The PKG™, from Global Kinetics Pty Ltd., is a wrist-worn device designed to continuously monitor the motor symptoms of Parkinson’s Disease. The patient uses the device for 6 days, and the data is then uploaded to a server where it is processed, and a report is sent to the neurologist. The device has been approved under CE (Class IIa), FDA (Class 2), and TGA.

The PKG™ algorithm was published in 2012 by Griffiths et al. (19). The authors introduced a method based on accelerometer signal analysis obtained from the wrist during 2 min windows. From this window, it is analyzed the frequency characteristics between 0.2 Hz and 4 Hz, the maximum acceleration reached, and the time without movement, from which two indices are generated. One is associated with bradykinesia (BKS) and the other with dyskinesia (DKS), which are then represented on a graph with interquartile ranges to determine the severity of one symptom or the other. There is no evidence of a training-evaluation data method, so it is not considered a machine-learning algorithm. Validation was performed through the median of all BKS samples over 9 h across 10 days and correlated with the UPDRS, obtaining a significant r value = 0.64, p < 0.0005. A third score called FDS was designed to measure motor fluctuations (20). This score, expressed as an algebraic combination of BKS and DKS, determines if a patient is fluctuating. The device has been widely tested in clinical conditions and compared with UPDRS (21) or diaries (8, 20). For example, the work of Santiago et al. (22) determined that in 41% of neurologists, PKG™ provided more information than classic routine visits. In a study by Nahab et al., the authors also indicate utility in clinical practice (23). Furthermore, the system has shown good results in usability (24), although there were some scores to be discussed. For example, only 27% of patients scored positively on the report of PKG, and 59% rated the use of PKG as valuable in offering useful information to clinicians. Interesting articles can be found, such as the one from Ossig et al. (8). In this study, the agreement was moderate to high in the total number of OFF and ON with and without dyskinesia (K = 0.404 in OFF with bradykinesia and K = 0.658 with ON with dyskinesia). However, a low to moderate agreement was found if the agreement on every single-hour-level (in OFF with bradykinesia K = 0.215 and in ON with dyskinesia correlation of K = 0.324). A similar situation is observed in Löhle et al. (25), where there is a moderate correlation between PKG and clinicians when total hours are compared (0.43 in bradykinesia and 0.51 in dyskinesia). However, a poor correlation is noted when outcomes are compared every 30 min (0.13 in bradykinesia and 0.21 in dyskinesia). According to Monje et al. (11), the PKG™ has been widely validated but needs further independent validation. In the same line, the NICE highlights in the published guideline that PKG is the device with more clinical validations but more evidence is needed.

Great Lakes Neurotechnologies (GLN) is the main manufacturer of Kinesia 360, a double-device system (wrist and ankle), and Kinesia U™, a wrist-worn device. The device incorporates a triaxial accelerometer and gyroscope. Among all the commercialized existing devices, GLN was the first company to achieve FDA 510(K) clearance as a tremor transducer device. In Europe, Kinesia 360 is considered a Class I device. Kinesia 360™ provides information about tremor, dyskinesia, slowness, mobility, posture, and steps (26). The quantification of bradykinesia conducted on an ankle-mounted device is based on the analysis of specific frequency characteristics derived from both the accelerometer and gyroscope and is calculated using linear regression models which are correlated with UPDRS scores (27, 28). The dyskinesia algorithm also relies on a linear regression model with sensors worn on the more affected side of the body (29). However, the number of features evaluated is significant, totaling 18. The obtained correlation with the modified Abnormal Involuntary Movement Scale (mAIMS) is significant (r = 0.77). Both models to determine bradykinesia and dyskinesia are more complex than the method presented by Griffiths et al. for the PKG due to the significant amount of features extracted from the accelerometer and gyroscope. The Kinesia 360 device also measures essential tremor, which was tested with 20 PD patients with intraclass correlation coefficients over 0.7 (30). Kinesia 360 has been evaluated in numerous studies and demonstrated its effectiveness in some therapies such as levodopa (31), rotigotine (32), deep brain stimulation (33), or subthalamic stimulation (34).

GLN also launched a device called Kinesia U™ for continuous monitoring but also for specific tasks. The main advantage of this new device was to eliminate the ankle-mounted device, improving the adherence of the patient to this technology by using only a wrist device. However, it is not clear how this affected the algorithm. The new app includes the capacity to fill in a diary and rate their symptoms. Results are provided in a 0 to 4 score indicating the severity of each symptom. In Pulliam et al. (31), it was shown that Kinesia presents a good accuracy for tremor, dyskinesia, and bradykinesia, being tremor and dyskinesia the best symptoms detected with the Area under curve (AUC) of 0.89 and 0.86, respectively. The bradykinesia algorithm has an AUC of 0.82 and a false positive rate of 0.34.

PDMonitor™ is a system composed of five devices designed to comprehensively characterize all motor symptoms of a Parkinson’s disease (PD) patient from any part of the body. The device, which is manufactured by PDNeurotechnology, is a CE Certificate device Class IIa. Each device includes an accelerometer, gyroscope, and magnetometer (7). This eliminates the need to choose the most affected side, allowing for the monitoring of movements from the upper limbs, lower limbs, and trunk. The system was designed within the PERFORM project (35, 36), and its algorithms are based on training a database of experts and utilizing machine learning algorithms. The algorithms are briefly described in different publications, including tremor (37), dyskinesia (38), bradykinesia (39), and Freezing of Gait (FoG) (40). Gait parameters and the ON and OFF states are also provided (41). Each algorithm employs a distinct classification method. For instance, the tremor detection algorithm uses Hidden Markov Models and achieves an accuracy of 0.87. The dyskinesia detection algorithm is based on a decision tree with a classification accuracy of 0.85. The bradykinesia detection algorithm utilizes Support Vector Machines and has a classification accuracy of 0.745. For this machine learning classifier, 20 patients participated in clinical settings and 24 in home environments. The algorithm for FoG, on the other hand, employs a Random Forest classifier achieving an accuracy of 0.96. The database was formed by 5 patients with FoG, 6 patients without FoG, and 5 healthy patients (40).

In another study in 2023, Antonini et al. published an article where the accuracy, sensitivity, and specificity of the different symptoms were evaluated against UPDRS and diaries. All the results on sensitivity, specificity, and accuracy for all the algorithms are over 0.8, except for the sensitivity of Gait (0.67). The accuracy and specificity achieve 0.96 or more in all the symptoms. Bradykinesia obtains a moderate-high correlation of 0.68 with UPDRS. In 2021, due to the appearance of COVID-19, a 2-cases study was presented showing the feasibility of PDMonitor™ for monitoring symptoms in home environments (42). PDMonitor™ was also evaluated against UPDRS, achieving moderate to high correlations in all their claimed detected movements/symptoms (41). This article questions the subjectivity of questionnaires, given that showing a device works does not need to correlate highly with questionnaires, which have doubtful outcomes. Although the device offers a comprehensive map of motor symptoms, there are doubts about the wearability of the device due to using five separate devices (13). However, in Antonini et al. (43), it should be mentioned that the scores obtained in wearability are significantly high except for two items: difficulty in putting on the device and that patients would wear the device if it was invisible. These items are of special importance because a patient in the morning usually suffers morning akinesia due to the deep OFF that they could suffer. Having to set up five sensors is an important barrier that could drastically reduce adherence to the technology and is something to consider.

STAT-ON™ is a medical device Class IIa which is worn on the waist (44). The device, which was designed under the project REMPARK (45, 46), has been commercialized by Sense4Care SL. The device aims to minimize the number of sensors while maximizing high accuracy by measuring symptoms from a position near the centre of the human body and using machine learning techniques trained with large databases.

The system detects bradykinesia (47, 48), dyskinesia (49), FoG (50, 51), gait parameters (48, 52), and falls (53), providing ON and OFF outcomes (54). The device is based on machine learning algorithms, more specifically in support vector machine classifiers and support vector regression models. The database was obtained in home environments, and the gold standard was the video and the UPDRS-III. For the validation of the ON and OFF algorithm, diaries were used, and a clinician called the patient every 2 h to ensure the motor state annotated by the patient, creating a robust database. A total of 92 PD patients participated in the database, and 10 extra PD patients participated in the dyskinesia algorithm. The algorithm for ON and OFF was evaluated with 91 PD patients in 3 studies (54–56) obtaining sensitivity and specificity values over 0.92. The bradykinesia algorithm was evaluated with 75 patients employing UPDRS subscales and achieving a moderate-high correlation of 0.67 p < 0.01 (57). The dyskinesia algorithm, evaluated with a leave-one-out method and 102 patients, achieved a sensitivity and specificity of 1 and 0.95 in trunk dyskinesia, 0.90 and 0.95 in strong dyskinesia in limbs and neck, 0.78 and 0.95 in mild dyskinesia in the trunk, and 0.39 and 0.95 in detecting mild dyskinesia in limbs [59]. Finally, the FoG algorithm was evaluated with 15 PD patients with FoG, achieving 0.92 and 0.87 in sensitivity and specificity, respectively.

In an Italian study performed by Zampogna et al. (58), the FoG and Dyskinesia were evaluated in 71 PD patients and an AUC was obtained of 0.92 and an accuracy of 0.8. The FoG evaluation obtained 0.83 on AUC and 0.81 on accuracy comparing FoG patients vs. PD patients without FoG. The number of FoG episodes AUC obtained a score of 0.87.

In Cabo-Lopez et al. (59), the correlation between STAT-ON™ and the clinical decision for considering a patient candidate for second-line therapy was 0.73 (p < 0.001) and a significant association in results between STAT-ON™ and the MANAGE-PD questionnaire (p = 0.004). Finally, a clinical trial with 84 PD patients comparing STAT-ON™ against diaries, having the UPDRS as the gold standard, obtained a correlation of 0.63 (p < 0.001) against the 0.24 obtained by the diaries (60).

Neptune™ from Orbit Health GmbH is a class IIa device that monitors bradykinesia, dyskinesia and ON/OFF states. The algorithm employed deep-learning techniques, specifically convolutional neural networks, for the detection of ON, OFF, and Dyskinesia states. The window size entered into the convolutional neural network is 1 min in length and obtained 0.654 accuracy in a three-class outcome problem. The sensitivity and specificity obtained for OFF was 0.64/0.89, for ON 0.67/0.67, and for Dyskinesia 0.64/0.89. However, the correlation between the complete data obtained and the MDS-UPDRS subscale 3.14 was high, with 0.83 for bradykinesia and with AIMS item 5 was 0.84 for dyskinesia detection. Correlations with lower windows (1 h, 30 min, 5 min, and 1 min) decrease but maintain a moderate/high correlation, going from 0.775 for dyskinesia and 0.735 for bradykinesia for 60 min to 0.703 for dyskinesia and 0.632 for bradykinesia with a window length of 1 min.

The main advantage of deep learning is the high performance of the classifier and the ability to compute complete raw inertial data without treating, extracting, and selecting the key features, being what is called a “black box” (61). However, this has been the focus of discussions due to the transparency of these methods in the field of computer science. Another characteristic of deep learning is that it is used to be a method for training large databases such as video or image with thousands or millions of signals (pixels) per sample. Unfortunately, a triaxial signal accelerometer in 30 PD patients is not considered a large database. This circumstance has the problem of overtraining a classification model leading to false positives in new evaluation data. Thus, more evidence is needed in this device to validate from a computer science point of view their approach.

PD-Watch from Biomedical Lab s.r.l. is a wrist-worn device classified as a Class I medical Device based on inertial sensors. The device is a 43 mm × 40 mm × 13 mm smartwatch and weighs 16 g. It has a battery life of 15 days. The data is recorded and then uploaded to a cloud server, from where it is processed and a report is returned. The device is capable of detecting ON/OFF, bradykinesia (62), tremor (63), dyskinesia (64), and inactivity with the absence of movement. One of the differences with other smartwatches devices is that PD-Watch can also detect dyskinesia severity. All the algorithms are based on extracting frequency features, and a structure of conditions that once met, the frequency response of the accelerometer is compared to a certain threshold (62). Results show more than 0.8 in sensitivity, specificity, and accuracy in all the algorithms. Further and external evidence is needed for this threshold-based algorithmic system.

Feetme™ from Feetme, is a wearable device that consists of two insoles, a device for connecting the insoles, and an app to manage the insoles. This device can continuously monitor movement and foot pressure in order to obtain gait parameters. The device includes a triaxial accelerometer and gyroscope and 18 capacitive pressure sensors (65). Feetme™ has received ClassIm in Europe and Class I FDA 510(k) exempt medical device.

The APP connects to the device which uploads the inertial and pressure data and processes it to provide different outcomes. Some of these outcomes are gait parameters such as stride velocity, cadence, stride length, step time, stride time, swing time, stance time, and single and double support time (66). All these features have been tested to evaluate stroke (67, 68), multiple sclerosis (69), or Parkinson’s Disease (66, 70). The device was evaluated against a recognized gold standard such as Gait-rite also in healthy adults (69, 71). Feetme™ has demonstrated to provide good results in tests such as the 6 min walk test (65). However, in the field of Parkinson’s Disease, Feetme™ has not shown any clinical utility for assessing motor symptoms such as bradykinesia, freezing, dyskinesia, or tremor.

Another similar intended-use device is Nushu, from Magnes, which has been registered in the FDA under a Class 2 device. The main difference is that Nushu are shoes instead of insoles and that the system uses a magnetometer but not pressure sensors (72). Values such as stride velocity, stride time, stride length, swing time, cadence, symmetry and variability of steps are provided. Nushu is based on inertial sensors (triaxial accelerometer and triaxial gyroscope) and a triaxial magnetometer. An orientation algorithm is set if it detects motion or not. The algorithm employed to estimate the orientation is based on Madgwick’s filter, which reduces computational burden compared to classical Kalman filters (73). A segmentation of the signal and event detection based on thresholds are executed to finally extract spatial and temporal gait parameters (74). Nushu’s algorithms also used SVM and reduced-SVM for detecting gait phases, which shows a high algorithmic level, which is used to detect Freezing of Gait and provide a vibrotactile biofeedback to avoid FoG on PD patients (75). The device has been only tested with healthy users, although authors claim that the device is under some clinical trials in the field of Parkinson’s Disease (72).

The device works with an app that allows configuring the type of activity to record. After the test or the recording, which is stored in an internal memory, the data is transferred via Bluetooth to the app and then uploaded to a cloud from where, through a dashboard, the results can be seen (76). Similarly to Feetme™, it is claimed the relation of gait parameters to motor symptoms of Parkinson’s Disease, however, further studies are needed to validate this idea rigorously.

In 2021, Powers et al. (77) published an article where a smartwatch named MM4PD could monitor tremor and dyskinesia in Parkinson’s Disease. The study was divided into three phases. The first one in clinics was video recorded with three movement disorders specialists rating. A subset of subjects participated in a 1-week out-of-clinic measurement period for obtaining data from daily living conditions. In the second phase, 225 PD patients participated, from which 143 were used to design the algorithm and 82 for validation. Finally, a third part was performed with 171 control users. For the gold standard, video was used, but also MDS-UPDRS-III was employed.

The device collected signals from a triaxial accelerometer and triaxial gyroscope at 100 Hz. The resting tremor algorithm was based on an analysis of the signal between 3 and 7 Hz in a 2.56 s window, but it’s not clear the classification method, which significantly detracts from the credibility of the method in the scientific field, as Bloem et al. reports (78). The algorithm classifies essential tremors into slight, mild, moderate, and severe tremors.

The dyskinesia algorithm uses a 10.24 s window overlapped every 2.56 s to not lose information between windows. The algorithm of dyskinesia is not explained although it is reported a low false positive rate despite a possible confusion with walking. Also, in 69 PD patients, the clinicians disagreed with the dyskinesia label.

A total of two solutions resulted from this method: StrivePD from Rune Labs and Parky from H2O Therapeutics. The two devices were FDA 510(K) cleared and claim tremor and dyskinesia in their intended use. Although the algorithms are the same and run on the Apple Watch, the layout of their solutions and their services are different.

On the other hand, NeuroRPM from NeuroRPM Inc. is another software that runs on an Apple Watch and besides detecting dyskinesia and tremor, it also detects Bradykinesia according to its intended use (FDA 510(k) Clearance number K221772). There is no scientific evidence apart from the summary report provided in the FDA clearance which shows a study with 36 PD patients and a sensitivity and specificity of 0.718 and 0.951 for tremor, 0.714 and 0.774 for bradykinesia, and 0.712 and 0.947 for dyskinesia. The outcomes are provided every 15 min and are binary for each of the three outcomes: tremor (yes/no), dyskinesia (yes/no), and bradykinesia (low to normal/severe). NeuroRPM has registered a clinical trial (NCT05680961) executed at the Parkinson’s & Movement Disorders Centre of Maryland. The lack of information on the algorithm is a major drawback, taking into account the significant amount of smartwatches that obtain these measurements.

Surface electromyography (sEMG) is a non-invasive technique used to measure the electrical activity produced by skeletal muscles, making it a valuable tool in the research and management of Parkinson’s disease (79–82). sEMG involves placing electrodes on the skin above specific muscles to detect the electrical signals generated during muscle contraction and relaxation. In the context of Parkinson’s disease, sEMG can be used to analyze the muscle activity patterns that are often disrupted by the condition (rigidity or dystonic dyskinesia).

sEMG can also help in monitoring tremors by providing detailed data on their frequency and intensity, or in assessing muscle rigidity and coordination (83). With the help of an accelerometer, the system could provide a comprehensive state of the patient (84). In the market, we can find Adamant Health (85), a company that commercializes a solution for capturing sEMG signals from different parts of the body with small wearable sEMG sensors and provide outcomes about the motor states of PD patients. Adamant Health has a wide scientific background in the recognition of motor symptoms in Parkinson’s Disease (81, 83, 84, 86–88). On the other hand, Paragit Neurotech provides a specific device with EMG that can be mounted on a patch at any part of the body (89). The device contains an accelerometer, a gyroscope, and an internal memory for data logging. The sensor fusion enables identifying the appearance of tremor, stiffness and according to the information on the webpage, bradykinesia and dyskinesia.

Other devices such as Ultium from Noraxon (90), Trigno from Delsys (91), Freeemg from BTS Bioengineering (92, 93), and Mini Wave from Cometa (94) have been used to characterize some motor symptoms or gait parameters in PD. These device systems are composed of several sensors and can be set at any part of the body. Given that there is no standardization on the point where the sensors must be located, these devices are mainly intended to be used in research and clinical studies. As far as authors know, only Adamant Health and Paragit Neurotech have incorporated algorithms that allow quantifying symptoms.