The review was conducted based on the Bargas-Avila and Hornbæk approach [11] and Cochrane methodology [12], and it included 5 phases.

The review identified 27 studies (Figure 1). All the reviewed studies tracked upper limb motion using wearable devices and were divided into 2 main categories according to the purpose of the study: (1) clinical motion tracking and (2) rehabilitation. Some studies examined both applications. More specifically, the majority of papers that were selected (15/27, 56%) [13-27] focused on wearable devices that monitor human movement, and data were collected to be employed for general use in a clinical setting. The remaining papers (12/27, 44%) [28-39] emphasized systems for upper limb rehabilitation. The PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) checklist is provided in Multimedia Appendix 1.

The key features of the studies are summarized and compared in Table 2, according to the purpose of the study, the type and number of sensors used, the placement location of the sensors, the placement method for the sensors, and the measurement method, with division into the 2 aforementioned categories.

As mentioned above, wearable devices are used to capture upper limb function for general health purposes to obtain information and data about physiological parameters to assist rehabilitation patients and therapists. The experiments in most of the studies (22/27) [13-19,21,22,24-26,28,29,31-37,39] were noncontrolled experiments, where participants were either healthy able-bodied individuals or patients who experienced upper limb disorder, and they were recruited for participation in the experimental evaluation. In some studies (4/27) [20,27,30,38], the experiments were controlled, where both patients and healthy participants were included and the results were compared statistically. The number of participants was not specified in 1 study [23].

The findings revealed that most of the studies (22/27) [13-19,23-29,31,32,34-39] had a positive outcome. In 2 cases [22,30], further studies need to be conducted. In the first study, the maximum estimated error exceeded the required accuracy for a typical clinical assessment (over 5 degrees). In the second study, the hand feedback for stroke patients was briefly modifiable, indicating no therapeutic benefit over a short period. Thus, these studies are considered neutral. The outcomes of the remaining 3 studies [20,21,33] were negative as they do not offer any major findings [20] on motion tracking and the application is difficult to implement [21,33] due to limitations such as the system’s measurement accuracy and large errors.

From the above information, it can be concluded that motion-tracking wearable devices have an overall positive impact on the interpretation of the data, and they provide useful assessments that can be used in future systems for clinical and rehabilitation purposes. The aforementioned data are presented in Table 3, categorized according to the purpose of the study.

Wearable upper limb devices provide important information about limb motion through motion analysis. Different signal processing techniques were used, and features were extracted to provide feedback to end users or therapists for better data assessment and interpretation of movement.

Signal processing is essential for such devices to reduce noise signals, transform data, and extract meaningful motion features. A variety of filters are used depending on the data provided by the sensors and the desired output of the system. In 13 studies [14,16-22,26,29,31,32,35], filters were used to remove noise from sensor signals. More specifically, the Kalman filter was used in 2 studies [20,32], which is an algorithm that uses observed measurements over time to produce estimated unknown variables that tend to be more accurate by estimating a joint probability distribution over the variable of each timeframe.

In 3 studies [18,21,32], the Butterworth filter was used to minimize low-frequency noise and high-frequency interferences. Another method to extract data from upper limb devices involves the infinite impulse response filter (IIR) [14,26]. This filter is digital with infinite impulse response and can be designed as a low-pass or high-pass filter. The Madgwick filter is an algorithm that reduces integration errors by the magnetometer and accelerometer, and it was used in 1 study [16]. A combination of 2 of the aforementioned filters was implemented in 1 study [21]. The accelerometer and gyroscope data were low-pass filtered using the Butterworth filter and then passed through the Madgwick filter to integrate the drift of the angular velocity. A complementary filter was implemented in 1 study [31] to reduce the integration drift by providing correct data on the orientation and position of each upper limb segment. More recently, research has been conducted on machine learning (ML) techniques to assist with predicting motion and improving tracking accuracy. For example, the Block Sparse Bayesian Learning (BSBL) algorithm was used in 1 study to reconstruct the accelerometer signal from compressed data [27]. Little et al [19], on the other hand, compared 10 different ML algorithms to predict angle trajectories through the fusion of physiological and kinematic features.

Another important feature of wearable devices is the sampling rates of the sensors, as they determine the quality of the captured data, providing a better understanding of upper limb motion. High sampling rates provide more precise data acquisition; however, battery life is significantly reduced as power consumption is increased. More specifically, in the revised studies, the smallest sampling rate was mentioned in the research conducted by Tolvanen et al [23], where a pressure sensitivity test for light finger touch by a stretch sensor was performed. In contrast, the highest sampling rate was at 2000 Hz in a paper by Pregnolato et al [36], where hand movements of stroke patients for rehabilitation were captured by a 9-DOF IMU with a surface EMG sensor. While both studies provide valuable insights into their respective applications, the battery life was not mentioned. With regard to the battery life, the shortest battery duration was observed in a study by Gu et al [15], where 10 hydrogel-elastomer hybrid ionic sensors were used to capture hand motion, joint bending, hand posture, and gesture. The battery life in this study was 1-2 hours, with an approximate sampling rate of 333.33 Hz, attributed to a 3.3-V rechargeable Li-ion battery. In contrast, the longest battery life was reported by Friedman et al [29], with a duration of 21.5 hours. Power was delivered to 2 triaxial magnetometers and 1 accelerometer to monitor wrist and hand movements, and the sensors were powered by a 3.7-V 450-mAh lithium polymer battery.

The relationship between battery size and sampling rate is crucial as it can affect the efficiency of the wearable sensors. Specifically, in the study conducted by Lee et al [17], a 170-mAh battery (sized 1-2 cm) was used to power an accelerometer with a sampling rate of 67 Hz, and the battery life was approximately 6 hours. In contrast, Wang et al [24] employed a larger 10-cm 3-V battery to power two 9-DOF IMUs with a sampling rate of 50 Hz. This variety of sampling rates and data acquisition reflects the diverse design choices between consumption and device portability in wearable sensor technology.

Feedback is a key feature of the rehabilitation process and motion tracking as it provides important and meaningful information to therapists to interpret the performance of the system. Moreover, it plays an important role in informing patients of their progress, which leads to better recovery chances. The majority of studies (23/27) [13-23,25-27,29-35,37,38] used visual feedback. Three studies provided visual and auditory information [24,36] and haptic feedback along with visual and auditory information [24,28], and 1 study [39] provided only haptic feedback.

More specifically, visual feedback provides therapists and users with more precise information about their tasks and training instructions to achieve the desired position through direct visual analysis of the movement. Haptic feedback is provided by vibrotactile actuators, and small transducers are designed to optimize skin response to vibration. The vibrotactile actuators must have a minimum critical distance for their vibration to be identifiable, and the subjects of the experiment had to rely exclusively on it to accurately rehabilitate their posture and eventually regain their lost muscular abilities. Wearable devices that used a combination of feedback approaches (visual, audio, and haptic) provided an objective outcome that contributed to increasing the effectiveness of training. Moreover, these approaches provide support to therapists, giving them additional information about the patients’ motions, and lastly, the quality of training is improved by detailed feedback signals.

Accuracy of system measurements is necessary for the effectiveness of proposed wearable devices as it plays a vital role in the interpretation of output data and the extraction of user features. Consequently, feedback should be quick enough to improve the operator’s performance in terms of reducing mental effort and informing therapists about movement characteristics. One of the most common ways to present the accuracy of the system and evaluate position and velocity based on a kinematic model is through the calculation of the mean error between the bending joint angle of the user and the proposed validated algorithm with good accuracy, which is used as a reference, under different speeds and magnitudes [14,16,17,19,22,24,25,27,29-37]. Moreover, a statistical analysis is conducted by calculating the “root mean square deviation” (RMSD) or “root mean square error” (RMSE), correlation coefficient (CC), mean absolute error (MAE), and P value. In some papers, accuracy was either not measured or the circumstances under which accuracy was calculated were not mentioned [13,15,18,20,21,23,26,28,35,38,39].

This paper provides a systematic literature review of upper limb wearable device technologies that were developed in the past decade. The screening of papers from 4 different electronic libraries yielded a total of 27 relevant papers that were included for analysis. The papers were classified into 2 major categories, according to the purpose of the study: clinical motion tracking and rehabilitation. The analysis of the findings suggests that upper limb motion is most often tracked using inertial sensors owing to their accuracy, compact size, and effectiveness in assessing range of motion (ROM). Advanced data processing techniques, such as Kalman and Madgwick filters, were used for data fusion, ensuring data accuracy. One of the key elements of wearable technologies is usability, which can be affected by how wearables are placed on a user, their form factor, and their energy consumption. Even though most studies used straps for placing sensors, the other characteristics have greater variability, with no clear consensus among the research community.

The papers were categorized according to the type of study (controlled and uncontrolled). Most of them were uncontrolled studies, which may lead to bias because of the absence of randomly selected control groups and a comparison between them.

The goal of upper limb wearable devices is to provide assistance to therapists and researchers through either monitoring upper limb function for clinical assessment or aiding rehabilitation training and reducing the recovery time [40]. Overall, the developed wearable devices can positively influence the motivation of users, while in the case of rehabilitation, patients can undergo treatment at home by aiming for a high level of independence [41].

Various sensing technologies were used for tracking upper limb motion. Inertial sensors tend to be the most used sensors as they are used to estimate joint angles of the upper limbs. These technologies provide data accuracy, and because of their size, they are used to monitor and provide feedback to patients and therapists on ROM and rehabilitation performance. However, the placement method needs to be considered owing to its essential role in the ROM assessment, as the interpretation of data influences the development of rehabilitation treatment. Moreover, a flexible and well-fitted design can improve signal quality and reduce measurement noise [42]. Consequently, the majority of devices were attached to the body with straps (mostly Velcro straps), as they provide flexibility and great strength (eliminating data bias) and require low maintenance. The hold that Velcro straps provide can be generally characterized as “firm.” Although this placement method is easy and noninvasive [39] and there is no need for external cameras, emitters, or markers [31], the feasibility of the system is not guaranteed for an extended period [25,39]. Similarly, clip-on straps can make the system modular in many aspects. This design can be adopted regardless of hand dimensions and the presence of deformities or inflammation. These straps are manufactured by using a stretchable and flexible material that yields extra comfort and causes minimum disruptions to movement [16].

The placement of wearable sensors significantly impacts their performance, user acceptance, and engineering demands. As sensor technology progresses from wearable to implantable and ingestible forms, challenges arise across regulatory, technical, and translational domains. Misplacement or misalignment of wearable sensors can reduce classifier accuracy; however, some approaches have maintained precision (97%) and recall (98%) at high levels during movement classification [7,9].

Despite the utility of straps, challenges remain in optimal sensor placement for maintaining user comfort and mobility and managing interference between sensors. Overcoming these challenges includes performing comprehensive user studies and data analyses to determine the best sensor placement, as well as designing an ergonomic and adaptable sensor housing for enhancing user comfort [43]. Thus, the ergonomic aspects of the system, such as dimension, weight, and undesirable contact, should be considered to accommodate various hand sizes and deformations [16]. Recent innovations have focused on achieving body compliance, ensuring comfort for the wearer, and maintaining accurate sensing performance [43].

Furthermore, some studies used smart textiles or e-textiles with embedded sensors as the sensing technology, which provided great feasibility regarding wearability. The use of textiles in health care and wellness applications has increased over the last few years and is expected to grow further in the future [41]. The primary benefit of using textile-based electrodes is that there is no direct contact with the skin, and this prevents problems like allergy and skin irritation [44]. Nevertheless, additional research needs to be conducted for integrating accuracy, improving usability, and implementing clinical validation. We expect that the benefits provided by textiles, especially related to ease of use and flexibility, and the advances in technology will make them essential for tracking motion and muscle activity and improving rehabilitation outcomes [44,45].

Wearable devices capture upper limb motion through data acquisition and processing. The examined studies used various measurement methods to record data such as body segment posture, amount of use, and ROM. Undoubtedly, data processing is necessary for developing wearable devices as the signals captured have to be interpreted. The most common filters are Butterworth, low-pass, and band-pass, which are used to weaken potential high-frequency noise in the accelerometer and gyroscope. Moreover, Kalman, Madgwick, and complementary filters are used to fuse the sensor readings and overcome potential biomechanical constraints. These algorithms combine the sensor readings to indicate the rotation and orientation of the arm; however, they tend to require high computational power.

Sampling rates across the analyzed studies varied significantly and ranged from as low as 2 Hz for pressure-sensing technologies to as high as 2000 Hz for motion-sensing applications. Balancing high sampling rates and energy efficiency remains a challenge for upper limb wearable devices, as higher rates enhance tracking precision but can reduce battery life. The wide range of rates indicates the adaptability and flexibility of the sensors used for upper limb applications, as they capture muscle movements as well as rapid motion movements with precision. However, the selection of the right battery to support the desired sampling rates remains a challenge, and there is a need for careful selection to ensure prolonged and uninterrupted data acquisition.

Different techniques are being developed that aim to improve energy efficiency and extend the battery life of wearable devices. For example, compressed sensing [46] allows the reduction of the sampling rate of a signal, which can then be transmitted using a compressed sparse representation and reconstructed with minimal loss compared to the original signal [47,48]. Another algorithm that can be used with time series data, such as motion and muscle data, and can improve energy efficiency is change point detection [49]. The algorithm is used to detect the point when a signal changes and is often used to assist with action recognition [50]. Therefore, by detecting the time point when a signal changes, the sampling rate can be adapted and energy consumption can be reduced during nonrelevant activities (eg, transition between motions, resting, etc) [49]. At the same time, recently, research has been conducted on extending battery life through energy harvesting [51]. The process relies on capturing energy either directly from the human body, including motion [43], or from the environment and converting it to power.

Feedback plays a crucial role in the overall system performance as it provides useful information to therapists and researchers and affects therapy outcomes by influencing motivation. Feedback is given frequently to users who are less proficient or whose posture needs to be improved. This is beneficial for future applications as users should not rely on external feedback, but instead follow their intrinsic feedback mechanics. Visual information on a computer or smartphone/tablet is usually provided, which is very useful, especially for systems that are remotely monitored. Ultimately, the main aim of feedback is to help users improve their performance while providing useful data to therapists for better information processing and future reference.

The improved computation power of electronics and advances in ML algorithms have increased the use of these algorithms for not only motion trajectory prediction [52] but also assessing motion quality and providing relevant feedback during rehabilitation [53]. Using ML with wearable devices through Tiny Machine Learning (TinyML) has gained popularity recently; however, many challenges remain [54]. Further research in this area can revolutionize the use of wearable technology in health care applications by providing greater accuracy, and improved and more personalized feedback. Feedback can also be further reinforced using haptic devices [55]. However, current haptic devices can be cumbersome, and more portable devices need to be developed to make such technologies easy to use in a clinical setting [56].

In this review, an effort was made to cover studies related to upper limb wearable devices, including study purpose, sensing technology, feasibility and wearability of the system, sensor placement, methodology, and feedback received. However, because of the variety of studies conducted in this area, every aspect could not be covered, and hence, a summary was provided for aspects with relatively more research. Limitations, such as the positioning, number, and possible disruption of sensors, are challenges that need to be overcome as they affect the limb’s computed trajectory. Moreover, when a magnetometer is used, ferromagnetic materials can affect its calibration, and the data may not be transmitted with accuracy. Additionally, the lack of uniformity of battery specifications (compared to sampling rates) and sensor specifications highlights the challenges in comparisons between these characteristics and standardized data in this field. Although the studies reviewed indicated a positive influence regarding the motivation of users, more clinical trials need to be conducted, as they are important to assess the effectiveness of the system. Another limitation is the search strategy employed in this review. The search was performed in only 4 specific databases: ACM Digital Library, IEEE Xplore, PubMed, and ScienceDirect. The search may have excluded relevant papers, and some studies might have been overlooked. The low number of studies analyzed might not fully capture the diversity of the current research in the field, which limits the comprehensiveness of the review.

Future studies should aim to reduce the weight and dimensions of the system and increase the sampling rate, which can facilitate quick motion tracking with high accuracy. Additionally, efforts should be made to fabricate upper limb devices that are more flexible, powerful, and compact. The progress of ML algorithms would be beneficial, particularly in IMU- and EMG-based devices, as the rehabilitation process is automatically guided in real-life settings while being able to provide remote intervention [55]. Moreover, researchers should focus on the further development of feedback as it should be more adaptable and provide more options to users and therapists. In addition, long-term clinical trials are essential for establishing the effectiveness of wearable devices in real-world rehabilitation settings. These trials should focus on larger and more diverse patient groups to better establish the system’s efficacy. Moreover, in broader health care frameworks, standardized protocols and consistent measurement methods should be developed to ensure the results are accurately compared, thus ensuring the reliability of the system. Behavioral measurements are also important, especially in training sessions. In conclusion, future research should focus on integrating sensors and improving usability and feasibility as upper limb wearable devices are predictably becoming powerful tools, enabling innovative rehabilitation treatments while improving the quality of health care.