Table A5.

Studies based on other types of sensors.

Study	Problem	Dataset Population	Input Data	Analysis/Algorithms	Evaluation
Rahman et al. (2020) [133]	Diagnosis	5 PD patients and 5 healthy controls	EEG signals from a portable headset with sensors placed at the forehead while watching 4 videos which provoke 4 different emotions	Feed-forward NN trained with Adam optimization algorithm	ACC = 96.5%, PREC = 95.5%, REC = 97%, F1-SCORE = 97.6%
Kleinholdermann et al. (2021) [134]	Severity (MDS-UPDRS III scores) estimation	45 PD patients	sEMG signals from a wrist-worn band while performing a simple tapping activity	Windowing + feature extraction + linear regression/poly-SVM/kNN/RF	Best performing: RF regression with r = 0.739
Capecci et al. (2019) [135]	Emotion (positive/negative) recognition	36 PD patients	Body temperature, heart rate and galvanic response from smartwatch sensors	Linear-SVM/poly-SVM/RBF-SVM	Best performing: RBF-SVM with ACC = 88.6–91.3%
Lacy et al. (2018) [136]	Diagnosis	49 PD patients and 41 healthy controls (1st dataset); 58 PD patients and 29 healthy controls (2nd dataset)	Position measures from 2 electromagnetic sensors located at the thumb and index finger while performing finger-tapping tests	Low-pass Butterworth filtering + velocity and acceleration features extraction from derivatives + ESN	AUROC = 0.802
Picardi et al. (2017) [137]	Diagnosis; classification between different cognition levels (PD patients with normal cognition-PDNC, PD patients with mild cognitive impairment-PDMCI and PD patients with dementia-PDD)	22 PDNC, 23 PDMCI, 10 PDD and 30 age-matched healthy controls	Flexion signals from a glove with finger-mounted sensors and position and orientation information from a wrist-worn tracking system	Feature extraction + Cartesian Genetic Programming/SVM/ANN	Similar performance for all the algorithms: AUROC = 0.72–0.99 for all the pair-wise classifications
Memedi et al. (2015) [138]	Symptom detection (bradykinesia/dyskinesia)	65 advanced PD patients	Spatiotemporal features from spiral drawings, produced with a touchscreen telemetry device	PCA + MLP/RF/RBF-SVM/linear-SVM/LR	Best performing: MLP with ACC = 84.0%, SENS = 75.7%, SPEC = 88.9%, AUROC = 0.86, weighted Kappa = 0.65
Pham et al. (2019) [139]	Diagnosis	42 PD patients and 43 healthy controls from newQWERTY MIT-CSXPD database (https://www.physionet.org/content/nqmitcsxpd/1.0.0/, accessed on 17 February 2022)	keystroke logs time series	CNN/LSTM/CNN-GoogleNet/CNN-AlexNet/LSTM-fuzzy recurrent plots (FRP)	Best performing: LSTM-FRP (m = 3) with ACC = 65.14–81.90%, SENS = 66.67–95%, SPEC = 63.33–66.67%
Matarazzo et al. (2019) [140]	Medication response detection (improved/not changed) and medication response prediction in 21 weeks	29 PD patients and 30 age-matched healthy controls	Keystroke logs with the help of neuroQWERTY software	RNN	For medication response detection: ACC = 76.5%, AUROC = 0.75, kappa = 0.47; for medication response prediction: AUROC = 0.69–0.75