Table 1.

Summary of highlighted research articles for tremor analysis using wearable sensors.

Work	Year	Sensors and Location	Participants	Analysis Methods	Main Aims or Findings	Main Results
Patel et al. [36]	2010	8 uniaxial accelerometers, located on the arms and legs	12 PD patients	Machine learning: Support Vector Machines	The results indicate that is possible to estimate clinical scores with a low error	Error values of 3.4% for tremor detection using hand-crafted features.
Rigas et al. [37]	2012	6 accelerometers located on the wrists, ankles, sternum, and the waist	23 subjects (18 PD patients and 5 healthy controls)	Hidden Markov model	High accuracy in tremor detection. It is possible to discriminate tremors from other PD symptoms.	Accuracy of 87% for tremor severity, maximum specificity 97%, and sensitivity 95%
Roy et al. [38]	2013	EMG and 4 accelerometers located on both forearms and shanks	11 PD for training, and 12 for testing (8 PD + 4 healthy controls)	Machine learning: Dynamic neural network	The use of a hybrid sensor and a neural network to detect tremor during unconstrained activities	Overall mean specificity of 90.2% and sensitivity of 92.9% for tremor detection
Tzallas et al. [39]	2014	4 triaxial accelerometers and 1 gyroscope on both wrists, ankles, and the waist	20 PD patients for short-term analysis and 24 for long-term analysis	Machine learning	The authors propose a system for continuous evaluation of tremor and motor symptoms	Accuracy of 87% for tremor classification
Ahlrichs et al. [40]	2014	Triaxial accelerometer located on the wrist	76 subjects (64 PD for testing and 12 PD for training)	Machine learning: Support vector machines	The results indicate that frequency domain features may be enough to detect tremor	Specificity of 88.4% and sensitivity of 89.4%
Kostikis et al. [57]	2015	Smartphone accelerometer and gyroscope, located on the wrists with a glove	25 PD patients and 20 age-matched healthy controls	Machine learning	The authors propose the use of consumer smartphones to assess tremor	Accuracy of 82% for PD patients and 90% for healthy controls
Braybrook et al. [43]	2016	A triaxial accelerometer located on the wrist of the most affected side	85 PD patients	Threshold method using a spectral analysis	The authors propose a system for ambulatory assessment of PD tremor	Specificity of 92.5% and selectivity of 92.9% for tremor detection
García-Magariño et al. [45]	2016	Smartphone with triaxial accelerometer used in unconstrained environments	11 PD with tremors 10 subjects without tremor	Algorithmic approach	The authors propose a smartphone-based application for detecting hand tremor	Specificity of 95.8% and Sensitivity of 99.5% for tremor detection
Jeon et al. [46]	2017	A watch-like device with an accelerometer and gyroscope located on the wrist and finger of both hands	85 PD patients	Machine learning: Several algorithms	The authors show an accurate scoring system for estimating the tremor severity	Accuracy of 85.5%, and an RMSE of 0.410 (on five classes according to UPDRS)
Pulliam et al. [47]	2018	Triaxial accelerometer and gyroscope located on both wrists and ankles	13 PD patients	Regression models	The use of wearable sensors to quantify the dose–response for several symptoms	AUC 0.89 for tremor detection.
Kim et al. [50]	2018	Triaxial accelerometer and gyroscope located on the wrist and finger of both hands	92 PD patients	Deep learning: Convolutional neural networks (CNNs)	The use of CNN outperforms machine learning approaches	Accuracy of 0.85, and RMSE of 0.35
López-Blanco et al. [51]	2019	Consumer smartwatch (accelerometer and gyroscope)	22 PD patients	The tremor intensity calculated through RMS of the gyroscope signal	The use of consumer smartwatches for tremor quantification is reliable and well-correlated with clinical scores.	Spearman coefficient between UPDRS scores and smartwatch measurements for the intensity of 0.81 (p < 0.001)
Hssayeni et al. [15]	2019	Triaxial gyroscope located on the wrist and ankle of the most-affected side	24 PD patients	Machine and deep learning: Recurrent neural networks	The authors propose the use of gradient tree boosting and Long short-term memory networks	A correlation r = 0.93 with LOSO using gradient tree boosting algorithm
Pierleoni et al. [52]	2019	Watch like device with a triaxial accelerometer, gyroscope, and magnetometer	40 PD patients	Threshold method	A method for continuous and real-time monitoring of PD using wearables and Cloud services	Accuracy of 97.7% for tremor detection
Battista et al. [53]	2020	Watch-like device with a triaxial accelerometer	20 PD patients	Threshold method	A device for continuous monitoring of PD tremor, increasing the discrimination with normal daily activities	Linear correlation with the UPDRS constancy r = 0.744 (p = 0.0004)
Van Brummelen et al. [54]	2020	Consumer devices with triaxial accelerometers	10 patients with PD and 10 with essential tremor	Spectral analysis	The authors evaluate the performance of the accelerometers of consumer devices.	Consumer devices could be suitable to analyze the peak frequency in PD tremor
Mahadevan et al. [55]	2020	Watch-like IMU (results reported with the use of the accelerometer)	31 PD and 50 healthy controls for evaluating the tremor detector	Threshold method and machine learning	Sensor measures of resting tremor present high agreement with clinical scoring	Tremor classifier accuracy of 83%. Pearson correlation of 0.97 for tremor constancy
San Segundo et al. [56]	2020	Wrist-worn triaxial accelerometers	12 patients with PD	Deep learning: Convolutional neural networks	Evaluation of novel preprocessing methods and algorithms in free-living and laboratory settings	Error lower than 5% when estimating the percentage of tremor in a laboratory setting