Upper limb assessment with inertial measurement units according to the international classification of functioning in stroke: a systematic review and correlation meta-analysis

Figures & data

Table 1. Selected domains of the World Health Organization’s International Classification of Functioning Disability and Health (ICF).

Subheading	ICF code	Domain	Definition	Classification criterion
Body functions (b)	b710	Mobility of joint functions	Functions of the range and ease of movement of a joint.	IMUs used to assess range of motion.
	b735	Muscle tone functions	Functions related to the tension present in the resting muscles and the resistance offered when trying to move the muscles passively.	IMUs used to assess spasticity.
	b760	Control of voluntary movement functions	Functions associated with control over and coordination of voluntary movements.	IMUs used to assess voluntary movements with or without object manipulation in a non-finalized task (i.e., Mingazzini test, finger-to-noise task).
Activities and participation (d)	d445	Hand and arm use	Performing the coordinated actions required to move objects or to manipulate them by using hands and arms, such as when turning door handles or throwing or catching an object.	IMUs used to assess daily activity task (i.e., making a cup of tea, turn a key).

Figure 1. Schematic representation of spatio-temporal and kinematic parameters recorded through inertial measurement units. The investigated spatio-temporal and kinematics variables are reported into the white boxes.

Table 2. Synoptic table of studies included in the passive range of motion assessment (b710) of the ICF domain.

Author	Participants, (gender), age	IMUs number (N) and location	“IMUs device,” Frequency acquisition	Source signals (+ processing technique)	Procedure	Results
Bai 2020 (1)^Citation29	4 (3 M, 1F), 75.5 ± 2.5	(2) Humerus, Forearm	(a) “Xsens MTx, Xsens Technologies, the Netherlands,” NR (b) “Sony Move,” 30 Hz	(a) 3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm) (b) 3D sensor orientation as estimated by a 9-axis. (complementary filter/Kalman filter)	Abduction test	IMUs are shown to be able to accurately measure upper limb joint orientation and position.
Bai 2020 (2)^Citation30	5 (4 M, 1F), 68.8 ± 8.7	(a) (4) Shoulder, Humerus, Forearm, Hand (b) (2) Humerus, Forearm	(a) “Xsens MTx, Xsens Technologies, the Netherlands,” NR (b) “Sony Move,” 30 Hz	(a) 3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm) (b) 3D sensor orientation as estimated by a 9-axis. (complementary filter/Kalman filter)	Shoulder f/e, shoulder abd/add, elbow f/e, forearm pronation/supination, wrist f/e, hand ulnar/radial deviation.	Low cost IMUs provide adequate accuracy in measurement UL orientation and position tracking.
Bertomeu-Motos 2018^Citation31	3 (NR) 61 ± 9.6	(2) Shoulder, Upper arm	“ShimmerSensing, Dublin, Ireland,” 100 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	The wrist joint location has been computed during the exercise with a 7-DoF arm model fed with the segments orientation as provided by two magneto-inertial sensors and the wrist position as provided by an end-effector robot.	Higher estimated ROM’s joints correlate with high FMA scores. The proposed algorithm can determine the ROM using only one accelerometer attached in the upper arm but reveals an instability when shoulder movements appear due to the inevitable trunk compensation in stroke patients.
Bhagubai 2020^Citation32	10 (6 M, 4F), 61.5	(7) Shoulder, Humerus, Forearm, Hand, Thumb, 2nd finger, 3rd finger	“Custom IMU based on ST LSM330DLC chip,” 100 Hz	3D sensor orientation as estimated by a 6-axis. (gradient descent sensor fusion algorithm)	Arm-related tasks (A1, A2, A3, B1) from the FMA, with the affected and non-affected side.	The measurement system was adequately sensitive to show significant differences in stroke subjects’ arm postures between the affected and non-affected side. The presence of pathological synergies can be analyzed using the measured joint angles.
Kang 2020^Citation33	18 (10 M, 8F), 49.78 ± 15.55	(6) Head, Sternum, Humerus BS, Forearm BS.	“Custom magneto-inertial measurement unit, MPU9250; InvenSense, San Jose, CA,” 100 Hz	3D sensor orientation as estimated by a 9-axis. (gradient descent sensor fusion algorithm)	ROM, movement time and variability was recorded during three phases of shoulder movement (abduction, holding and adduction). Parameter extraction was performed on the Euler angles obtained from the frontal plane around the sagittal axis.	During the holding phase, MAC significantly increased the minimum Euler angle and decreased the ROM compared with the other types of cueing. Further, the root mean square error in the angle measurements was significantly smaller and the duration of movement execution was significantly shorter during the holding phase when MAC was provided than when the other types of cueing were used.
Li 2016^Citation34	35 (NR), NR	(4) Head, Upper arm, Forearm, Hand.	“Custom magneto-inertial measurement unit, MPU9150; InvenSense, San Jose, CA,” NR	3D sensor orientation as estimated by a 9-axis. (Kalman filter sensor fusion algorithm)	shoulder flexion (90°) in a sitting position and held on for 2–3 seconds.	All kinematics showed a significant statistical difference between patients and healthy people, while the feature values showed a high correlation with FMA scores.

Abbreviations: NR = Not Reported; ROM = Range Of Motion; UL = Upper Limb; BS = Both Side; f/e = flexion/extension; abd/add = abduction/adduction; FMA = Fugl-Meyer Assessment; ARAT = Action Research Arm Test; ADL = Activities of Daily Living; IMU = Inertial Measurement Unit; MAC = Melodic Auditory Cueing. In Bai et al., 2020 (1) and (2), only configuration (b) was used to performed kinematic analysis. Column five reports the source signal from which the kinematic variable of interest has been derived: 6-axis = the sensor fusion algorithm runs using accelerometer and gyroscope data; 9-axis = the sensor fusion algorithm runs using accelerometer, gyroscope, and magnetometer data.

Table 3. Synoptic table of studies included in the muscle tone functions (b735) of the ICF domain.

Author	Participants, (gender), age	IMUs number (N) and location	“IMUs device,” Frequency acquisition	Source signal (+ processing technique)	Procedure	Results
Ang 2018^Citation54	15 (7 M, 8F), 56.9 ± 10	(3) Thorax, Humerus, Forearm	“APDM Opal™ wireless,” 128 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	Joint torques have been computed using inverse dynamics with measurements from three IMUs to calculate the tonic stretch reflex threshold. Data have been subjected to correlation analysis with the EMG and MAS.	The estimated muscle activation profiles have a high correlation to the EMG signal profiles. Spasticity severity calculated with IMUs showed a high correlation with the MAS score.
Chen 2021^Citation55 (1)	9 (8 M, 1F) 51.8 ± 12.4	(1) Forearm	“Custom magneto-inertial measurement unit, MPU9250; InvenSense, San Jose, CA, 200 Hz	3D sensor orientation as estimated by a 9-axis. (complementary filter sensor fusion algorithm)	f/e of the elbow supported by a guide-track has been performed. A slider was used to assist patients to achieve the maximum ROM.	The RF algorithms exhibited excellent classification performance in detecting and categorizing four grades of spasticity.
Kim 2020^Citation53	48 (26 M, 22F), 61.2 ± 13.7 (M) 77.8 ± 10.1 (F)*	(1) Wrist	“Shimmer Sensing, Dublin, Ireland,” 256 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	Results of an IMU device during a passive stretch test with Machine-learning algorithms (eg. RFs) have been combined. Measurement of elbow spasticity was based on the MAS.	Among the Machine-learning algorithms, a RF performed well, achieving up to 95,4% accuracy.
Paulis 2011^Citation56	13 (7 M, 6F), 70.2 ± 12.3	(2) Humerus, Wrist	“MTx, Xsens Technologies, Enschede, Netherlands,” 100 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	Tardieu Scale measurements have been performed in two sessions, using both IMUs and goniometry, to quantify spasticity in elbow flexors of stroke patients.	For goniometry, test – retest and inter-rater reliability proved to be excellent and fair to good, respectively. For IMUs, both test – retest and inter-rater reliability were excellent. IMUs are reliable and accurate to use in Tardieu Scale measurements to quantify spasticity in the elbow flexors of hemiplegic stroke patients.

Abbreviations: IMU = Inertial Measurement Unit; MAS = Modified Ashworth Scale; EMG = Electromyography; f/e = flexion/extension; RF = Random Forest.×45 stroke and 3 spinal cord injuries. Column five reports the source signal from which the kinematic variable of interest has been derived: 6-axis = the sensor fusion algorithm runs using accelerometer and gyroscope data; 9-axis = the sensor fusion algorithm runs using accelerometer, gyroscope, and magnetometer data.

Table 4. Synoptic table of studies included in the voluntary movements assessment (b760) of the ICF domain.

Author	Participants, (gender), age	IMUs number (N) and location	“IMUs device,” Frequency acquisition	Source signal (+ processing technique)	Procedure	Results
Bai 2020 (1)^Citation29	4 (3 M, 1F), 75.5 ± 2.5	(2) Humerus, Forearm	(a) “Xsens MTx, Xsens Technologies, the Netherlands,” NR (b) “Sony Move,” 30 Hz	(a) 3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm) (b) 3D sensor orientation as estimated by a 9-axis. (complementary filter/Kalman filter)	BBT and the NHPT have been performed.	IMUs are shown to be able to accurately measure upper limb joint orientation and position.
Bai 2020 (2)^Citation30	5 (4 M, 1F), 68.8 ± 8.7	(a) (4) Shoulder, Humerus, Forearm, Hand (b) (2) Humerus, Forearm	(a) “Xsens MTx, Xsens Technologies, the Netherlands,” NR (b) “Sony Move,” 30 Hz	(a) 3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm) (b) 3D sensor orientation as estimated by a 9-axis. (complementary filter/Kalman filter)	BBT and the NHPT have been performed.	Low cost IMUs provide adequate accuracy in measurement UL orientation and position tracking.
Chen 2021^Citation57 (2)	37 (28 M, 9F) 49.8 ± 10.3	(4) Head, Upper Arm, Forearm, Hand	“IMU, Noraxon, USA Inc.,” 100 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	Finger-To-Nose task. UL motor function evaluated with FMA, ARAT and MBI.	Results showed that Finger-To-Nose task kinematic variables measured via IMU were associated with UL motor function.
Held 2018^Citation37	4 (NR), 48–55 range	(14) Shoulder BS, Humerus BS, Sternum, Sacrum, Feet BS, Lower leg BS, Upper leg BS.	“Xsens Technologies, Enschede, Netherlands,” 20 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	Kinematic data Have been recorded with a full body motion capture suit during clinical assessment (including FMA and ARAT).	These results demonstrate the feasibility of the method to meas- ure upper-limb kinematics, with an IMU-based motion capture system at different stages of stroke rehabilitation and during ADL and the concordance to standard clinical assessment.
Hesam-Shariati 2019^Citation51	24* (16 M, 8F), 57.9 ± 12.1	(6) Trapezius, Upper arm BS, Forearm BS, Hand	“Triaxial accelerometer Trigno, Delsys, USA,” 148,15 Hz	gravity-free 3D linear acceleration and jerk as estimated from high-pass filtered measured sensor-embedded accelerations, as well as linear velocity as computed from numerical integration gravity-free 3D linear acceleration	Kinematic parameters have been investigated during Wii-baseball swing.	Increased acceleration magnitude and decreased normalized velocity during a complex movement.
Nikmaram 2019^Citation58	41(12)** (30 M, 11F), 67,6 ± 11,4	(2) Upper arm, Wrist	“Xsens MTx, Xsens Technologies, the Netherlands,” 200 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	In the experiment the patients sit in front of a wooden frame labeled with musical note-pitch. The final goal was to teach them to play several simple nursery rhymes only by moving their affected arm in 3D sonification space.	IMUs data showed slight improvements in movement smoothness.
Nie 2021^Citation59	8 (6 M, 2F) 58 ± 12.6	(2) Upper Arm, Forearm	“Trigno IM Sensor, Delsys Inc.,” 200 Hz	3D sensor orientation as estimated by a 9-axis. (improved explicit complementary filter)	3D reaching movements	Compared to a traditional optical tracking system, IMUs accurately tracked the wrist movement during reaching.
Pan 2021^Citation60	34 (26 M, 8F) 59.8 ± 11.2	(4) Middle of Waist, Upper Arm, Forearm, Hand	“Custom magneto-inertial measurement unit, MPU9150; InvenSense, San Jose, CA,” 50 Hz	3D sensor orientation as estimated by a 9-axis. (particle filter)	IMU data and EMG signals have been collected from the UL during voluntary upward reaching. Five features have been assessed: max shoulder joint angle, peak and average speeds, and torso balance calculated from The FMA score of each patient.	Statistically significant differences were observed among severe, mild-to-moderate, and the control group. The features varied as the level of UL motor function changed since these features significantly correlated with the FMA. Moreover, the Bland – Altman method showed high consistency between the evaluation method of five features and the FMA scale.
Park 2020^Citation52	15*** (10 M, 5F) 68.6 ± 16.1	(4) Wrist BS, Lower Leg BS	NR	measured 3D linear acceleration	Mingazzini test upper limbs for 20 seconds.	The automatic grading system quantified proximal weakness in real time and assessed symptoms through automatic grading.
Rau 2013^Citation61	8 (5 M, 3F), 62.9 ± 13.8	(2) Acromion, Forearm	“Triaxial accelerometer (sensor chip is not reported) NR	NR	Five repetitions of forward-reaching movements.	High correlation has been found in reaching displacement, velocity, and acceleration measurements obtained using the tele-assessment system and the standardized kinematic system. Differences in the maximum reaching distance and the maximum reaching velocity of forward reaching movements were observed among the study groups.
Salazar 2014^Citation62	4 (2 M, 2F), 53.2 ± 6.5	(3) Back (T12), Acromion, Elbow	“Custom triaxial accelerometer (ADXL345), Analog Devices, Norwood, MA” 100 Hz	limb 3D rotation as estimated from measured tilt angles as derived from accelerometer signal (gravity-based)	Trunk, upper arm, and forearm angular displacements during a reach-press-return task.	Preliminary studies revealed acceleration profiles of stroke patients through which it is possible to quantitatively assess the functional movement, identify compensatory strategies, and help define proper movement.
Schwartz 2021^Citation44	26 (17 M, 9F) 62.2 ± 12.1	(9) Spina Scapulae BS, Sternum, Upper arm BS, Forearm BS, Wrist BS	“XSens MVN Awinda, Xsens Technologies, the Netherlands,” 60 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	Four movements with both ULs: (1) isolated shoulder flexion, (2) pointing ahead, (3) reach-to-grasp a glass, and (4) key insertion. The validity of metrics compared to clinically measured interjoint coordination (FMA) has been done by correlation analysis.	The movement task and the tested arm showed significant effects on all kinematic parameters. Hand dominance resulted in significant effects on shoulder f/e and curve efficiency. Relations with the FMA revealed the strongest and significant correlation for curve efficiency, followed by shoulder f/e, elbow f/e, and shoulder abd/add. Curve efficiency additionally correlated significantly with the arm subsection, focusing on synergistic control.
Tedim-Cruz 2014^Citation63	5 (5 M), 35–73 range	(4) Acromion, Humerus, Forearm, Wrist	“Custom magneto-inertial measurement unit (sensor chip is not reported)” 50 Hz	3D sensor orientation as estimated by a 9-axis. (non-specified sensor fusion algorithm)	Patients have been tested on BS across five selected tasks of WMFT: forearm-to-table (task 1), forearm-to-box (task 2), extend-elbow (task 3), hand-to-table (task 4) and hand-to-box (task 5). Each task was evaluated according to performance time (seconds) and Func- tional Ability Score (FAS) based on joint kinematics.	The prototype tested was able to automatically classify UL movement, according to WMFT, in a clinical setting.
Wade 2014^Citation48	10 (8 M, 2F), 59.2 ± 15.3	(2) Wrist BS	“Triaxial accelerometers (Gulf Coast Data Concepts, LLC).” 40 Hz	measured 3D linear acceleration	Unimanual goal-directed movements using both hands, with and without task objects.	Main effects of the hand on all 3 temporal measures and main effects of objects on movement duration and peaks have been found. Spectral measures of accelerometry data are sensitive to performance differences with the nonparetic and paretic limbs and with the object present and absent, for ADL-inspired tasks.
Zollo 2011^Citation64	24 (15 M, 9F), 55.9 ± 15.4	(1) Forearm	“Xsens MTx, Xsens Technologies, the Netherlands” NR	Smoothness (i.e., jerk) as computed from the absolute (gravity-free) linear acceleration derived from the 3D sensor orientation (proprietary sensor fusion algorithm)	Point-to-point movements with robot-assisted movement.	Significant correlation has been found between kinematic variables and FMA and motor strength scores during unperturbed and resistive robot-assisted tasks.

Abbreviations: IMU = Inertial Measurement Unit; NR = Not Reported; BS = Both Side; BBT =Bean Bag Test; NHPT = Nine Hole Peg Test; FAS = Functional Ability Score; UL = Upper Limb; f/e = flexion/extension; abd/add = abduction/adduction; ARAT = Action Research Arm Test; FMA = Fugl-Meyer Assessment; MBI = Modified Barthel Index; EMG = Electromyography; WMFT = Wolf Motor Function Test; MARG = Magnetic, Angular Rate, and Gravity sensor.

*13 at follow-up; **12 patients analyzed, ***13 stroke patients, 1 myelitis, and 1 myasthenia gravis. In Bai et al., 2020 (1) and (2), only configuration (b) was used to performed kinematic analysis. Column five reports the source signal from which the kinematic variable of interest has been derived: 6-axis = the sensor fusion algorithm runs using accelerometer and gyroscope data; 9-axis = the sensor fusion algorithm runs using accelerometer, gyroscope, and magnetometer data.

Table 5. Synoptic table of studies included in the object manipulation and daily life activity assessment (d445) of the ICF domain.

Author	Participants, (gender), age	IMUs number (N) and location	“IMUs device,” Frequency acquisition	Source signal (+ processing technique)	Procedure	Results
Bai 2020 (1)^Citation30	5 (4 M, 1F), 68.8 ± 8.7	(a) (4) Shoulder, Humerus, Forearm, Hand (b) (2) Humerus, Forearm	(a) “Xsens MTx, Xsens Technologies, the Netherlands,” NR (b) “Sony Move,” 30 Hz	(a) 3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm) (b) 3D sensor orientation as estimated by a 9-axis. (complementary filter/Kalman filter)	Patient performed a “drinking” task.	This study demonstrates that multi-sensor inertial sensing systems can provide additional insights for motion quantification.
Biswas 2014^Citation35	4 (NR but both sexes), 45–73 range	(1), Wrist	“Shimmer, Dublin, Ireland,” 50 Hz	Limb posture as estimated from measured tilt angles as derived from accelerometer signal (gravity-based)	Activities involved in a representative ADL: “making a cup of tea.”	The results showed that the IMU can independently recognize all three of the elementary UL movements investigated (reach and retrieve, lift cup to mouth, rotation of the arm) with accuracy in the range 91%-99% for healthy subjects and 70%-85% for stroke patients.
Bochniewicz 2017^Citation36	10 (8 M, 2F), 56 ± 10.4	(2), Wrist-worn BS	“Custom magneto-inertial measurement unit (ADIS16400BMLZ, Analog Devices, Norwood, MA).” 200 Hz	Limb measured 3D linear acceleration and angular velocity	Series of ADL tasks (doing the laundry, kitchen activities, shopping and making the bed).	A significant correlation between ARAT scores of the stroke’s patients and the percentage of time spent in functional use has been found.
Held 2018^Citation37	4 (NR), 48–55 range	(14) Shoulder BS, Humerus BS, Sternum, Sacrum, Feet BS, Lower leg BS, Upper leg BS	“Xsens Technologies, Enschede, Netherlands,” 20 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	FMA, ARAT, and self-directed ADL.	A change in data kinematics from the daily-life recording was seen in all patients, increasing the number of reaches performed during daily life in their home environment.
Lin 2017^Citation38	15 (9 M, 6F), 59.3 ± 16.3	(16) Hand, Fingers (one for each side of the interphalangeal joint)	“Custom IMU (LSM330DLC, STMicroelectronics, Geneva, Switzerland),” NR	3D sensor orientation as estimated by a 6-axis. (gradient descent algorithm)	Grip task, thumb task, and card turning task.	The IMUs glove provides quantitative data useful for medical treatments.
Mahmoud 2021^Citation39	9 (NR) 58.2 ± 4.8	(2) Forearm, Hand	“Xsens DOT, Xsens Technologies, the Netherlands,”800 Hz	3D sensor orientation as estimated by a 9-axis. (proprietary sensor fusion algorithm)	To simulate the daily activity of moving objects to a high level. To simulate the action of cleaning the window. To control action 1, the actual main investigation of the patient’s arm in the state of low-load movement. To simulate the situation of daily horizontal movement of objects.	The results showed a high overall accuracy. In addition, the fusing of Kinect and Xsens data revealed that muscle strength was highly correlated with the kinematics data.
Mazomenos 2016^Citation40	4 (NR -mixed gender-), 45–73 range	(2) Elbow, Wrist	“Shimmer, Dublin, Ireland,” 50 Hz	3D sensor orientation as estimated by a 9-axis. (gradient descent algorithm)	3 elementary UL movements: reach and retrieve, bend the arm at the elbow, rotation of the arm about the long axis, such as subtasks of an ADL (“making a cup of tea”).	The results showed a high overall accuracy for all three movements for the “preparation of a cup of tea” task.
Michaelsen 2013^Citation41	8 (4 M, 4F), 59.4 ± 6.9	(1) Forearm	“Triaxial accelerometer, EMG Systems, Brazil,” 100 Hz	measured 3D linear acceleration	Ten reach-to-grasp movements of grabbing a 500 ml-size bottle.	The IMU allowed identification of changes in reaching-to-grasp movements of subjects with hemiparesis. Movements were slower in the paretic UL with increased movement time, time to peak acceleration and deceleration duration.
Repnik 2018^Citation42	28 (18 M, 10F), 57 ± 9.1	(7) Sternum, Humerus BS, Forearm BS, Hand BS	“Magneto-inertial measurement unit (brand/chip is not reported),”200 Hz	3D sensor orientation as estimated by a 9-axis. (unscented Kalman filter)	ARAT.	Movement quantification enables differentiation between different subject groups within movement phases as well as for the complete task. Strong correlations between ARAT scores and movement time as well as movement smoothness was found. Weak to moderate correlations were observed for parameters that describe hand trajectory similarity and trunk stability.
Schwarz 2020^Citation43	10 (6 M, 4F), 60.8 ± 11.4	(8) Sternum, Shoulder, Upper arm, Lower arm, Hand, Thumb, 2° and 3° fingers	“Custom IMU (ST LSM330DLC, ST Microelectronics, Geneva, Switzerland,” 100 Hz (accelerometer) and 200 Hz (Gyroscope)	3D sensor orientation as estimated by a 6-axis. (not specified sensor fusion algorithm)	Reach-to-grasp activities with the affected and non-affected UL. It was investigated whether the factors, tested arm, object weight, and target height, affect the expressions of ROM in trunk compensation and f/e of the elbow, wrist, and finger during object displacement.	The tested arm and target height showed strong effects on all metrics, while an increased object weight showed effects on trunk compensation. High inter- and intrasubject variability was found in all metrics without clear relationships to clinical measures. Relating all metrics to each other resulted in significant negative correlations between trunk compensation and elbow f/e in the affected arm.
Schwartz 2021^Citation44	26 (17 M, 9F) 62.2 ± 12.1	(9) Spina Scapulae BS Sternum Upper arm BS Forearm BS Wrist BS	“XSens MVN Awinda, Xsens Technologies, the Netherlands,” 60 Hz	3D sensor orientation and position as estimated by a 9-axis. (proprietary sensor fusion algorithm with a kinematic model)	Four movements with both ULs: (1) isolated shoulder flexion, (2) pointing ahead, (3) reach-to-grasp a glass, and (4) key insertion.	The movement task and the tested arm showed significant effects on all kinematic parameters. Hand dominance resulted in significant effects on shoulder f/e and curve efficiency. The level of UL function showed influences on curve efficiency and the factor age on the median slope. Relations with the FMA revealed the strongest and significant correlation for curve efficiency.
Thies 2009^Citation45	6 (4 M, 2F), 59.1 ± 16	(1) Forearm	“Xsens Technologies B.V., Enschede, Netherlands (sensor model not reported),” 4 Hz (cutoff)	measured 3D linear acceleration	ADL (drinking from a glass and moving a plate). These activities involved movements such as a forward reach, followed by hand opening, hand grasp, object manipulation and finally object release and arm retraction.	For “drinking from a glass” significant group differences were obtained on both days for the timing variability of the acceleration signals’ characteristics; all stroke patients showed increased signal timing variability as compared to their corresponding control subjects. “Moving a plate” provided less distinct group differences.
Van Meulen 2015^Citation46	13 (8 M, 5F), 63.8 ± 8.5	(17) Head, Sternum, Sacrum, Shoulder BS, Upper arm BS, Lower arm BS, Hand BS, Upper leg BS, Lower leg BS, Feet BS	“Xsens MVN Awinda, Xsens Technologies, the Netherlands,” 120 Hz	3D sensor orientation and position as estimated by a 9-axis. (proprietary sensor fusion algorithm with a kinematic model)	Multiple ADL. Each task was repeated three times.	Significant correlations with the FMA scores were found in the hand movements (working area, maximum reached distance, and the range in vertical hand elevation.
Van Meulen 2016^Citation47	2 (NR), NR	(12) Head, Sternum, Pelvis, Upper arm BS, Lower Arm BS, Upper Leg BS, Lower Leg BS, Feet BS	“Xsens MVN Awinda, Xsens Technologies, the Netherlands,” 20 Hz	3D sensor orientation and position as estimated by a 9-axis. (proprietary sensor fusion algorithm with a kinematic model)	Clinical assessment consisted of 10-mWT for the P1 and a predefined arm task (patient had to reach as far as possible and to make a circle as wide as possible over a table) for the P2. The second session included a measurement session of 3 h of movement data, which was captured at the patient’s home 4 weeks after discharge.	Quality of movement can be evaluated in a daily life setting. Differences between in-clinic measurements and measurements during daily life are observed by applying the presented metrics and visualization methods. For the upper extremities, the P2 was able to reach a larger area.
Wade 2014^Citation48	10 (8 M, 2F), 59.2 ± 15.3	(2) Wrist BS	“Triaxial accelerometers (Gulf Coast Data Concepts, LLC),” 40 Hz	measured 3D linear acceleration	Unimanual goal-directed movements using both hands, with and without task objects.	Spectral measures of IMUs data are sensitive to performance differences with the nonparetic and paretic limbs and with the object present and absent, for ADL-inspired tasks.

Abbreviations: IMU = Inertial Measurement Unit; NR = Not Reported; BS = Both Side; DAS = Disability Assessment Scale; NHPT = Nine Hole Peg Test; DAS = Disability Assessment Scale; MAS = Modified Ashworth Scale; ADL = Activities of Daily Living; UL = Upper Limb; BBT = Box and Block Test; SIS = Stroke Impact Scale; TLT = Thumb Localizing Test; ARAT = Action Research Arm Test; ROM = Range Of Motion; DoF = Degrees of Freedom; RF = Random Forest; FMA = Fugl-Meyer Assessment; OT = Occupational therapy; MARG = Magnetic, Angular Rate, and Gravity sensor; EMG = Electromyography; EEG = Electroencephalogram; 10mWT = 10 Meters Walking Test; P1 = First Patient; P2 = Second Patient. In Bai et al., 2020 (2), only configuration (b) was used to performed kinematic analysis. Column five reports the source signal from which the kinematic variable of interest has been derived: 6-axis = the sensor fusion algorithm runs using accelerometer and gyroscope data; 9-axis = the sensor fusion algorithm runs using accelerometer, gyroscope, and magnetometer data.

Figure 2. Flow diagram of studies selection process (PRISMA ver. 2020).²⁸ *All records were excluded by humans.

Figure 3. Forest plot for the random effect model correlation meta-analysis between kinematics and clinical data. The continuous line indicates no correlation (right and left, positive and negative correlation, respectively). The dashed line indicates the pooled Z-score. ES: effect size; CI: confidence interval; W: weight; V: variance; Sig.: statistical significance; N: sample number. Studies^30,44 allocated to more than one ICF’ subgroups were considered for meta-analysis only in their dominant subgroup.

Bai L, Pepper MG, Yan Y, Phillips M, Sakel M. Low cost inertial sensors for the motion tracking and orientation estimation of human upper limbs in neurological rehabilitation. IEEE Access. 2020;8:54254–54268.

 Web of Science ®Google Scholar

Bai L, Pepper MG, Yan Y, Phillips M, Sakel M Quantitative measurement of Upper limb motion pre- and post-treatment with Botulinum Toxin. Meas. 2021 Jan 15; 168:108304. 10.1016/j.measurement.2020.108304

 Web of Science ®Google Scholar

Bertomeu-Motos A, Blanco A, Badesa FJ, Barios JA, Zollo L, Garcia-Aracil N. Human arm joints reconstruction algorithm in rehabilitation therapies assisted by end effector robotic devices. J Neuroeng Rehabil. 2018 Feb 20;15(1):10. doi:10.1186/s12984-018-0348-0.

 PubMedGoogle Scholar

Bhagubai MMC, Wolterink G, Schwarz A, Held JPO, Van Beijnum BF, Veltink PH Quantifying pathological synergies in the upper extremity of stroke subjects with the use of inertial measurement units: a pilot study. IEEE J Transl Eng Health Med. 2020 Dec 7; 9:2100211. 10.1109/JTEHM.2020.3042931

 PubMedGoogle Scholar

Kang S, Shin JH, Kim IY, Lee J, Lee JY, Jeong E. Patterns of enhancement in paretic shoulder kinematics after stroke with musical cueing. Sci Rep. 2020 Oct 22;10(1):18109. doi:10.1038/s41598-020-75143-0.

 PubMed Web of Science ®Google Scholar

Li J, Pan B, Jin T, et al. A single task assessment system of upper-limb motor function after stroke. Technol Health Care. 2016 Apr 29;24(Suppl s2):S707–15. doi:10.3233/THC-161199.

 PubMedGoogle Scholar

Ang WS, Geyer H, Chen I-M, Ang WT. Objective assessment of spasticity with a method based on a human upper limb model. IEEE Trans Neural Syst Rehabil Eng. 2018;26(7):1414–1423. doi:10.1109/TNSRE.2018.2821197.

 PubMed Web of Science ®Google Scholar

Chen Y, Yu S, Cai Q, et al. A spasticity assessment method for voluntary movement using data fusion and machine learning. Biomedical Signal Processing and Control. 2021;65:65, 102353. doi:10.1016/j.bspc.2020.102353.

 Web of Science ®Google Scholar

Kim JY, Park G, Lee SA, Nam Y. Analysis of machine learning-based assessment for elbow spasticity using inertial sensors. Sensors (Basel). 2020 Mar 14;20(6):1622. doi:10.3390/s20061622.

 PubMed Web of Science ®Google Scholar

Paulis WD, Horemans HL, Brouwer BS, Stam HJ. Excellent test-retest and inter-rater reliability for Tardieu Scale measurements with inertial sensors in elbow flexors of stroke patients. Gait Posture. Feb, 2011;33(2):185–189. doi:10.1016/j.gaitpost.2010.10.094.

 PubMed Web of Science ®Google Scholar

Chen ZJ, He C, Gu MH, Xu J, Huang XL Kinematic evaluation via inertial measurement unit associated with upper extremity motor function in subacute stroke: a cross-sectional study. J Healthc Eng. 2021 Aug 19; 2021:4071645. 10.1155/2021/4071645

 PubMed Web of Science ®Google Scholar

Held JPO, Klaassen B, Eenhoorn A, et al.Inertial sensor measurements of upper limb kinematics in stroke patients in clinic and home environmentFront Bioeng Biotechnol2018 Apr 12610.3389/fbioe.2018.00027

 PubMed Web of Science ®Google Scholar

Hesam-Shariati N, Trinh T, Thompson-Butel AG, Shiner CT, Redmond SJ, McNulty PA. Improved kinematics and motor control in a longitudinal study of a complex therapy movement in chronic stroke. IEEE Trans Neural Syst Rehabil Eng. Apr, 2019;27(4):682–691. doi:10.1109/TNSRE.2019.2895018.

 PubMed Web of Science ®Google Scholar

Nikmaram N, Scholz DS, Großbach M, et al. Musical sonification of arm movements in stroke rehabilitation yields limited benefits. Front Neurosci. 2019 Dec 20; 13: 1378. 10.3389/fnins.2019.01378.

 PubMed Web of Science ®Google Scholar

Nie JZ, Nie JW, Hung NT, Cotton RJ, Slutzky MW. Portable, open-source solutions for estimating wrist position during reaching in people with stroke. Sci Rep. 2021 Nov 18;11(1):22491. doi:10.1038/s41598-021-01805-2.

 PubMed Web of Science ®Google Scholar

Pan B, Huang Z, Jin T, Wu J, Zhang Z, Shen Y. Motor function assessment of upper limb in stroke patients. J Healthc Eng. 2021;2021:6621950. Feb 24; 2021. doi:10.1155/2021/6621950.

 PubMed Web of Science ®Google Scholar

Park E, Lee K, Han T, Nam HS. Automatic grading of stroke symptoms for rapid assessment using optimized machine learning and 4-limb kinematics: clinical validation study. J Med Internet Res. 2020 Sep 16;22(9):e20641. doi:10.2196/20641.

 PubMed Web of Science ®Google Scholar

Rau CL, Chen YP, Lai JS, et al. Low-cost tele-assessment system for home-based evaluation of reaching ability following stroke. Telemed J E Health. Dec 2013;19(12):973–978. doi:10.1089/tmj.2012.0300.

 PubMed Web of Science ®Google Scholar

Salazar AJ, Silva AS, Silva C, et al. Low-cost wearable data acquisition for stroke rehabilitation: a proof-of-concept study on accelerometry for functional task assessment. Top Stroke Rehabil. Jan-Feb 2014;21(1):12–22. doi:10.1310/tsr2101-12.

 PubMed Web of Science ®Google Scholar

Schwarz A, Veerbeek JM, Held JPO, Buurke JH, Luft AR Measures of interjoint coordination post-stroke across different upper limb movement tasks. Front Bioeng Biotechnol. 2021 Jan 28; 8:620805. 10.3389/fbioe.2020.620805

 PubMed Web of Science ®Google Scholar

Tedim Cruz V, Bento VF, Ribeiro DD, Araújo I, Branco CA, Coutinho P. A novel system for automatic classification of upper limb motor function after stroke: an exploratory study. Med Eng Phys. Dec, 2014;36(12):1704–1710. doi:10.1016/j.medengphy.2014.09.009.

 PubMed Web of Science ®Google Scholar

Wade E, Chen C, Winstein CJ. Spectral analyses of wrist motion in individuals poststroke: the development of a performance measure with promise for unsupervised settings. Neurorehabil Neural Repair. Feb, 2014;28(2):169–178. doi:10.1177/1545968313505911.

 PubMed Web of Science ®Google Scholar

Zollo L, Rossini L, Bravi M, Magrone G, Sterzi S, Guglielmelli E. Quantitative evaluation of upper-limb motor control in robot-aided rehabilitation. Med Biol Eng Comput. Oct, 2011;49(10):1131–1144. doi:10.1007/s11517-011-0808-1.

 PubMed Web of Science ®Google Scholar

Biswas D, Corda D, Baldus G, et al. Recognition of elementary arm movements using orientation of a tri-axial accelerometer located near the wrist. Physiol Meas. Sep 2014;35(9):1751–1768. doi:10.1088/0967-3334/35/9/1751.

 PubMed Web of Science ®Google Scholar

Bochniewicz EM, Emmer G, McLeod A, Barth J, Dromerick AW, Lum P. Measuring functional arm movement after stroke using a single wrist-worn sensor and machine learning. J Stroke Cerebrovasc Dis. Dec, 2017;26(12):2880–2887. doi:10.1016/j.jstrokecerebrovasdis.2017.07.004.

 PubMed Web of Science ®Google Scholar

Lin BS, Hsiao PC, Yang SY, Su CS, Lee IJ. Data glove system embedded with inertial measurement units for hand function evaluation in stroke patients. IEEE Trans Neural Syst Rehabil Eng. Nov, 2017;25(11):2204–2213. doi:10.1109/TNSRE.2017.2720727.

 PubMed Web of Science ®Google Scholar

Mahmoud SS, Cao Z, Fu J, Gu X, Fang Q Occupational therapy assessment for upper limb rehabilitation: a multisensor-based approach. Front Digit Health. 2021 Dec 17; 3:784120. 10.3389/fdgth.2021.784120

 PubMedGoogle Scholar

Mazomenos EB, Biswas D, Cranny A, et al. Detecting elementary arm movements by tracking upper limb joint angles with MARG sensors. IEEE J Biomed Health Inform. Jul 2016;20(4):1088–1099. doi:10.1109/JBHI.2015.2431472.

 PubMed Web of Science ®Google Scholar

Michaelsen SM, Gomes RP, Marques AP, et al. Using an accelerometer for analyzing a reach-to-grasp movement after stroke. Motriz: Revista de Educação Física. Oct/Dec 2013;19(4):746–752. doi:10.1590/S1980-65742013000400012.

 Google Scholar

Repnik E, Puh U, Goljar N, Munih M, Mihelj M. Using inertial measurement units and electromyography to quantify movement during action research arm test execution. Sensors (Basel). 2018 Aug 22;18(9):2767. doi:10.3390/s18092767.

 PubMed Web of Science ®Google Scholar

Schwarz A, Bhagubai MMC, Wolterink G, Held JPO, Luft AR, Veltink PH. Assessment of upper limb movement impairments after stroke using wearable inertial sensing. Sensors (Basel). 2020 Aug 24;20(17):4770. doi:10.3390/s20174770.

 PubMed Web of Science ®Google Scholar

Thies SB, Tresadern PA, Kenney LP, et al. Movement variability in stroke patients and controls performing two upper limb functional tasks: a new assessment methodology. J Neuroeng Rehabil. 2009 Jan 23;6(1):2. doi:10.1186/1743-0003-6-2.

 PubMedGoogle Scholar

van Meulen FB, Reenalda J, Buurke JH, Veltink PH. Assessment of daily-life reaching performance after stroke. Ann Biomed Eng. Feb, 2015;43(2):478–486. doi:10.1007/s10439-014-1198-y.

 PubMed Web of Science ®Google Scholar

van Meulen FB, Klaassen B, Held J, et al. Objective evaluation of the quality of movement in daily life after stroke. Front Bioeng Biotechnol. 2016 Jan 13; 3: 210. 10.3389/fbioe.2015.00210.

 PubMed Web of Science ®Google Scholar

Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. J Clin Epidemiol. Jun 2021;134:178–189. doi:10.1016/j.jclinepi.2021.03.001.

 PubMed Web of Science ®Google Scholar