Table A3.
Studies based on image and depth sensors.
| Study | Problem | Dataset Population | Input Data | Analysis/Algorithms | Evaluation |
|---|---|---|---|---|---|
| Guayacán et al. (2020) [111] | Diagnosis | 11 PD patients and 11 healthy controls | Video recordings while walking | 3D spatio-temporal CNN | ACC = 88–90% |
| Reyes et al. (2019) [109] | Diagnosis | 88 PD patients and 94 healthy controls | Gait samples from MS Kinect | Cropping noisy parts + LSTM/1D-CNN/CNN-LSTM | Best performing: CNN-LSTM with ACC = 83.1%, PREC = 83.5%, REC = 83.4%, F1-SCORE = 81%, Kappa = 64% |
| Buongiorno (2019) [110] | Diagnosis and severity estimation (mild vs. moderate) | 16 PD patients and 14 healthy controls | Postural and kinematics features from MS Kinect v2 sensor while performing 3 motor exercises (gait, finger and foot tapping) | Feature selection + SVM/ANN | Best performing for diagnosis: gait-based ANN with ACC = 89.4%, SENS = 87.0%, SPEC = 91.8%; for severity estimation: ACC = 95.0%, SENS = 90.0%, SPEC = 99.0% |
| Grammatikopoulou et al. (2019) [117] | Severity estimation (UPDRS scores classification) | 12 advanced PD patients and 6 PD patients in initial stage | Skeletal features from MS Kinect v2 RGB videos while playing an exergame | Transformation to local coordinate system + two parallel LSTMs (the 1st trained with raw joint coordinates and the 2nd with joint line distances) | ACC = 77.7% |
| Tucker et al. (2015) [122] | Medication adherence estimation (on/off medication classification) | 7 PD patients | Skeletal joints 3D position, velocity and acceleration from MS Kinect | C4.5 DT for generalized model; C4.5 DT, RF, SVM, IBk for personalized models | for generalized model: ACC = 36.2–77.9%; for personalized models: ACC = 67.7–100% |
| Li et al. (2018) [120] | TUG subtasks segmentation and time estimation for each subtask | 24 PD patients | Video recordings while performing TUG tests | Pose estimation with OpenPose/Iterator Error Feedback + SVM/LSTM | Best performing: OpenPose + LSTM with ACC = 93.1%, PREC = 80.8–97.5%, REC = 86.3–97%, F1-SCORE = 83.5–97.3% for subtasks segmentation and MAE = 0.32–1.07 for time estimation |
| Wei et al. (2019) [123] | Development of a virtual physical therapist: movement recognition (repetitions and sub-actions detection), patient’s errors identification (satisfactory/non-satisfactory performance), task recommendation (regress/repeat/progress) | 35 PD patients | Motion data recorded by MS Kinect v2 sensor while performing 3 balance/agility tasks | HMM for repetitions and sub-actions detection + linear-SVM for movement errors identification + {majority undersampling/minority oversampling/decision threshold adjustment/hybrid oversampling with feature standardization and interpolation + RF} for task recommendation | For repetitions detection: ACC = 97.1–99.4%; for sub-actions segmentation: SENS = 88.4–96.9%, SPEC = 97.2–98.8%; for errors identification: ACC = 86.3–94.2%; best performing for tasks recommendation: hybrid oversampling + RF with ACC = 81.8–95.7%, FPR = 2.8–5.4% |
| Hu et al. (2019) [121] | FoG detection | 45 PD patients | Videos collected while performing TUG tests | Graph representation of videos + pretrained features (Res-Net 50 vertex, C3D vertex and context features) + graph sequence-RNN (Bi-directional GS-GRU/Bi-directional GS-LSTM/forward GS-GRU/forward GS-LSTM) + fusion | Best performing: linear fusion of Bi-directional GS-GRU with context model with AUC = 0.90, SENS = 83.8%, SPEC = 82.3%, ACC = 82.5%, Youden’s J = 0.66, FPR = 17.7%, FNR = 16.2% |
| Li et al. (2018) [118] | Binary classification between pathological (PD/LID) and normal motion; multiclass classification (PD with LID, PD without LID and normal); levodopa-induced dyskinesia severity (UDysRS-III scores) estimation; parkinsonism severity (UPDRS-III scores) estimation | 9 PD patients | 2D-Video recordings while performing communication and drinking tasks (for dyskinesia detection) and while performing leg agility and toe tapping tasks (for parkinsonism detection) | Convolutional pose estimators + RF | For binary classification: AUC = 0.634–0.930, F1-SCORE = 50–90.6%; for multiclass classification: ACC = 71.4%, SENS = 83.5–96.2%, SPEC = 68.4–95.7%; for UDysRS-III estimation: RMSE = 2.906, r = 0.741; for UPDRS-III estimation: RMSE = 7.765, r = 0.53 |
| Vivar-Estudillo et al. (2018) [112] | Diagnosis | 18 PD patients and 22 healthy controls | Position, velocity and rotation data regarding hand movements from leap motion sensor | Texture features extraction with SDH + kNN/SVM/DT/LDA/LR/ensembles | Best performing: bagged tree with ACC = 98.62%, SENS = 98.43%, SPEC = 98.80%, PREC = 98.80% |
| Moshkova et al. (2020) [113] | Diagnosis | 16 PD patients and 16 healthy controls | Signals from leap motion sensor while performing hand motor tasks according to the MDS-UPDRS-III | Features extraction + kNN/SVM/DT/RF | Best performing: SVM with ACC = 98.4% when features are extracted from all the tasks |
| Ali et al. (2020) [114] | Diagnosis; classification between PD patients with medication, without medication and healthy controls | 87 PD patients with medication, 119 PD patients without medication and 139 healthy controls | Videos while performing hand motor tasks | Segmentation to frames + temporal segmentation with CNN + spatial segmentation with CNN-AE + FFT for feature extraction + SVM | Best performance when combining 2 tasks for diagnosis: ACC = 91.8%; for 3-class classification: ACC = 73.5% |
| Liu et al. (2019) [119] | Severity estimation (Bradykinesia-related MDS-UPDRS scores classification) | 60 PD patients | Video recordings while performing hand motor tests | Pose estimator NN + feature extraction + kNN/RF/linear-SVM/RBF-SVM | Best performing: RBF-SVM with ACC = 89.7%, PREC = 20–100%, REC = 60–100%, F1-SCORE = 33.3–100% |
| Rajnoha et al. (2018) [116] | Diagnosis | 50 PD patients and 50 age-matched healthy controls | Face images extracted from video recordings | HOG for face detection + CNN for embeddings generation + kNN/DT/RF/XGBoost/SVM | Best performing: DT:ACC = 67.33% with leave-one-out cross validation, RF: ACC = 60.7–85.92% with train-test split |
| Jin et al. (2020) [115] | Diagnosis | 33 PD patients and 31 elderly healthy subjects | Short videos while imitating images of smiley people | Splitting videos to frames + coordinate points extraction with Face++ + transformation from absolute to relative coordinates + features extraction + LASSO + LR/SVM/DT/RF/LSTM/RNN | Best performing: SVM with PREC = 99%, REC = 99%, F1-SCORE = 99% |