This model (46/63, 73% studies) measures the awkward movements of the trunk for the affected UE [15]. Trunk compensatory movements can occur in 3 anatomical planes (sagittal, transverse, and coronal) of the human body. The sagittal plane (41/46, 89% studies) was the most common, which was described as trunk lean forward, trunk lean backward, trunk displacement, trunk flexion, and trunk anteriorization. The transverse plane (24/46, 52% studies) included trunk rotation and trunk twist. A total of 9 (20%) studies discussed the coronal plane, including trunk leans from side to side, trunk contralateral and ipsilateral flexion, trunk lateral bending, and trunk lateral shift (Table 3). The most common task (35/46, 76%) was the reaching task, followed by shoulder, elbow, and wrist exercises (4/46, 9%) [20,59-61], daily life activities [62-65], drinking tasks [66], simulated therapy activities [67], instrumented trunk impairment scale (version 2) tasks [21], Fugl–Meyer Assessment (FMA) items [22], occupational therapy tasks [68], and the Graded Repetitive Arm Supplementary Program (GRASP) [69], which is a set of UE exercises completed without the presence of a therapist. In 13 (28%) studies, tasks were completed using a robot-assisted device. In 4 (9%) studies, the tasks were conducted using VR [23-25] and mixed reality (MR) training systems [26].

This model (29/63, 46% studies) measures awkward movements of the shoulder of the affected UE [15], involving complex movements of the shoulder girdle and shoulder joint. The most observed shoulder compensation was shoulder elevation (17/29, 59% studies), followed by shoulder abduction [27,28,64,70,71], shoulder girdle compensatory movement [29-31,64,72], shoulder forward (protraction) [26,32], and shoulder overflexion [22,33,73]. The most commonly used task was reaching task (18/29, 62% studies). Other tasks involved hand-to-mouth tasks [33,70,73], drinking tasks [66,72], elbow flexion-extension task [59,71], daily life activities [64,65], counterclockwise cyclic motions [31], FMA items [22], and GRASP [69]. In all, 12 studies were conducted using a robot-assisted device and 1 with an MR training system [26].

This model measures awkward elbow movements of the affected UE [15]. A total of 6 studies found that stroke survivors had insufficient elbow extension during reaching tasks [27,28,74,90], drinking tasks [66], or daily life activities [65].

This model (3/63, 5% studies) measures the compensation among finger joints [34,75,76]. A study assessed the compensation among the joints in a finger in reaching tasks [34], whereas 2 other studies assessed compensation among multiple fingers in repetitive force-pulse tasks [75] and index finger movements [76].

Other types of compensation models included joint coordination [35,36], slouching [69], and muscle synergies [33], which were measured in reaching tasks, GRASP, and hand-to-mouth tasks, respectively.

No standard measurement has been applied across studies for this type of compensation. For the arm use model, Ballester et al [83] and Hung et al [85] computed the mean squared sum of the acceleration over 1 minute. Thrane et al [82] calculated the arm movement ratio, that is, the ratio of arm use duration between the impaired arm and less impaired arm. Johnson et al [84] quantified the arm use by comparing the torque generated by the tangential forces of the 2 arms on the steering wheel.

For the arm nonuse model, Bakhti et al [86] computed proximal arm nonuse, which was the difference between the Euclidean distance between the trunk and hand to the target during the reaching movement in both spontaneous and maximal proximal arm use conditions. Johnson et al [87] used 4 kinematic metrics, including movement time, peak velocity, total displacement, and movement smoothness, to predict learned nonuse (LNU). Johnson et al [88] compared the root mean square of the rotation angle of the wheel in steering tasks in 3 different steering modes (unilateral nondominant, unilateral dominant, and bilateral) to quantify LNU. Barth et al [89] computed the total movement duration of each limb and the activity ratio, which was the movement duration of the paretic limb to the nonparetic limb to assess LNU.

Miller et al [81] created an array of numerical values, including amplitude, time, and frequency data from acceleration signals on the sternum, right wrist, left wrist, right upper limb, and left upper limb to characterize the interlimb coordination model.

The parameters for measuring trunk compensation in the sagittal plane included trunk angular displacement (12/41, 29% studies), trunk linear displacement (8/41, 20% studies), trunk contribution slope [37,38,62], acceleration of trunk motion [28,64], surface electromyogram (sEMG) signals [39,77], and face orientation [67]. The parameters used to measure trunk compensation in the transverse plane included trunk angular displacement (7/24, 29% studies), acceleration of trunk motion [27,28,64], trunk linear displacement [34,40], and sEMG signal [39]. A total of 2 parameters, trunk angular displacement (6/9, 67% studies) and trunk linear displacement [34,41] were measured to assess trunk compensation in the coronal plane.

The shoulder elevation compensation measurements included the elevation angle of the scapula, acromion, or acromioclavicular joint (4/17, 24% studies); acceleration of shoulder joint motion [27,28]; shoulder vertical translated distance [37]; and sEMG signal [39]. Shoulder abduction compensation was assessed by acceleration of shoulder joint motion [27,28,64], shoulder abduction angle [70], and functional magnetic resonance imaging (fMRI) [71]. Shoulder girdle compensatory movement measurements included acceleration of shoulder joint motion [29,64], sEMG signals [30], shoulder position [31], and the coefficient of the elbow joint extension to the shoulder joint flexion ratio [72]. A total of 3 studies [22,33,73] used the shoulder flexion angle to assess the shoulder overflexion compensation. The parameter for measuring shoulder forward compensation was shoulder forward liner displacement [26,32].

The elbow extension angle [66,74,90] and acceleration of elbow joint motion [27,28] were used to measure elbow compensation.

In all, 3 kinds of measurements were used to assess finger compensation. Fluet et al [34] measured the finger extension angle to assess the individual finger compensation. Kim et al [75] measured the covariance of individual finger impulses across multiple pulses, and Furudate et al [76] measured the pressure force of fingers to assess the compensation among multiple fingers.

As for other compensation models, Reisman and Scholz [36] measured multiple parameters including joint angles (ie, scapula, shoulder, elbow, and wrist), movement time, goal-equivalent variance, and nongoal-equivalent variance to evaluate joint coordination; Nibras et al [35] measured only joint angles to assess joint coordination. Lin et al [69] captured joint positions to assess slouching. Belfatto et al [33] measured sEMG signals to assess the muscle synergy.

A total of 6 types of sensors were identified as shown in Tables 5 and 6.

Body-worn sensors (25/72, 35% studies) were the most commonly used technology to detect compensatory movements, including accelerometers, inertial measurement units (IMUs), strain sensors, and CyberGlove. In all, 9 studies used accelerometers, including uniaxial [82] and triaxial [27,28,60,64,65,83,85,89]. Accelerometers were attached to different parts of the body. Some were worn on the wrists of both arms [82,83,89] or only on the wrist of the affected arm [85] to measure arm movement quantity, such as movement duration and acceleration magnitude, to evaluate arm use. Some were placed on the trunk (chest, middle back, or T12 vertebrae) [27,28,60,64], shoulder [27,28,64], elbow [27,28], and wrist [65] to measure time and movement variables, such as accelerations and joint angles of UE movement to detect trunk, shoulder, and elbow compensation. Among these studies, Antonio et al [27] and Carla et al [28] compared quantitative detection results using accelerometers with therapist-based visual analysis of video records. The results showed that the compensatory movements detected by the accelerometers, including shoulder abduction and elevation, insufficient elbow extension, and trunk forward displacement and rotation, were consistent with the therapists’ observations.

In all, 13 used IMUs. Each IMU typically consists of 1 or 2 triaxial accelerometers, a triaxial gyroscope, and a triaxial magnetometer [21,22,25,31,42,59,68,70,73,81]. The magnetometer was not included in some cases [43,44,90]. Accordingly, each IMU yielded 3D measurements of acceleration, angular velocity, and magnetic field vector (when using a magnetometer) in its intrinsic coordinate system [59]. In the reviewed studies, 1 to 9 IMUs were placed on the upper body parts, including the sternum [21,25,43,44,68,81,90], spine [21,22], pelvis [22], scapula [70,90], upper arms [22,25,31,42-44,59,70,73,81,90], forearms [22,25,43,44,59,70,90], wrists [73,81], and hands [22,70,90]. The original IMU signals [43,44,81] representing the movements of these body segments or the orientation in the form of Euler angles [21,22,25,31,42,59,68,70,73,90], of these body segments were output for compensation monitoring. It has been reported that IMUs can be used to detect trunk [21,22,25,42-44,59,68,90], shoulder [22,31,59,70,73], and elbow [22] compensation, as well as the interlimb coordination [81]. Furthermore, Ranganathan et al [43] proved that using IMUs could effectively detect compensatory trunk movements (approximately 90% accuracy) when compared with using an 8-camera motion capture system (Motion Analysis Corporation) as ground truth.

Moreover, 2 studies used changes in the electrical resistance of strain sensors printed on a garment [78,79] to identify different compensatory postures during UE movements. A study used CyberGlove to assess finger compensation by measuring the angles of finger joints [34]. Overall, 4 studies were conducted using robot-assisted therapies [42,59,70,87], 1 [25] using VR therapy, and 10 were conducted in homes [25,43,44,60,64,65,78,79,81,83].

The second most commonly used technology was the marker-based motion capture system (24/72, 33% studies). A total of 3 types of marker-based motion capture systems were used: an optical motion capture system (22/24, 92% studies), electromagnetic motion capture system [34], and ultrasound 3D motion capture system [86]. For this sensor, markers were attached to the upper body landmarks, which traditionally included the sternum, spinal process (C7 and T4), acromion processes, middle part of the humeri, lateral epicondyle, styloid process of the ulna, and bilateral thumbnails [62,63,87]. The participants were asked to perform the tasks while the positions of the markers were captured. The position and orientation of the trunk, shoulder, and elbow were then calculated according to the joint coordinate system method and used to characterize different compensation models.

For a long time, marker-based motion capture systems have been used as gold standard motion capture devices for clinical motion analysis [86]. Similarly, in the reviewed studies, marker-based motion capture systems were proven to be able to effectively identify compensation. In all, 5 studies have been used as the ground truth for the measurement of the effectiveness of other sensors in compensation detection [42,43,45,86,87]. Several interesting findings were reported using marker-based motion capture systems: (1) pre- and posttests showed that both robotic [32,38,63] and MR therapies [26] elicited benefits on reducing trunk compensatory movements, and stroke survivors showed less trunk compensatory movements during VR reaching [24]. However, Belfatto et al [33] argued that robotic therapy promoted the adoption of compensatory movements when stroke survivors performed training tasks; (2) therapist-based therapy [62] or a combination of robotic therapy and constraint-induced therapy [46] demonstrated more significant improvements in reducing trunk compensatory movements compared with robot-assisted therapy; (3) trunk displacement and shoulder elevation compensatory movements could discriminate between mild and moderate stroke paresis [66], whereas shoulder girdle compensatory movement could differentiate between mild or moderate and severe or pronounced stroke impairments [72].

Among all the studies, 13 [31-33,38,39,42,45-47,59,62,63,87] monitored compensation with robot-assisted upper limb devices, one [24] was conducted in VR therapy and one [26] was in MR therapy.

A total of 16 studies used marker-free vision sensors, including Microsoft Kinect depth sensors (versions 1 and 2) and a simple simple camera, as motion capture tools. Most (11/16, 69%) of these studies used Kinect, which is usually placed approximately 2.0 meters in front of the user to capture the 3D space (x-y-z) coordinates of 20 (version 1) or 25 (version 2) skeleton joint positions in the user’s body at 30 frames per second. In the reviewed studies, the locations and orientations of the upper body parts (ie, hip, spine, shoulder, elbow, wrist, and hand) [48-50,61,69,86] or spine [20,23,41,51,52] were recorded, and 2 studies have verified the effectiveness of this sensor technology for monitoring compensation. Bakhti et al [86] proved that Kinect can be used to accurately assess proximal arm nonuse when compared with an ultrasound 3D motion capture system (CMS20s, Zebris). The agreement between Kinect and CMS20s was measured using intraclass correlation coefficients (0.96), linear regression (r²=0.92), and Bland and Altman plots (Kinect: −4.25, +6.76 to –6.76); CMS20s: −4.71, +7.88 to –7.88). Lin et al [69] found substantial agreement of detected compensation, such as shoulder elevation and hip extension, between annotated videos and Kinect (Cohen κ 0.60-0.80) and almost perfect agreement for trunk rotation and flexion (Cohen κ 0.80-1).

Overall, 6 studies used RGB cameras and 2 (33%) of them [67,74] used a camera to collect motion images to extract kinematic data through third-party data extraction algorithms for quantitative compensation assessment. The other 4 (67%) studies collected motion videos for clinicians’ visual evaluation of compensation.

In all, 8 studies [41,48-52,74,84] were conducted using robot-assisted upper limb devices, 2 studies [23,69] were conducted using VR therapy, and 1 study [67] was conducted in a home using a single low-cost camera.

Physiological signal sensing technologies include electromyogram (8/72, 11% studies), electroencephalogram (EEG) [33,80], and fMRI [71] systems. According to the reviewed studies, sEMG signals of upper limb muscles (including, but not limited to, biceps, triceps, upper trapezius, pectoralis major, brachioradialis, anterior, middle, and posterior deltoids) and trunk muscles (left or right rectus abdominis, left or right obliquus externus abdominis, left or right thoracic erector spinae, left or right lumbar erector spinae, and descending part of the trapezius) not only helped to discriminate true recovery and compensation [29,30,33,84] but also could be used as features for automatic compensation detection [39,53,77]. Chen et al [77] confirmed that using a generative adversarial network with sEMG signals as features could achieve excellent detection performance (accuracy=94.58%, +1.15% to –1.15%) of trunk compensatory movements.

A study used fMRI [71] to study the cortical activation pattern of compensatory movements and demonstrated that compensatory movements require a greater recruitment of cortical neurons. A total of 2 studies [33,80] showed that brain scalp EEG signals could help researchers gain more insight into the relationship between motor compensation and underlying brain activities. Among all studies, electromyogram [29,30,33,39,54,84] and EEG [33] systems were used along with robot-assisted devices for compensation assessment.

In all, 8 studies directly selected sensors embedded in the rehabilitation training system as compensation evaluation tools. Nibras et al [35] used the measurement information in an exoskeleton to distinguish between recovery and compensation in stroke survivors. They found 2 compensatory patterns in stroke survivors: atypical decoupling of the shoulder elevation and forearm joints and atypical coupling of the shoulder horizontal rotation and elbow joints, by analyzing 4 ArmeoSpring angles when stroke survivors performed reaching movements with the ArmeoSpring exoskeleton. In contrast, a simpler and less complex UE rehabilitation robot, such as an end-effector robot, may not have the capacity to provide detailed UE measurement information as the exoskeleton. Therefore, additional sensors, such as inertial sensors [31,42,59], are required with the sensors in the end-effector robot to satisfy compensation assessment needs. In addition, Johnson et al [84,88] used the force sensors of a UE rehabilitation system, a driver simulation system, to quantify impaired arm activity [84] and LNU [88]. Kim et al [75] and Furudate et al [76] used force sensors in hand rehabilitation systems to evaluate the compensation among individual fingers.

A total of 5 studies used ambient sensors, including a pressure distribution mattress (Body Pressure Measurement System, Model 5330, Tekscan, Inc) [45,54,91] and position measurement sensors [55,56]. A pressure distribution mattress was used to measure a person’s body pressure distribution in a seated position for the automatic detection of compensatory postures [45,54,91]. Cai et al [45] verified the effectiveness of using pressure distribution data together with machine learning (ML) algorithms to detect compensatory patterns using a 3D motion capture system (VICON, Oxford Metrics) as the ground truth. When using a pressure mattress or VICON, the average F₁ score (an evaluator of the ML algorithm performance) was >0.95.

The position measurement sensors used were either a force sensor placed anterior to the back of the chair [56] or a contactless first-reflection ultrasonic echolocation sensor placed on the edge of a table [55] to monitor the position of the trunk. As only the trunk position was monitored, the researchers only realized a rough detection of compensatory trunk flexion movement. In addition, 3 studies [45,54,91] used robot-assisted upper limb devices, and 4 studies have proposed that these systems could be used in a home environment [45,54,55,91].

Overall, 56 studies used statistical methods and 15 adopted AI-based methods as shown in Table 7 and Table 8, respectively.

Statistical methods were used to assess compensation from 4 perspectives: real time changes of compensation measurements in body movements, group variance in compensation measurements, effects of an intervention on compensation measurements, and the statistically significant associations of physiological signals with compensation measurements.

A total of 23 studies used mean and SD, canonical correlation analysis, Spearman correlation, step-wise multiple regression, or ANOVA to test the real time changes of compensation measurements in body movements. For instance, Wittmann et al [25] used repeated measures 1-way ANOVAs to test trunk orientation changes during rehabilitation training to assess trunk compensation in real time.

In all, 20 studies tested the differences among groups to assess compensation. The most commonly used statistical methods were ANOVA and Mann–Whitney U test. For instance, Kim et al [75] compared all compensation measurements between groups (stroke survivors vs healthy participants) and between hands (within-group factor: more affected hand vs less affected hand in stroke survivors and nondominant hand vs dominant hand in healthy participants) with ANOVA for compensation assessment.

In addition, 10 studies analyzed the differences in compensation measurements before and after the intervention. Wilcoxon signed-rank test, ANOVA, Spearman rank correlation coefficient, paired 1- and 2-tailed t test and 1- and 2-tailed Tukey honestly significant difference tests were used in these studies. For instance, Fluet et al [34] used ANOVA to analyze how 2 different training models (traditional vs VR-based training) affect upper limb compensation in the dimensions of peak reaching velocity, finger extension excursion, shoulder excursion, elbow excursion, and trunk excursion.

Overall, 3 studies tested the associations between physiological signals, such as fMRI and sEMG, and compensation parameters [30,39,71]. For instance, Lee et al [71] used Spearman rank correlation coefficient to test the relationship between the brain activation area and shoulder abduction angle. They found that greater activation of the supplementary motor area was required for a larger shoulder abduction angle. Huang et al [39] applied the Pearson correlation test and found a positive correlation between muscle fatigue (measured by sEMG median frequency) and compensation. They concluded that sEMG median frequency was a good indicator of compensation due to muscle fatigue.

A total of 15 studies used AI-based methods to detect compensatory postures, and 9 studies classified 3 common compensatory postures: trunk lean forward, trunk rotation, and shoulder elevation [45,48-50,53,54,74,77,91]. In addition, 4 studies discriminated trunk compensatory movements in the sagittal, transverse, and coronal planes [43,44,55,61], and 2 studies did not mention the type of compensatory posture that was classified [78,79]. Dolatabadi et al [50] made the compensation data set public for other researchers. A total of 2 studies used this data set to train their ML models to improve the accuracy of compensation detection [48,49]. The remaining studies collected their own data to detect compensation.

Various ML algorithms were applied to train the classification model (Table 6). The most commonly used ML algorithm was the support vector machine (SVM). Cai et al [45] reported the highest average F₁ score (0.99) for recognizing trunk lean forward, trunk rotation, and shoulder elevation based on 5 features extracted from the pressure distribution data. Nordin et al [61] reported the highest accuracy (99.15%) for detecting the 3D trunk compensatory postures.

Notably, 8 studies [44,48,49,54,61,74,77,79] used more than one ML algorithm to compare the classification results for compensatory postures. For example, Zhi et al [48] used both SVM and recurrent neural network classifiers to classify shoulder elevation, trunk rotation, and lean forward. The results demonstrated high accuracy in healthy participants, but low accuracy in stroke survivors. Cai et al [54] applied the k-nearest neighbor and SVM algorithms to detect and categorize shoulder elevation, trunk rotation, and lean forward in stroke survivors, and both algorithms yielded high classification accuracies (F₁ score >0.95). Nordin et al [61] used 4 different classification algorithms with 10-fold cross-validation to assess the 3D trunk compensatory movements. The results showed accuracy of 99%, 98%, 96%, and 83% with SVM, k-nearest neighbor, random forest, and logistic regression, respectively.