237, PsD:10	9	Gait monitoring data	Realistic activity monitoring	Very few subjects in dataset	Acceleration sensors at the hip and leg
Parkinson Dataset with replicated acoustic features dataset	[141, 145, 146]	I-240, PsD:40, HC:40	46	Voice recording replications	No missing values	ML cannot be applied	Recorder
Parkinson Speech Dataset with Multiple Types of Sound Recordings Dataset	[147, 148]	I-1040, PsD:20, HC:20	26	Sound recordings	Useful for classification as well as regression	Fewer subjects	-
Parkinson's Disease Classification Data Set	[149, 150, 151, 152]	I-756, PsD:188, HC:64	754	Speech recordings	Features extracted through many speech signal processing algorithms	Imbalanced dataset	Microphone
Parkinson Disease Spiral Drawings Using Digitized Graphics Tablet dataset	[153, 154]	I:77, PsD:62, HC:15	7	Handwriting-Spiral drawings	Three different kinds of tests to collect data	Fewer subjects, and imbalanced dataset	Wacom Cintiq 12WX graphics tablet
NewHandPsD dataset	[155, 156]	PsD:31, HC:35	-	Handwriting	Balanced dataset	Fewer subjects	Smart pen (BiSP)
PC-GITA Spanish dataset	[157, 158]	PsD:50, HC:50	-	Voice samples	A well-balanced corpus with respect to age and gender	Recorded under optimal recording conditions	Fast Track C400 sound card and a Professional microphone
NTUA Parkinson Dataset	[158, 159]	I:44007, PsD:55, HC:23	-	MRI and DaTscans	Sufficiently large dataset	Imbalanced dataset	-
Neurovoz	[160, 161, 162]	PsD:47, HC:32	-	Spanish speech samples	No Noise	Fewer subjects	AKG C420 headset microphone coupled with a preamplifier
OpenfMRI	[163, 164]	37 studies with 1411 subjects	-	MRI and EEG scans	NIFTI format	-	-
PsD-BioStampRC21 dataset	[165, 166]	PsD:17, HC:17	-	wearable sensor accelerometry data	24-hour monitoring data	Fewer subjects	MC 10 BioStamp RC sensors
Italian Parkinson's Voice and Speech	[167, 168]	PsD: 28, HC: 37	-	Voice and Speech data	-	Fewer subjects	-
Physio Bank	[94, 169, 170]	PsD: 93, HC:73	-	Database of Gait	Multichannel recordings, demographic information, and disease severity	-	Ultraflex Computer Dyno Graphy, Infotronic Inc. sensors
Neurocon	[171, 172]	I: 303, PsD: 27, HC:16	-	T1 and resting-state MRI dataset	Includes both T1 and resting-state scans	Fewer subjects of only initial stage PsD	1.5-Tesla Siemens Avanto MRI scanner

A relative contrast of the proposed survey with the PsD detection state-of-the-art surveys_

Author	Year	Conclusion	1	2	3	4	5	6	7	8	9	10	11	12	13
Proposed Survey	2023	Examines PsD patients’ datasets and PsD diagnosis techniques on various modalities while addressing the open research challenges too	✓	✓	9	✓	✓	✓	✓	✓	✓	✓	✓	✓	✓
Salari et al. [33]	2023	SVM is proposed to be a bridge between psychological research and the clinical facts	✓	✓	5			✓				✓			✓
Shafiq et al. [39]	2023	Flower Pollination Algorithm and Extreme Gradient Boost Algorithm pair obtained the accuracy of 93%	✓	✓	1			✓				✓	✓		✓
Shaban [40]	2023	ANN is the most widely used PsD detection technique and sensory signals, EEG and handwriting are the most utilized data modalities			4			✓	✓			✓			✓
Kumar et al. [41]	2023	Accelerometer is the most widely used sensor for gait analysis for PsD detection	✓	✓	1		✓	✓				✓	✓	✓	✓
Dixit et al. [42]	2023	Review of PsD diagnosis along with an extensive discussion of future scope	✓		6	✓	✓	✓	✓			✓			✓
Khanna et al. [43]	2023	No existing technique is yet suitable for practical use clinically. MRI, PET and SPECT scans are the most popular neuroimaging modalities	✓	✓	3		✓	✓				✓	✓	✓	✓
Zhang [44]	2022	SVM and ANN show best accuracy on SPECT scan for PsD detection	✓		3			✓				✓			✓
Chandrabhatla et al. [45]	2022	Technological evolution in detecting PsD, in-lab to in-home	✓		3			✓				✓
Giannakopoulou and Roussaki [46]	2022	ML + IoT can revolutionize PsD diagnosis, mostly used best performing AI models are MLPs, LSTMs, CNNs, other ANN and ensemble learning techniques	✓		5			✓	✓		✓	✓	✓	✓	✓
Rana et al. [47]	2022	AI has revolutionized detecting early-stage PsD	✓		7			✓	✓			✓			✓
Tanveer et Al. [48]	2022	CNN and RNN give better performance than ML	✓		8			✓	✓			✓	✓	✓	✓
A. ul Haq et al. [49]	2022	DL techniques are the most appropriate			4		✓		✓			✓	✓	✓	✓
Rana et al. [50]	2022	Best accuracy to diagnose PsD for voice data was achieved using L1-norm SVM and k-fold cross-validation, for handwritten patterns using bagging ensemble, and for gait analysis by SVM	✓	✓	3			✓				✓	✓	✓	✓
Alzubaidi et al. [51]	2021	Best performing and widely used DL models and datasets according to various modalities are identified			9	✓	✓		✓			✓	✓	✓	✓
Loh et al. [52]	2021	CNN model achieves higher accuracy in various modalities for PsD detection	✓		8				✓			✓			✓
Noor et al. [53]	2020	CNN is used mostly in the detection of PsD			1		✓		✓			✓			✓
Khachnaoui et al. [54]	2020	DL methods show better performance than ML			2			✓	✓		✓	✓
Di Biase et al. [55]	2020	Very few algorithms are accurate enough to be potentially used clinically for PsD diagnosis and monitoring symptoms			1			✓				✓	✓	✓	✓

Comparative analysis of PsD private datasets_

Dataset/Source/Author Name	References	Instances and subjects	Attributes	Datatype	Merit	Demerit	Used Device/Sensor
Hospital Universiti Kebangsaan Malaysia, Kuala Lumpur	[124]	I: 1440 PsD: 20 HC: 20	-	EEG signals	Non-invasive detection	Fewer subjects	14-channel wireless Emotiv Epoc headset
Wearable Bio mechatronics Laboratory at Western University	[173]	I: 409,380 PsD: 13	-	Video recordings while performing the trials	Resting, postural, and action tremors are monitored	-	Wearable assistive devices
Tuncer T.	[174]	I:756 PsD:188 HC:64	-	Vowel voice recordings	High classification performance even with heterogeneous dataset, less expensive	Imbalanced dataset	Microphone
M. Lu	[114]	I:200 PsD:55	-	Video recordings of MDS-UPDRS exams	Both gait and finger tapping experiments are performed	Less data	-
Pacific Parkinsons Research Centre (PPRC)	[175]	PsD:20 HC:21	-	Resting EEG	High Performance metrics	Fewer subjects	64-channel EEG cap and a Neuroscan SynAmps2 acquisition system
Iakovakis et al.	[176, 177]	PsD:18 HC:15	913	Keystroke dynamics information	Combined discriminative potential of enriched keystroke variables	Fewer subjects	Touchscreen smartphone
Hospital at Sun Yat-sen University	[178]	I:7000 PsD:40 HC:30	-	EEG signals	Large number of instances	Fewer hyperparameters	64-electrode Geodesic Sensor Net (Electrical Geodesics Inc.)
RMIT University, Melbourne, Australia	[179]	PsD:41 HC:40	-	Voice samples	all recordings were normalized	Fewer number of samples	Apple iPhone 6S plus

Comparative analysis of PsD detection using emotional intelligence_

Author (s)	Year	Purpose	Algorithm	Dataset	Output	Merits	Demerits
Pegolo et al. [127]	2022	Measuring hypomimia to distinguish between PsD and HC and to classify the emotions	Random Forest showed best results	Frontal face videos; PsD:50 HC:17	AUC values from 94.3 to 91.6; F1 scores from 76.2 to 71.5	A stand-alone methodology for quantifying the degree of impairment	Imbalanced data
Parameshwara et.al. [124]	2022	PsD detection	KNN and Naïve Bayes Classifiers	EEG data; PsD: 20 HC: 20	Acc = 99%	Non-invasive use of emotional EEG signals	Inadequate number of PsD patients with mild-to-moderate severity, relation in EEG and motor symptoms is unknown
Dar et al. [20]	2022	EEG based emotional classification	ELM (extreme learning machine)	PsD dataset, AMIGOS, and SEED-IV datasets	Acc = 83.2%	Implemented on cross-subject and cross-dataset, robust to real-time applications	Datasets with fewer number of subjects
Anusri et al. [126]	2021	Early detection of PsD using facial emotional intelligence	Alexnet and VGG16	PPMI dataset PsD:188 HC:50	Acc = 96.5%	Early diagnosis justified using different performance metrics	Imbalanced dataset
Sechidis et al. [128]	2021	Assessment of PsD speech characteristics	Machine learning	Speech dataset	NA	Main focus is on happy and sad emotion only	No dataset of PsD and HC exists thta has labeled emotions
Justyna and Burget [22]	2020	Analysis of Emotional changes during quick pronunciation exercise	XG Boost	Video recording of speech exercise PsD:70 HC:45	Acc = 69%; sadness and surprise are negatively correlated with PsD	Seven emotion classes are considered	Low accuracy
Yuvaraj et al. [125]	2014	To discriminate PsD patients and HC while emotional information processing	Discriminant analysis	EEG scan; PsD: 20 HC: 30	Acc = 95%	Distributed spectral powers in EEG frequency bands supply information about emotion processing in PsD patients	Fewer number of subjects
Zhao et al. [129]	2014	Classification of emotions through speech in PsD patients	Naïve Bayes, Random Forest, and SVM	Recorded speaking short statements	Acc = 3.33%	First automatic classification of emotions in the voice of PsD patients	Low Accuracy

List of acronyms_

Acronym	Meaning	Acronym	Meaning
PsD	Parkinson's Disease	FER	Facial emotion recognition
HC	Healthy Controls	MRI	Magnetic Resonance Imaging
PPMI	Parkinson's Progression Markers Initiative	TNR	True negative rate
TPR	True positive rate	SVM	Support Vector Machine
PET	Positron emission tomography	MSA	Multiple system atrophy
AI	Artificial Intelligence	DL	Deep Learning
ML	Machine Learning	DICOM	Digital Imaging and Communications in Medicine
SPECT	Single-photon emission computerized tomography	EMG	Electromyogram
CNN	Convolutional Neural Network	CRNN	Convolutional Recurrent Neural Network
SWEDD	Scans Without Evidence of Dopaminergic Deficit	KNN	K-nearest neighbors
ECG	Electrocardiogram	PSP	Progressive supranuclear palsy
LC	Locus coeruleus	LDA	Linear discriminant analysis
SNpc	Substantia nigra pars compacta	DaTscan	Dopamine Transporter Scan
MFCC	Mel-Frequency Cepstral Coefficients	OCSA	Optimized crow search algorithm
OCT	Optical coherence tomography	CCSA	Chaotic crow search algorithm
EEG	Electroencephalogram	RBD	REM sleep behavior disorder
PCA	Principal component analysis	MDS	Movement Disorder Society
ROI	Region of interest	RGNN	Regularized graph neural network
ANOVA	Analysis of variance	FAWT	Flexible analytic wavelet transform
RF	Random forests	Emotion DL	Emotion-aware distribution learning
RBF	Radial basis functions	LIWC	Linguistic Inquiry and Word Count
RQA	Recurrence quantification analysis	MVLL	Maximum vertical line length
MDLL	Maximum diagonal line length	FMI	Face mobility index
STRNN	Spatial–temporal recurrent neural network	LDA	Latent Dirichlet Allocation
FACS	Facial Action Coding System	VGRF	Vertical ground reaction force
AUC	Area under the ROC Curve	EI	Emotional intelligence
MEIS	Multi-factor emotional intelligence scale	FNN	Fuzzy neural network
ANN	Artificial Neural network	GCNN	Graph Convolutional Neural Network
FL	Federated learning	UPDRS	Unified Parkinson's disease rating scale
FDG	Fluorodeoxyglucose	MCI	Mild cognitive impairment
DN	Dopaminergic Neurons	DT	Dopamine transporter
DBN	Deep belief network	LSTM	Long short-term model
TL	Transfer learning	SVR	Support vector regression
Node DAT	Node-wise domain adversarial training	GSR	Galvanic skin response
DNA	Deoxyribonucleic acid	RBC	Red blood cell
WHARMPA	Weighted HARMonization Parameters	RNA	Ribonucleic acid
FoG	Freezing of Gait	CC	Correlation coefficients
MCC	Merkel cell carcinoma	RMSE	Root mean square error

Comparative analysis of techniques used to detect PsD using SPECT images_

Author (s)	Year	Purpose	Algorithm	Dataset	Output	Merits	Demerits
Hathaliya et al. [73]	2022	PsD classification and monitoring of the DT (dopamine transporter) level inside the brain	CNN-based scheme	58692, 11738, and 7826 SPECT images from PPMI for training, validation, and testing, respectively	Accuracy = 88%	Accurate diagnosis of PsD and its progression	-
Adams et al. [32]	2021	DL algorithm to accurately predict motor-based symptoms	CNN model	252 subjects DAT SPECT images	Average absolute error = 6.0 ± 4.8	Enhanced diagnosis of UPDRS_III score with longitudinal data	Features related to PsD are not identified
Magesh et al. [74]	2020	ML-based early detection of PsD	CNN and VGG16-based TL scheme	642 DaTscan SPECT images	Accuracy = 95.2%	Quick diagnosis for PsD	Dearth of conclusive diagnostic tests for PsD
Pereira and Ferreira [75]	2020	To classify PsD, SWEDD, and HC	CNN models	SPECT and MRI images	Accuracy = 65.7%	Classifies PsD, SWEDD, and HC subjects	Lower accuracy
Wenzel et al. [76]	2019	Robust classification algorithm to diagnose the dopamine transporter	CNN and ImageNet-based TL, semi-quantitative SBR analysis	645 subjects	Accuracy = 97%	Accurate diagnosis of PsD patient with DT analysis	Trained model with few samples
Ortiz et al. [77]	2019	To classify HC and PsD using isosurface-based feature extraction	CNN model	269 DaTscan images	Accuracy = 95%	Low complexity of the input data	Increases overall system complexity

Comparative analysis of techniques used to detect Parkinson's disease progression detection_

Author (s)	Year	Purpose	Algorithm	Dataset	Output	Merits	Demerits
Kleinholdermann et al. [112]	2021	To predict UPDRS scores	Random Forest regression model outnumbered	Surface electromyography (sEMG) electrodes data; PsD: 45	Correlation between true and estimated UPDRS values = 0.739	Support the usage of wearables to detect PsD	Other performance metrics not considered
Hssayeni et al. [113]	2021	Severity level estimation (UPDRS-III scores)	Dual-channel LSTM and TL Ensemble	Angular velocity data from inertial sensors; PsD:24	1D-CNN-LSTM used for raw signals and 2D-CNN-LSTM used for time-frequency data with r = 0.79, MAE = 5.95	Unobtrusive analysis at home	Fewer subjects
Lu et al. [114]	2021	Detection and severity estimation for PsD patients	OF-DDNet	Video recordings of MDS-UPDRS exams; PsD: 55	AUC = 0.69, F1-score = 0.47, prec. = 0.47, and balanced acc. = 48%	Handles model uncertainty, label noise and small dataset and imbalance issues	-
Raza et al. [110]	2021	Progression of PsD	XGBoost	Speech recordings; I:5875; PsD:42	MAE = 5.09 ± 0.16	Enables continuous monitoring and efficient communication	Uses only voice recordings
Rupprechter et al. [89]	2021	Severity estimation (UPDRS scores)	Ordinal Random Forest classifier	729 videos of gait assessments	Balanced accuracy = 50%, Binary sensitivity = 73%; Binary specificity = 68%	Provides interpretability	3D pose estimation not used
Abujrida et al. [115]	2020	Detection and severity estimation in walking balance	Wavelet (DWT) domain + fine tree/Boosted trees/RF/bagged trees/weighted LR/KNN/LDA, cubic SVM	Signals from gyroscopes, accelerometers, demographics, and lifestyle data PsD:152, HC:304	Bagged trees Acc = 95%, Prec = 95% AUROC = 0.92 RF: Acc = 93%, Prec = 92% AUROC = 0.97;	Large no. of participants, and amalgam of frequency, time, and statistical features with sway area and lifestyle parameters too	Lack of signal segmentation strategies
Mehrbakhsh Nilashi et al. [111]	2020	Remote tracking of PsD progression	DBN with SVR	Parkinson's telemonitoring dataset; I: 5875	Average output for testing and training sets are 0.521 and 0.537 respectively for RMSE in detecting motor-UPDRS	Supervised and unsupervised learning techniques used	Feature selection techniques are not used
Li et. al. [116]	2019	Diagnosis and severity (H&Y scores) estimation	Random Forest regression	Acceleration and angular velocity signals; PsD:13 HC:12	MAE = 0.166, r = 0.97	Records user's motion information unobtrusively	-
Buongiorno et al. [117]	2019	Detection and severity rating of PsD	ANN and SVM	Postural and kinematics parameters from MS Kinect v2 sensor; PsD:16 HC:14	ANN: Acc. = 95%, SVM: Acc. = 81%, Spec = 78%	Low-cost and easy and fast setup	Only mild vs. moderate severity is measured
Yoon et al. [118]	2019	PsD Severity estimation (UPDRS 0–176 scores)	TL	Phonation and speech recordings; PsD:42	MAE = 2–3 Approximately	Recording done at home, and better accuracy	-
Grammatikopoulou et al. [119]	2019	PsD Severity estimation	Parallel two LSTMs trained with raw joint-coordinates and the combined line distances	Skeletal features of MS Kinect v2 RGB videos; PsD (advanced stage): 12 and PsD (initial stage):6	Acc. = 77.7%	-	Very few subjects, only 2 stages

Automated Parkinson's Disease Detection: A Review of Techniques, Datasets, Modalities, and Open Challenges

Figures & Tables

Figure 1:

Figure 2:

Figure 3:

Figure 4:

Figure 5:

Figure 6:

Comparative analysis of techniques used to detect PsD using MRI images_

Comparative analysis of PsD public datasets_

A relative contrast of the proposed survey with the PsD detection state-of-the-art surveys_

Comparative analysis of PsD private datasets_

Comparative analysis of PsD detection using emotional intelligence_

List of acronyms_

Comparative analysis of techniques used to detect PsD using SPECT images_

Comparative analysis of techniques used to detect Parkinson's disease progression detection_

Paradigm

My account