Table A1.
Studies based on inertial sensors.
| Study | Problem | Dataset Population | Input Data | Analysis/Algorithms | Evaluation |
|---|---|---|---|---|---|
| Rastegari et al. (2020) [49] | Diagnosis | 10 PD patients and 10 healthy controls from another study [166] | Accelerometer and gyroscope data from both ankles using SHIMMER sensors | Segmentation + bag-of-words features extraction based on sub-sequences clustering with k-medoid + SVM/DT/RF/kNN | Best performing: DT with ACC = 90%, PREC = 90%, REC = 90% |
| Zhang et al. (2020) [54] | Diagnosis | 656 PD patients and 2148 healthy controls from the mPower study [126] | Gait features from smartphone sensors located at the pocket | Deep CNN | AUROC = 0.87 |
| Juutinen et al. (2021) [50] | Diagnosis | 29 PD patients and 29 healthy controls | Accelerometer and gyroscope signals from a waist-mounted smartphone while performing walking tests | Interpolation + low-pass filtering + smoothening + segmentation to strides + feature extraction + feature selection with mRMR/SFS/SBS + classification of strides with DT/Gaussian kernel/LDA/ensemble/kNN/LR/NB/SVM/RF + majority voting for subject classification | Best performing: SFS + kNN with ACC = 84.5%, SENS = 88.5%, SPEC = 81.3% |
| Fernandes et al. (2018) [51] | Parkinsonism diagnosis and differential diagnosis between idiopathic PD (IPD) and vascular parkinsonism (VaP) | 15 IPD, 15 VaP patients and 15 healthy controls | Gait signals from wearable Physilog motion sensors placed on both feet and MOCA scores | Normalization + feature selection based on Kruskal–Wallis and Mann–Whitney tests with Bonferroni correction + features ranking with RFs + MLP/DBN | Best performing for parkinsonism detection: MLP with ACC = 99.33%, SENS = 93.33%; for IPD-VaP differential diagnosis: DBN with ACC = 73.33%, SENS = 73.33%, SPEC = 73.33% |
| Cuzzolin et al. (2017) [52] | Diagnosis and severity estimation in the SF-36 (0–100) scale | 156 PD patients and 424 healthy controls | Signals from an IMU attached to the lower spine while walking | HMM + kNN | For diagnosis: ACC = 85.51%, F1-SCORE = 81.54%; for severity estimation: MAE = 27.81 |
| Borzì et al. (2020) [70] | Severity estimation (MDS-UPDRS related to leg agility scores classification) | 93 PD patients | Acceleration, angular velocity and orientation data from a smartphone application | Low-pass Butterworth filter + FFT + feature selection + DT/kNN/SVM/ANN | Best performing: ANN with ACC = 77.7%, AUROC = 0.92, r = 0.92, RMSE = 0.42, ICC = 0.82 |
| Hssayeni et al. (2018) [96] | Fluctuation identification (on/off states detection) | 12 PD patients (1st dataset) and 7 PD patients (2nd dataset) | Signals from an ankle-mounted triaxial gyroscope | FIR bandpass filter + LSTM | ACC = 73–77%, SENS = 63–75%, SPEC = 78–83% |
| Aich et al. (2020) [97] | Fluctuation identification (on/off states detection) | 20 PD patients | Statistical and gait parameter features from 2 knee-worn accelerometers | Feature selection + RF/SVM/kNN/NB | Best performing: RF with ACC = 96.72%, PREC = 96.92%, REC = 97.35%, F1-SCORE = 0.97, AUROC = 0.99 |
| Abujrida et al. (2020) [53] | Diagnosis and severity estimation regarding walking balance (MDS-UPDRS-2.12), shaking/tremor (MDS-UPDRS-2.10) and FoG | 152 PD patients and 304 healthy controls from the mPower study [126] | Signals from smartphone embedded accelerometers and gyroscopes, demographics and lifestyle data | Segmentation + smoothening + feature extraction in time, frequency (FFT + PSD), wavelet (DWT) domain + RF/bagged trees/boosted tress/fine tree/cubic SVM/weighted kNN/LR/LDA | Best performing for diagnosis: RF with ACC = 95%, PREC = 94%, AUROC = 0.99; for walking balance: RF with ACC = 93%, PREC = 92%, AUROC = 0.97; for tremor: bagged trees with ACC = 95%, PREC = 95%, AUROC = 0.92; for FoG: bagged trees with ACC = 98%, PREC = 96%, AUROC = 0.98 |
| Kim et al. (2015) [84] | FoG detection | 15 PD patients | Gyroscopic and accelerometer data from smartphones placed at the waist, pocket and ankle | AdaBoost | SENS = 81.1–86%, SPEC = 91.5–92.5% |
| Ashour et al. (2020) [87] | FoG detection | 10 PD patients | Accelerometer data from sensors placed on the ankles, knees and hips | Patient-dependent LSTM/{DWT/FFT + Patient-dependent SVM/ANN} | Best performing: LSTM with ACC = 68.44–98.89% |
| Torvi et al. (2016) [89] | FoG prediction in 1 s, 3 s and 5 s horizons | 10 PD patients from Daphnet dataset | Accelerometer data from sensors placed on the ankles, the thighs and the trunk | LSTM/RNN with or without transfer learning | Best performing: LSTM + transfer learning with ACC = 85–95% |
| Arami et al. (2019) [90] | FoG prediction (2-class classification FoG/no-FoG and 3-class classification pre-FoG/FoG/no-FoG) | 10 PD patients from Daphnet dataset | Accelerometer data from sensors placed on the ankles, the thighs and the trunk | Windowing + filtering + feature extraction + feature selection with mRMR/BE + features time series prediction with AR/ARMA + RBF-SVM/PNN | For 2-class classification: ACC = 94%, SENS = 93%, SPEC = 87%; for 3-class classification: ACC = 77% |
| Kleanthous et al. (2020) [91] | FoG detection (FoG/walking/transition from walking to FoG classification) | 10 PD patients from Daphnet dataset | Accelerometer data from sensors placed on the ankles, the thighs and the trunk | Low pass Butterworth filter + Boruta algorithm + GBM + XGB for feature selection + XGB/RF/GBM/RBF-SVM/MLP | Best performing: RBF-SVM with ACC = 79.85%, SENS = 72.34–91.49%, SPEC = 87.36–93.62% |
| Li et al. (2020) [88] | FoG detection | 10 PD patients from Daphnet dataset | Accelerometer data from sensors placed on the ankles, the thighs and the trunk | Filtering + segmentation + data augmentation + CNN with squeeze-and-excitation blocks for feature extraction + attention-enhanced LSTM | ACC = 98.1–99.7%, SENS = 95.1–99.1%, SPEC = 98.8–99.8% for generalized and personalized models with 10-fold cross-validation; AUC = 0.945, ACC = 91.9%, EER = 10.6% with LOSO validation |
| Halder et al. (2021) [92] | FoG states classification (pre-FoG, FoG, pre-post-FoG, no-FoG) | 10 PD patients from Daphnet dataset | Accelerometer data from sensors placed on the ankles, the thighs and the trunk | Second-order Butterworth low-pass filtering + PCA + kNN/MLP/SVM | Best performing: kNN with ACC = 98.92%, SENS = 94.97%, SPEC = 99.19%, F1-SCORE = 95.25%, PREC = 95.55% |
| Palmerini et al. (2017) [93] | Pre-FoG detection (classification between gait and pre-FoG) | 11 PD patients | Accelerometer and gyroscope signals from sensors placed at the lower-back and at the ankles | Windowing + feature extraction + LDA | SENS = 83%, SPEC = 67%, AUROC = 0.76 |
| Borzì et al. (2021) [94] | FoG and pre-FoG detection | 11 PD patients on and off therapy | A single angular velocity signal from 2 shins-mounted inertial sensors while performing TUG tests | Normalization + segmentation + wrapper feature selection + SVM/kNN/LDA/LR | Best performing: SVM with ACC = 85.5–86.1%, SENS = 84.1–85.5%, SPEC = 85.9–86.3%, F-SCORE = 73.4–74.6% |
| Shi et al. (2020) [85] | FoG detection | 63 PD patients | Accelerometer, gyroscope and magnetometer signals from IMUs placed on both ankles and the spine while performing TUG tests | Time-series segmentation with overlapping windows + Morlet CWT/FFT/raw data + 1D-CNN/2D-CNN/LSTM | Best performing: CWT + 2D-CNN with ACC = 89.2%, SENS = 82.1%, SPEC = 96%, GM = 88.8% |
| Camps et al. (2018) [86] | FoG detection | 21 PD patients | Signals from a waist-mounted IMU with accelerometer, gyroscope and magnetometer while performing several walking tests and ADLs | Spectral window stacking (with FFT) + RUSBoost/SVM-RBF/CNN | Best performing: CNN with ACC = 89%, SENS = 91.9%, SPEC = 89.5% and geometrical mean of SENS-SPEC = 90.6% |
| Ghassemi et al. (2018) [95] | Gait segmentation (strides detection) | 10 PD patients for the 1st experiment and 34 PD patients for the 2nd experiment | Acceleration and angular velocity signals from foot-worn IMUs while walking in straight line (1st experiment) and walking, turning or performing other movements (2nd experiment) | Peak detection algorithm/Euclidean DTW/Probabilistic DTW/hierarchical HMM | Best performing for the 1st experiment: all with F-SCORE = 99.8–100%; For the 2nd experiment: hHMM with PREC = 98.5%, REC = 93.5%, F-SCORE = 95.9% |
| Kostikis et al. (2015) [55] | Diagnosis | 25 PD patients and 20 healthy controls | Tremor measurements from accelerometer and gyroscope smartphone sensors | NB/LR/SVM/AdaBoost/C4.5/RF | Best performing: RF with SENS = 82%, SPEC = 90%, AUROC = 0.94 |
| Williamson et al. (2021) [59] | Diagnosis | 202 PD patients and 178 healthy controls from the UK Biobank dataset (https://www.ukbiobank.ac.uk/, accessed on 17 February 2022) | Acceleration signals from a wrist-worn sensor | Segmentation + automatic segments labelling (gait or low movement) + segmentation into frames + features extraction + GMM | SENS = 65–75%, AUROC = 0.85 |
| Park et al. (2021) [60] | Diagnosis | 25 PD patients and 21 healthy controls | Signals from IMUs attached to the thumb and index fingers while performing finger tapping, hand movements and rapid altering movements | Linear regression + correlation between motor parameters and UPDRS scores + DNN/LR for PD diagnosis | For motor parameters-UPDRS scores correlation: r = 0.838–0.849; Best performing for diagnosis: DNN with AUROC = 0.888–0.950 |
| Talitckii et al. (2021) [65] | Differential diagnosis between PD and other extrapyramidal disorders | 41 PD patients and 15 patients with other extrapyramidal disorders | Accelerometer, gyroscope and magnetometer signals form a dorsal-mounted sensor while performing UPDRS-related tasks | STFT + feature extraction + linear-PCA/RBF-PCA/poly-PCA/LDA/FA + RF/SVM/LR/LightGBM/stacked ensemble model | Best performing: standard classifier with ACC = 72–85%, PREC = 72–85%, REC = 77–100%, F1-SCORE = 76–88% |
| Varghese et al. (2020) [66] | PD or other movement disorders (DD) patients and healthy controls classification; PD patients and DD patients or healthy controls classification | 192 PD patients, 75 DD patients and 51 healthy controls | Accelerometer data from 2 smartwatches placed at both hands and answers from electronic questionnaires distributed via smartphones | FFT + PCA + RBF-SVM/RF/ANN | Best performing for PD/DD detection: ANN with ACC = 89%, PREC = 94%, REC = 92%, F1-SCORE = 93%; for PD detection: RBF-SVM with ACC = 79%, PREC = 81%, REC = 89%, F1-SCORE = 85% |
| Loaiza Duque et al. (2020) [67] | Healthy and trembling subjects classification; PD-ET differential diagnosis | 19 PD patients, 20 ET patients and 12 healthy controls | Angular velocity signals from smartphone built-in triaxial gyroscope with the help of a smartphone application | Kinematic features extraction + feature selection based on chi-square and unbiased tree + linear-SVM/LR/DA/NB/DT/ensemble subspace kNN | Best performing for trembling patients detection: ensemble subspace kNN with ACC = 97.2%, SENS = 98.5%, SPEC = 93.3%; for PD-ET differential diagnosis: linear-SVM with ACC = 77.8%, SENS = 75.7%, SPEC = 80% |
| Channa et al. (2021) [83] | Classification between PD patients with tremor and PD patients with bradykinesia and healthy controls | 10 PD patients with tremor, 10 PD patients with bradykinesia and 20 healthy controls | Accelerometer and gyroscope signals from a smart bracelet | Butterworth bandpass IIR filter + FFT to extract both time and frequency domain features + feature selection based on ANOVA test + NN-SOMs clustering/kNN | Best performing: kNN with ACC = 91.7%, SENS = 83–100%, SPEC = 89–100% |
| Li et al. (2019) [57] | Diagnosis and severity (H&Y scores) estimation | 13 PD patients and 12 healthy controls | Acceleration and angular velocity signals from a prototype handle for spoons with embedded inertial sensors | IIR filtering + windowing + feature extraction + normalization + kNN/Adaboost/RF/linear-SVM for diagnosis; RF regression for severity estimation | Best performing for diagnosis: linear-SVM with ACC = 92%, SENS = 92.31%, SPEC = 91.67%, AUROC = 0.98; for severity estimation: MAE = 0.166, r = 0.97 |
| Koçer et al. (2016) [56] | Diagnosis and severity (H&Y scores) estimation | 35 PD patients and 20 healthy controls | Resting tremor acceleration signals from the Nintendo Wii Remote (Wiimote) | Windowing + feature extraction (with FFT) + SVM | For diagnosis: ACC = 89%, PREC = 91%, REC = 94%; for severity estimation: ACC = 33–77% |
| Bazgir et al. (2015) [71] | Severity estimation (UPDRS scores classification) | 52 PD patients | Accelerometer and gyroscope data from a smartphone placed at the wrist | Filtering + STFT + MLP trained with the back propagation algorithm | ACC = 91%, SENS = 89.6%, SPEC = 90.64% |
| Kim et al. (2018) [72] | Severity estimation (UPDRS scores classification) | 92 PD patients | Accelerometer and gyroscope signals from a wrist-worn device | High pass filter + FFT + RF/NB/linear regression/DT/MLP/SVM/CNN | Best performing: CNN with ACC = 85%, kappa = 0.85, r = 0.93, RMSE = 0.35 |
| Dai et al. (2021) [73] | Severity estimation (MDS-UPDRS scores classification) | 42 PD patients and 30 healthy controls | Accelerometer, gyroscope and geomagnetic data from a finger-mounted sensor while measuring rest and postural tremor and during finger tapping | Denoising with an IIR bandpass filter + FFT + SVM/RF/kNN | Best performing: SVM with ACC = 96–97.33%, SENS = 96.36–100%, SPEC = 95–96.67% |
| Khodakarami et al. (2019) [74] | Severity estimation (UPDRS scores classification) and prediction of response to levodopa (absolute value and percentage) | 151 PD patients and 174 healthy controls | Signals from the wrist-worn smartwatch of the Parkinson’s Kinectigraph system | Feature extraction + JMIM-based feature selection + (+PCA) LR/RBF-SVM/gradient boosting DTs | Best performing for severity estimation: LR with AUROC = 0.79–0.88, AUPR = 0.65–0.88; for absolute levodopa response estimation AUROC = 0.92 and AUPR = 0.87; for levodopa response percentage estimation AUROC = 0.82, AUPR = 0.73 |
| Javed et al. (2018) [58] | Diagnosis and treatment response index (TRIS) estimation | 19 PD patients and 22 healthy controls | Accelerometer and gyroscope data from both wrists using SHIMMER3 sensors while performing hand rotation tests before and after the dose administration | PCA/stepwise regression/LASSO regression + SVM/linear regression/DT/RF | Best performing for diagnosis: stepwise regression + SVM with ACC = 89%; for TRIS estimation: RMSE = 0.69, r = 0.84 |
| Watts et al. (2021) [100] | PD patients classification according to their levodopa regimens and response | 26 PD patients | Bradykinesia and dyskinesia-related signals from Personal KinetiGraph (PKG), demographics and MDS-UPDRS-III scores | k-means for clustering based on regimen features + RF for classification based on PKG features, demographics and MDS-UPDRS-III scores | ACC = 86.9%, SENS = 86.5%, SPEC = 87.7%, PPV = 95.3%, F1-SCORE = 90.7%, AUROC = 0.871 |
| Pfister et al. (2020) [98] | On/Off/Dyskinesia motor states classification | 30 PD patients | Data from a wrist-worn accelerometer in a free-living environment | Data augmentation + CNN/SVM/kNN/RF/MLP | Best performing: CNN with ACC = 65.4%, Kohen’s Kappa = 0.47, SENS = 64.45–66.68%, SPEC = 66.72–89.48%, F1-SCORE = 62.4–69.01%, 1vsALL ACC = 66.7–82.56% |
| Eskofier et al. (2016) [79] | Bradykinesia detection | 10 PD patients | Accelerometer data from IMUs mounted on both hands | Timeseries segmentation + AdaBoost/PART/kNN/SVM/CNN-DNN | Best performing: DNN ACC = 90.9% |
| Shawen et al. (2020) [80] | Tremor and bradykinesia detection; severity estimation (UPDRS scores classification) | 13 PD patients | Accelerometer and gyroscope signals from a flexible skin-mounted sensors and accelerometer signals from a wrist-worn smartwatch | Cubic spline interpolation + segmentation + high-pass filtering + time, frequency, entropy, correlation and derivative-based feature extraction + RF | For tremor detection: AUROC = 0.68–0.79; for bradykinesia detection: AUROC = 0.61–0.69; for tremor severity estimation: AUROC = 0.67–0.77; for bradykinesia severity estimation: AUROC = 0.59–0.66 |
| San-Segundo et al. (2020) [81] | Tremor detection; tremor duration estimation | 12 PD patients for laboratory set and 6 PD patients for in-the-wild set | Accelerometer signals from wrist-worn sensors | Downsampling + FFT + unsupervised non-negative tremor factorization + feature extraction manually/with a CNN + RF/MLP | Best performing for tremor detection: CNN + MLP with AUC = 0.887; for tremor duration estimation: MAE = 4.1–9.1% |
| Ibrahim et al. (2020) [82] | Tremor onset detection | 13 PD patients | Signals from hand-mounted IMUs while performing 6 different rest, postural and motor tasks | Butterworth filtering + zero-phase shifting + Hilbert-Huang Transform + MLP | ACC = 92.9%, PREC = 98.7%, REC = 86.7%, SPEC = 98.9%, F1-SCORE = 0.923 |
| Som et al. (2020) [61] | Diagnosis | 152 healthy subjects for pre-training and 18 PD patients and 16 healthy controls for the final classification | Accelerometers signals from a wrist-worn sensor while performing various ADL for the pre-training + accelerometer signals from 6 sensors located at the sternum, the lumbar, both ankles and wrists while walking | Zero-centering + normalization + segmentation + feature extraction with pre-trained convolutional AE + PCA/global-average-pool layer + SVM/MLP | Best performing: MLP with ACC = 68.64–73.81%, PREC = 69.27–76.53%, REC = 68.64–73.81%, F1-SCORE = 67.65–73.89% |
| Ricci et al. (2020) [62] | Diagnosis | 30 newly diagnosed untreated PD patients and 30 healthy controls | Acceleration, angular velocity and orientation signals from a network of 14 IMUs distributed in the whole body | Feature selection with ReliefF ranking and Kruskal–Wallis + NB/kNN/SVM | Best performing: SVM with ACC = 95%, PREC = 95.1%, AUROC = 0.95 |
| De Vos et al. (2020) [69] | PD and PSP differential diagnosis | 20 PD patients and 21 PSP patients | Accelerometer, gyroscope and magnetometer signals from 6 IMUs placed on the lumbar spine, the sternum, both wrists and feet | ANOVA + LASSO + LR/RF | Best performing: RF with ACC = 88%, SENS = 86%, SPEC = 90% |
| Moon et al. (2020) [68] | PD and ET differential diagnosis | 524 PD patients and 34 ET patients | Balance and gait characteristics from 6 IMUs placed on the lumbar spine, the sternum, both wrists and feet | NN/SVM/kNN/DT/RF/LR | Best performing: NN with ACC = 89%, PREC = 61%, REC = 61%, F1-SCORE = 61% |
| Kuhner et al. (2017) [63] | Diagnosis; correlation with severity metrics | 14 PD patients and 26 healthy controls | Fusion of accelerometer, gyroscope and magnetometer signals from XSens motion capture suit while performing several motor tests | RF with probability distributions for classification PCA for correlation |
For diagnosis: ACC = 86–94.6%, SENS up to 91.5% and SPEC up to 97.2%; for correlation between the 1st pc and the UPDRS scores: r = 0.79 |
| Caramia et al. (2018) [64] | Diagnosis and severity estimation (H&Y scores classification) | 25 PD patients and 25 healthy controls | Accelerometer, gyroscope and magnetometer signals from 8 IMUs attached to both feet dorsum, thighs, shanks and to the chest and lumbar | Extraction of range of motion parameters/spatio-temporal parameters (+PCA) + LDA/NB/kNN/linear-SVM/RBF-SVM/DT + majority voting with equal weights/weights analogue to the individual accuracies | Best performing for diagnosis: majority voting with weights analogue to the individual accuracies with ACC = 96%; for severity estimation: RBF-SVM with ACC = 87.75–94.5% |
| Hssayeni et al. (2021) [76] | Severity (UPDRS-III scores) estimation | 24 PD patients | Angular velocity from one wrist-mounted and one ankle-mounted inertial sensor | Gradient tree boosting/dual-channel LSTM with hand-crafted features and with or without transfer learning/1D-CNN-LSTM for raw signals/2D-CNN-LSTM for time-frequency data + ensemble learning | Best performing: ensemble of dual-channel LSTM with hand-crafted features and transfer learning, 1D-CNN-LSTM for raw signals and 2D-CNN-LSTM for time-frequency data with r = 0.79, MAE = 5.95 |
| Butt et al. (2020) [77] | Severity (MDS-UPDRS III scores) estimation | 64 PD patients and 50 healthy controls | Gyroscope and geomagnetic data from 4 wrist-mounted IMUs and 1 foot-mounted IMU | Kolmogorov–Smirnov test + Mann–Whitney U-test + normalization + CFS/PCA ranker/correlation attribute evaluation/chi-square attribute evaluation/wrapper subset evaluation + SVR/RF/LR/ANFIS | Best performing: CFS + ANFIS with r = 0.814, RMSE = 0.101 |
| Mirelman et al. (2021) [75] | Severity estimation (H&Y stages classification) | 332 PD patients and 100 healthy controls | Accelerometer and gyroscope signals from sensors placed on the ankles, the wrists and the lower back while performing various walking tests + demographic data | Low-pass Butterworth filter + feature selection based on RF permutation importance/neighborhood component analysis/mRMR + RUSBoost + DT/QDA for weak learner | SENS = 72–84%, SPEC = 69–80%, AUROC = 0.76–0.90 |
| Stamate et al. (2018) [78] | Identification of failures to follow the UPDRS-III movement protocol | 12 PD patients | Motor signals from smartphone sensors with the cloudUPDRS application | Filtering + frequency transformations + extra tree/Bernoulli NB/Gaussian NB/MLP/RF/Gradient Boosting/Bagging/AdaBoost/RCNN | ACC = 78%, F1-SCORE = 82%, AUROC = 0.87 |
| Belgiovine et al. (2018) [99] | L-dopa induced dyskinesia detection | 18 PD patients | Accelerometer and angular velocity data from smartphone placed on the wrist (for upper-limb experiment) or on the hip (for lower-limb experiment) | z-score normalization + DT/Gaussian-SVM/linear-SVM | Best performing: SVM (with both kernels) with ACC = 65.0–82.0%, MACRO F1-SCORE = 0.65–0.82 |