Accuracy of Wearable Sensor Technology in Hand Goniometry: A Systematic Review

Introduction

Evaluation of hand function comprises the assessment of a diverse array of physiologic characteristics, from strength to temperature sensitivity. Acquisition and interpretation of these data provide clinical decision guidance for rehabilitation planning and insight on the patient’s treatment progress. One of the most relevant characteristics required to be measured is range of motion (ROM), ¹ most commonly measured with protractor-based goniometers. ² Although evidence supports the superiority of manual goniometry to other forms of ROM measurement (eg, visual assessment), ³ there are still a variety of factors that may hinder its accuracy. Assessment by numerous providers with different levels of expertise and in different periods is a crucial component influencing the reliability of measurements, ⁴ supporting the necessity for a more reliable, accurate technique whose measurements do not solely rely on the provider’s skill.

Despite its apparent flaws, the criterion standard for ROM measurement is manual goniometry. The advent of new technologies to evaluate hand ROM may soon leave this technique in the past. For example, in a recent study by Reissner et al, ⁵ the authors found a lower minimal detectable difference and standard error of measurement (SEM) for 3-dimensional (3D) motion analysis of the joints when compared with manual goniometry. Although accurate measurements without the need for human intervention seem possible with the previous approach, the 3D system needs 11 infrared cameras to analyze hand movements, and therefore, it is only possible in a controlled environment. A technique that overcomes this hindrance is photogoniometry, which has shown acceptable agreement levels both with goniometry and when tested for interobserver and intraobserver measurements.^6,7 Although photogoniometry decreases the amount of equipment required, making at-home measurement possible, static images cannot provide a full ROM evaluation. As the hand comprises numerous small joints, another potential obstacle to manual hand goniometry is the length of the process, which can take as long as 30 minutes. ⁸ An ideal technique would be one that can provide continuous accurate measurements without the need for an office visit.

Wearables, such as wrist- based or glove-based devices, can solve the discussed obstacles by integrating sensor technology to provide uninterrupted goniometric readings. Decreasing interobserver measurement errors, ⁸ shortening the evaluation time, and providing ROM measurements during a patient’s daily activities are just a few of the advantages that sensor technology can bring to hand therapy, potentially aiding in the development of new rehabilitation protocols. However, these devices must be accurate enough to provide reliable measurements that serve this purpose. Therefore, this review aims to identify the most accurate wearable sensor technologies for hand ROM measurement.

Methods

Eligibility criteria included studies evaluating hand ROM, studies using goniometric assessments, studies using wearable sensors (ie, gloves, wristbands), and studies providing accuracy values. Only articles written in English were considered for this review. No particular time frame or publication status was considered. Exclusion criteria included studies of isolated sensors and studies that did not provide angle correlations. The review focuses on easily wearable technology, and therefore, studies using exoskeletons or those requiring additional materials (eg, cameras) were not considered.

The acceptable error margin of 5° in ROM clinical assessments proposed by Gajdosik and Bohannon ⁹ was used in this review to determine whether the reported wearable devices were sufficiently accurate. In the evaluated studies, the SEM was taken as representative of accuracy. Accuracy refers to the maximum difference that exists between the measured and actual values. ¹⁰ This value was calculated by comparing the device’s measurements with another measurement technique.

Studies were identified by inquiring PubMed, CINAHL, and EMBASE from conception to April 1, 2020. The following terms were used in combination in all databases: “wearable device,” “wearable technology,” “garment,” “glove,” “textile,” “wristband,” “goniometry,” “goniometer,” “range of motion,” “joint measurement,” “arthrometer,” “inertial sensor,” “motion sensor,” “gyroscope,” “hand,” “wrist,” “finger,” “shoulder,” “arm,” “elbow,” and “forearm.” Search terms were arranged as follows: ((((((((wearable device OR wearable technology OR garment OR glove OR textile OR wristband)) AND (goniometry OR goniometer OR range of motion OR joint measurement OR arthrometer)) AND (inertial sensor OR sensor system OR motion sensor OR gyroscope)) AND (hand OR wrist OR finger)) NOT shoulder) NOT arm) NOT elbow) NOT forearm. In addition, on April 19, 2020, 2 articles meeting the inclusion criteria that did not appear in the first inquiry were added from PubMed.

This systematic review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. A description of the study selection process, following the PRISMA flowchart, is presented in Figure 1. Eligibility assessment was performed by 1 reviewer (F.R.A.), starting with the title and abstract of the studies, followed by full-text evaluation. Data were extracted by the same reviewer. The risk of bias of included studies was assessed using the Risk Of Bias In Non-randomized Studies of Interventions (ROBINS-I) tool of the Cochrane Library. A summary and a graph were created using RevMan 5.3 (Cochrane Collaboration), which allows for bias stratification in several domains (Figures 2 and 3). Statistical analyses are reported when stated by the studies’ authors.

OPEN IN VIEWER

Figure 1.

Study selection process. Process of study selection following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses flowchart for systematic reviews.

OPEN IN VIEWER

Figure 2.

Risk of bias graph.

OPEN IN VIEWER

Figure 3.

Risk of bias summary.

Results

The inquiry identified 71 studies, of which only 11 fulfilled the eligibility criteria (see Table 1). Nine (82%) were observational descriptive studies, whereas the rest were cross-sectional studies that included more than 1 group of patients (2 [18%]). All studies used gloves to evaluate ROM, except for the study by Rowe et al ¹⁴ that used a wristband.

OPEN IN VIEWER

Table 1.

Summary of Included Studies.

Author	Type of study	Type of sensor	Type of wearable	Wearable substrate	Variable	Evaluated joint(s)	Evaluated movements	Intended use	Measurement error	Comparison technique
Wise et al 8	Observational descriptive	Bend sensors	Glove	Cotton/synthetic fabric	Light intensity	MCP, PIP, and IP joints	Flexion/extension	Clinical evaluation	2°-6°	Virtual motion tracking
Kessler et al 11	Observational descriptive	Bend sensors	Glove	Unspecified fabric	Electrical resistance	MCP, PIP, and IP joints	Flexion/extension, abduction/adduction, thumb rotation	Clinical evaluation	5.8°	—
Dipietro et al 1	Observational descriptive	Hall effect sensor	Glove	Unspecified fabric	Electromagnetic field	MCP, PIP, DIP, and CMC joints	Abduction/adduction, flexion/extension	—	1°	—
Oess et al 12	Observational descriptive	Bend sensors	Glove	Stretchable polyamide/Lycra	Electrical resistance	MCP, DIP, and PIP joints of the index finger; IP joint of the thumb; and RC joint	Flexion/extension	Neurologic rehabilitation	3°	Manual goniometry
Oess et al 13	Cross-sectional	Bend sensors	Glove	Stretchable polyamide/Lycra	Electrical resistance	MCP, DIP, and PIP joints of the index finger; IP joint of the thumb; and RC joint	Flexion/extension	Neurologic rehabilitation	5°	Manual goniometry
Rowe et al 14	Observational descriptive	Magnetometers	Wristband	—	Electromagnetic field	RC joint and MCP joint	Flexion/extension of wrist and finger, radial and ulnar deviation of wrist	Neurologic rehabilitation	2.4° radial/ulnar deviation, 5.8° wrist flexion/extension, 4.7° finger flexion/extension	Exoskeleton goniometry
Carbonaro et al 15	Observational descriptive	Bend sensor	Glove	Knitted piezoresistive fabric	Electrical resistance	MCP joint of thumb, index, and middle fingers	Flexion/extension	Neurologic rehabilitation	3°	Camera-based motion tracking
Park et al 16	Observational descriptive	Liquid metal strain (bend) sensor	Glove	Silicone	Electrical resistance	MCP, PIP, and DIP joints of the fingers; CMC, MCP, and IP joints of thumb	Flexion/extension, abduction/adduction	VR	0.66°-2.55° with a mean of 1.39°	Camera-based motion tracking
Li et al 17	Observational descriptive	Bend sensors + inertial sensor	Glove	Silicone	Electrical resistance + acceleration, angular velocity	MCP, PIP, DIP, and IP joints of thumb and wrist	Flexion/extension	Clinical evaluation	Inertial sensor 0.49°, bend sensor 2.44° (overall error of 6.35° ± 0.92° when evaluated in patients)	Manual goniometry
Lin et al 18	Observational descriptive	Inertial sensors	Glove	Unspecified fabric	Acceleration, angular velocity, electromagnetic field	MCP, PIP, DIP, and IP joints of thumb and wrist	Flexion/extension	Rehabilitation	Less than 0.2° static; ±3° in motion	Manual goniometry
Lin et al 19	Cross-sectional	Inertial sensors	Glove	Unspecified fabric	Acceleration, angular velocity, electromagnetic field	MCP, PIP, DIP, and IP joints of thumb and wrist	Flexion/extension	Neurologic rehabilitation	Less than 2°	Manual goniometry

MCP = metacarpophalangeal; PIP = proximal interphalangeal; IP = interphalangeal; DIP = distal interpalangeal; CMC = carpometacarpal; RC = radiocarpal.

Ten of the 12 studies providing accuracy estimates stated the technique they used to compare their devices, with 5 using manual goniometry, 3 using infrared camera–based motion tracking, 1 using virtual motion tracking, and 1 using exoskeleton goniometry. The following paragraphs provide a brief discussion of the studies’ methodologies and their results.

The first published studies on the use of wearable sensors for hand ROM measurement all described similar methodologies, focused on the correct placement of the hand in prespecified positions to ensure reproducibility of the measurements. Wise et al ⁸ and Dipietro et al ¹ achieved this by creating a mold of the patient’s grip with plaster, integrating it in a set of tests that consisted of measurements of the hand in a relaxed, extended position over a flat surface, followed by its closure over the mold. The work by Wise et al ⁸ evaluated the DataGlove (VPL Research, California, USA). Fiber optic bend sensors were embedded in the glove and connected to an interface board in the wrist, allowing data on the light intensity changes to be transferred to an interface box and translated to degrees of movement by a computer. Formerly designed for gesture recognition, the glove showed an error between 2° and 6° when compared with the 3D motion tracker Polhemus (Polhemus, Vermont, USA) by the manufacturer. ⁸ When tested by the authors, an overall error of 5.6° and a standard deviation of 2.3° were obtained, but with significant intersubject variability (P < .05). ⁸

Dipietro et al ¹ evaluated the Humanglove (HumanWare, Tuscany, Italy), which has 20 Hall effect sensors positioned in the joints, which can translate the electrical current changes into degrees of motion by a tracker positioned around the wrist. The Hall effect refers to the distortion of an electrical current caused by a magnetic field (Figure 4). ²⁰ The SEM provided by the manufacturer was 1°. ¹ Intraclass correlation coefficients ranged between 0.7 and 1, placing the consistency of the Humanglove between moderate and excellent. ¹

OPEN IN VIEWER

Figure 4.

Hall effect sensor glove.

Kessler et al ¹¹ evaluated CyberGlove (CyberGlove Systems Inc, California, USA), a device that used bend sensors for which changes in electrical resistance when bent correlated to degrees of movement. The authors tested the glove for flexion and extension, abduction, and thumb rotation, each using different constraints to hold the fingers in the required angle. ¹¹ Excluding thumb abduction and rotation, for which the SEMs were excessively high, the error of this device almost reached 6°. ¹¹

Similar to Kessler et al, ¹¹ Oess et al ¹² asked patients to keep their hand fixed in specific angles but without constraints to keep them in place. The NeuroAssess Glove, which they created, works similarly and uses the same sensor type as the CyberGlove. Accuracy was calculated by comparison to manual goniometry, showing a measurement difference of 0.3°. ¹² The use of the Bland-Altman method showed an agreement between manual goniometry and bend sensors (P < .05).

Park et al ¹⁶ created a silicone glove that was able to measure degrees of movement based on changes in the electrical conductance of sensors made of microchannels filled with liquid metal (Figure 5). The deformation of the microchannels altered their radius, changing the metal’s resistance. ¹⁶ This value was measured and compared with an optical tracking system that provided the authors with the goniometric measurements for calibration. The average SEM of this glove was 1.39°. ¹⁶

OPEN IN VIEWER

Figure 5.

Liquid metal bend sensor glove.

Li et al ¹⁷ evaluated the accuracy with fixed preestablished positions and also repeatability with handgrip measurement. The authors developed a silicone glove that measured ROM with bend sensors constructed from electrical semiconducting ethylene-propylene rubber tape over a flexible printed circuit. They used an inertial motion unit (IMU) in the dorsum of the hand to measure wrist and finger movement. Both the IMU output and the bending sensor’s resistance change showed strong negative linear correlations with the actual bending angle (R² = 0.9997 and R² = 0.9963, respectively). ¹⁷ When translated into angle measurements, the IMU had an error of 0.49°, whereas bend sensors had an error of 2.44°. ¹⁷ However, when measuring handgrip, the average error in measurement increased to 6.35°. ¹⁷

The aim of the second study by Oess et al ¹³ was to assess their devices in patients undergoing neurologic rehabilitation. The authors evaluated the devices’ data obtained while patients were performing hand movements related to activities of daily living, such as pouring water into a glass from a bottle or unscrewing the lid of a jar. Oess et al ¹³ adjusted their previous prototype by implementing short sensors over every finger joint instead of using a long sensor to cover all joints and by increasing voltage input to increase the sensors’ resolution over a 100° of flexion. The mean difference between the glove’s output and manual goniometry was −0.59 (95% confidence interval = −4.91 to 3.74, P < .05). ¹³ Comparing a group of healthy patients with a group of patients with cervical spinal cord injuries, they showed that the glove could significantly differentiate between hand function levels with an accuracy of 5° (P < .01). ¹³

Lin et al^18,19 followed methodologies similar to those of Kessler et al. ¹¹ The accuracy of the glove in both studies was 0.2° in static measurements and 3° in dynamic measurements.^18,19

Finally, 2 studies let patients explore the full ROM of their hands to evaluate their devices. Carbonaro et al ¹⁵ developed a prototype glove to assess the recovery of poststroke patients. The glove’s goniometers consisted of a layer of insulating material in between 2 layers of knitted piezoresistive fabric, and changes in resistance when bent correlated to curvature angles. The glove’s measurements were compared with an optical tracking system that served as the criterion standard for motion recognition and angle calculation. ¹⁵ The accuracy of the glove’s goniometers was within 3.6° of the camera-based motion tracking. ¹⁵ Rowe et al ¹⁴ created a wristband and a magnetic ring with an accelerometer and magnetometers that could measure both activity and degrees of motion of the joints. The device by Rowe et al ¹⁴ was the only wristband found in the reviewed literature, achieving accuracies below 5° for wrist radial or ulnar deviation and finger flexion and extension. Measurement errors in wrist flexion or extension were higher than 5°. ¹⁴

Discussion

Wearable technology is a rapidly advancing field. From smart wound dressings to pain intensity prediction, these devices allow physicians and health care workers to use objective measurements in their everyday practice. Eliminating subjectivity from clinical assessment is one of the main advantages these devices bring to medicine. In addition, the uninterrupted objective monitoring of physiological parameters and activity may offer comprehensive information about a person’s health status over time, ²¹ potentially improving safety for both hospitalized and ambulatory patients. ²² Hand ROM is a difficult parameter to evaluate as visual assessments and goniometry are not reliable measurement techniques. The studies included in this review provide insight into the advancement of wearable sensors to replace today’s methods. These new measurement techniques require a high level of accuracy and reproducibility to advise their use, and therefore, it is crucial to be aware of measurement errors when evaluating their performance.

As previously stated by Gajdosik and Bohannon, ⁹ an acceptable measurement error in goniometry is 5°. Most of the devices evaluated in the reviewed studies achieved an accuracy of 5° or less. The wearable sensor that had the best accuracy had an error of 0.2° with a standard deviation of 3° and used inertial sensors to calculate the degrees of movement based on acceleration, angular velocity, and electromagnetic changes. ¹⁸ The studies that used IMUs tended to report better accuracy.^17-19 An advantage of IMUs over conventional bend sensors is their modularity, allowing malfunctioning sensors to be easily replaced. However, one of these studies combined inertial and bend sensors, which impacted its accuracy. ¹⁷

Li et al ¹⁷ used a different material for their bend sensors (ie, an electrical semiconducting ethylene-propylene rubber tape) owing to deficiencies found in conventional bend sensors. As these sensors have a layer of carbon ink that grants them electroconductive properties, they are prone to microcracks that may alter resistance patterns and increase measurement instability.^12,13 Several studies have also found that larger glove sizes lead to slippage and uneven bending of the sensors, leading to the formation of microcracks. These data are in accordance with the findings by Li et al, ¹⁷ as the performance of their glove was better with hand lengths of 17 to 19 cm. Furthermore, the thickness of conventional bend sensors increases the feeling of tightness and makes patients uncomfortable. ¹³

A solution for microcracking and sensor stiffness is the introduction of sensors such as those proposed by Park et al, ¹⁶ composed of microchannels filled with conductive liquid metal. When compared with other studies using classic bend sensors, the authors’ glove accuracy was superior. The most important drawback reported in the study was slippage of the silicone glove, highlighting the importance of choosing the correct materials both for the sensors and for the supporting substrate. Therefore, silicone and fabric gloves share a similar obstacle, where slippage between the skin and the sensors might be a relevant source of error, influencing accuracy.

A study that was not included in the review but proposes a solution for slippage is the one by Gu et al. ²³ The authors created an ionic hydrogel embedded in polydimethylsiloxane, resembling previous works with silicone, with the novelty of tight skin attachment. Although they found significant resistance changes when flexed at different angles, the accuracy of the device was not evaluated due to the study’s objective. The authors reported that the material adhered effectively to the skin, preserving its functionality for up to 7 days, allowing continuous monitoring of the patient’s hand movements. ²³ Given this device’s advantages, further studies evaluating devices analogous to this in clinical settings are warranted.

Most of the authors compared their devices with manual goniometry for accuracy assessment. Prone to human error, this goniometric technique is not reliable enough to provide precise measurements, yet its extensive use is comprehensible due to difficulties in obtaining the equipment for camera-based or virtual goniometry. However, this raises doubts about the validity of accuracy measurements for some of these devices. Although evidence on the validity of infrared cameras, electromagnetic tracking of the fingers, and photogoniometry for hand ROM measurement is scarce, it is hypothesized that these measurement techniques might be more accurate by eliminating human subjectivity. Future studies estimating the accuracy of wearables in hand ROM measurement should use these measurement techniques to evaluate performance.

Limitations

This systematic review has multiple limitations. First, it reported, within the constraints of the selected databases, all studies in English to date describing the use of wearable devices for hand ROM measurement, which hampers the quantity of extracted information. Other limitations of our review include the scarcity of studies reporting on this topic and the potential bias of misinterpreting data and results. The process of selecting studies to be included in the review should also be regarded as a potential source of bias common to systematic reviews.

Conclusion

Most of the wearable devices were gloves using electroconductive bend sensors, with changes in resistance correlating to degrees of bending. One of the main disadvantages observed with these sensors was the increasing error, probably due to microcracking of the carbon-based ink they were covered in. Malleable materials with the same electrical characteristics embedded in a tightly adhering glove can adequately solve this issue, providing a good starting point for future studies. Inertial motion units had the lowest measurement errors, making wearables with these sensors the most accurate. Although most of the devices reached adequate levels of measurement error, accuracy evaluation in the reviewed studies might be subject to bias owing to the use of poorly reliable measurement techniques for comparison. Future studies should use different comparison techniques, such as infrared camera–based goniometry or virtual motion tracking, to evaluate the performance of wearable devices.