Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 48 of the 48 references it cites were found in a registry or an index; none of the 50 sources its claims cite is checked for support yet.

A Low-Cost, Accurate, and Easily-Worn E-Skin and IMU Hand Kinematic Measurement System

Open access

Abstract

Stroke and other neurological injuries impair hand function. Although rehabilitation therapists encourage reintegration of the affected hand into daily activities, there are few tools that can be worn during everyday life that provide quantitative feedback on how much or how effectively the hand is used. Wearable sensors that can accurately track hand movements and are easily applied and removed can present intuitive feedback that could motivate hand use similarly to how pedometers encourage walking. While tracking all the hand and finger joints is needed for scientific studies, under-sensorization, or using fewer sensors than required for tracking all degrees of freedom, may suffice for providing users feedback about hand use in everyday life. Further, it may enable a wearable device to be more easily donned and doffed, more power efficient, and more cost efficient. Here we develop a low-cost, multi-sensor, wireless wearable system for tracking selected hand and wrist movements during everyday life. The system includes fabricated soft, stretchable “e-skin” bend sensors and off-the-shelf inertial measurement units (IMUs) that accurately measure finger bend angles and wrist movements. The system also includes an application and removal protocol that enabled naïve unimpaired participants to apply and remove the system in ~5 min and ~4 min, respectively. The system cost was $111 per device, with prices falling to an estimate of $55 when manufactured at scale. This hand wearable demonstrates accurate kinematic tracking and user-friendly donning/doffing workflows for unimpaired participants, making it a promising platform for everyday hand tracking. Future work will extend this platform to the movement-impaired population for neurorehabilitation applications.

Overview

CoChat Summary

TL;DR

The paper presents a low-cost, wireless, under-sensorized hand/wrist kinematic system combining soft liquid-metal e-skin bend sensors and IMUs, with a user-friendly donning/doffing workflow. It reports benchtop and on-hand accuracy, usability in unimpaired participants, a ~9 h wear demonstration, and a $111 prototype cost ($55 estimated at scale).

Key findings

  • Benchtop strain-sensor accuracy: mean absolute error (MAE) 1.39° and RMSE 1.75° across 13 sensors; 99.7% of predictions had absolute error ≤5°. Under worst-case misalignment, MAE was 2.31° (RMSE 2.80°), with 91.3% of Model 1 predictions ≤5°.
  • On-hand strain-sensor accuracy for index MCP: overall MAE 4.91° (mean RMSE 7.30°) across first-application self-trained and second-application Model 1 conditions; 88% of pooled predictions were ≤10°. On-hand wrist flexion/extension with two IMUs: MAE 6.73° (RMSE 8.80°); 79.4% of predictions ≤10°, 92.7% ≤15°, and 98% ≤20°.
  • Usability: naïve unimpaired participants applied the full system in ~5 min and removed it in ~4 min. Toupee tape (TT) was significantly faster to apply/remove than Derma-Tac (DT) and had fewer failures: TT 0/24 total failures and 1/24 partial failures vs DT 4/24 total failures and 4/24 partial failures.
  • System cost: $111 per prototype device, estimated $118 for a future custom-PCB iteration, and estimated $55 per unit at manufacturing scale. A single participant wore the system continuously for ~9 h during daily activities, exceeding the conservative ~8.2 h battery-life estimate.

Methods

  • The system used in-lab-fabricated soft silicone microfluidic “e-skin” strain sensors with eutectic gallium–indium channels, plus off-the-shelf BNO055 IMUs, custom data acquisition electronics, a watch-style wrist unit, and a dorsal hand-back unit; data streamed wirelessly via Bluetooth.
  • Strain sensors were characterized on a custom benchtop “autobender” against a rotary encoder, then tested on-hand in N=4 unimpaired participants using an overhead camera and DeepLabCut as ground truth for index MCP and wrist angles; calibration used quadratic fits between normalized ADC values and known bend angles.
  • A separate N=4 unimpaired usability study compared TT vs DT attachment over repeated application/removal cycles, including splinted dexterity and simulated ADLs; a single participant completed a ~9 h long-term wear protocol. Statistical analysis used ANOVA, paired t-tests, and linear mixed-effects models with effect sizes and confidence intervals.

Limitations

  • The paper states that accuracy and usability were tested only in small unimpaired populations; future studies in larger movement-impaired populations are required, and the on-hand accuracy of the thumb MCP remains uncharacterized.
  • The long-term wear demonstration was N=1 and qualitative; real-world artifacts such as sensor slip, wire motion, direct pressure, impacts, temperature-dependent drift, and electromagnetic interference were not fully characterized.
  • The authors note risk of underpowered statistics in small-sample device testing; cost figures for the custom PCB and scaled manufacturing are engineering estimates, and the comfort study’s participant numbering appears inconsistent in the text (N=4 but listed as P5–P9).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/s26154795

Cites

48 of 48 references verified0 citation errors · 0 not checkable

Statistics

Figures

Figure 1 Liquid-metal strain sensor and fabrication workflow. (A) Assembled sensor showing wire routing and watch connector. (B) Sensor geometry with liquid-metal channel cross-sectional area (CXA), channel length (L0), sensor length (L1), width, thickness, and bending axis. (C) Stepwise fabrication: (1) Fill mold with Ecoflex 00-30 (1.2 mL/mold). (2) Conduct 3 × 5 min degas pulses at −25 inHg (15 min total). (3) Cure for 30 min at 60 °C. (4) Place ABS sacrificial inlay with alignment ears and add additional Ecoflex 00-30 (1.4 mL/mold). (5) Conduct 3 × 5 min degas pulses at −25 inHg (15 min total). (6) Cure for 30 min at 60 °C. (7) Demold, trim alignment ears, and soak in acetone (~12 h) to dissolve ABS. (8) Flush microchannel with acetone. (9) Inject eutectic gallium–indium (EGaIn; conductive liquid metal) to fill channel. (10) Insert leads and insulate ports with Sil-Poxy. (11) Add wire management (Velcro wrap) and solder magnetic connectors. (12) Optional: Bond fabric-tape anchors with Sil-Poxy. Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/cyy7xqe (accessed on 19 June 2026).
Figure 1 Liquid-metal strain sensor and fabrication workflow. (A) Assembled sensor showing wire routing and watch connector. (B) Sensor geometry with liquid-metal channel cross-sectional area (CXA), channel length (L0), sensor length (L1), width, thickness, and bending axis. (C) Stepwise fabrication: (1) Fill mold with Ecoflex 00-30 (1.2 mL/mold). (2) Conduct 3 × 5 min degas pulses at −25 inHg (15 min total). (3) Cure for 30 min at 60 °C. (4) Place ABS sacrificial inlay with alignment ears and add additional Ecoflex 00-30 (1.4 mL/mold). (5) Conduct 3 × 5 min degas pulses at −25 inHg (15 min total). (6) Cure for 30 min at 60 °C. (7) Demold, trim alignment ears, and soak in acetone (~12 h) to dissolve ABS. (8) Flush microchannel with acetone. (9) Inject eutectic gallium–indium (EGaIn; conductive liquid metal) to fill channel. (10) Insert leads and insulate ports with Sil-Poxy. (11) Add wire management (Velcro wrap) and solder magnetic connectors. (12) Optional: Bond fabric-tape anchors with Sil-Poxy. Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/cyy7xqe (accessed on 19 June 2026).CC BY · PubMed Central
Figure 2 Benchtop testing of strain samples using an autobender. (A) Schematic of the autobender. The strain sensor was attached to both the fixed and moving plates using Velcro. As the moving plate was actuated by a servo motor, the sensor bent over a curved radius in 15 deg increments from 0 deg to 90 deg and back, completing five full cycles. The ground truth bend angle was measured using a rotary encoder mounted coaxially with the servo motor’s axis of rotation. (B) Change in resistance (ΔR/R0) versus estimated mechanical strain (ϵXX). Experimental data (points) are compared against the predicted theoretical model (dashed line) for three different cross-sectional areas (CXA). The inset equation shows the theoretical relationship between the normalized resistance and mechanical strain. In the legend, “thick” corresponds to the thickness of the sample. Agreement between the experimental measurements and the theoretical model was quantified using variance accounted for (VAF), which was 0.999, 0.994, and 0.993 for the three samples. (C) Normalized sensor values versus bend angle and the quadratic fit. The data shown are the same experimental data as the black curve in (B). (D) Absolute angular error box plots for 13 different sensor samples grouped into five sensor-length categories. A one-way ANOVA did not detect a statistically significant effect of sensor length on mean absolute error (F(4, 8) = 1.41, p = 0.31, ω² = 0.11). (E) Schematic illustrating the reapplication process. Numbers indicate the application numbers: (1) apply sensor for the first time, and train Model 1, (2–3) apply sensor for the second and third times and train “Self-Trained Model i” models, and apply self-trained models and Model 1. (F) Reapplication absolute angular error box plots for three different samples in the sensor reapplication experiments. The split blue/orange boxes correspond to application (1) in (E). Blue-only boxes correspond to errors from self-trained models in reapplication steps 2–3, and orange-only boxes correspond to errors obtained by applying the first-application calibration model (Model 1). When MAEs were calculated by averaging within each sample across reapplications, Model 1 produced significantly higher errors than the corresponding self-trained models (paired t-test, t(2) = 6.72, p = 0.02). The estimated mean paired difference was 0.64° (95% CI [0.23°, 1.06°], Cohen’s d = 3.83, Hedges’ g = 2.19). No statistically significant difference was detected between the second and third reapplications when using Model 1 (paired t-test, t(2) = −0.82, p = 0.50). The estimated mean difference was −0.43° (95% CI [−2.66°, 1.81°], Cohen’s d = −0.47, Hedges’ g = −0.27). (G) Schematic of misalignment angular error testing on the autobender. The same sample was intentionally misaligned by 10 deg and 20 deg to evaluate the resulting angular error. (H) Misalignment absolute angular error box plots for three different samples. Split blue/orange boxes correspond to the 0° aligned condition. Blue-only boxes represent errors from self-trained models, while orange-only boxes represent errors obtained by applying the 0° calibration model (“Model 1”) to the misaligned data. No statistically significant difference was detected between errors derived from self-trained models and those obtained using Model 1 (paired t-test, t(2) = 1.29, p = 0.33). The estimated mean difference was 1.67° (95% CI [−3.79°, 7.13°], Cohen’s d = 0.74, Hedges’ g = 0.42). Partially created in BioRender. Oppenheim, T. (2026). https://biorender.com/86pe6ky (accessed on 19 June 2026).
Figure 2 Benchtop testing of strain samples using an autobender. (A) Schematic of the autobender. The strain sensor was attached to both the fixed and moving plates using Velcro. As the moving plate was actuated by a servo motor, the sensor bent over a curved radius in 15 deg increments from 0 deg to 90 deg and back, completing five full cycles. The ground truth bend angle was measured using a rotary encoder mounted coaxially with the servo motor’s axis of rotation. (B) Change in resistance (ΔR/R0) versus estimated mechanical strain (ϵXX). Experimental data (points) are compared against the predicted theoretical model (dashed line) for three different cross-sectional areas (CXA). The inset equation shows the theoretical relationship between the normalized resistance and mechanical strain. In the legend, “thick” corresponds to the thickness of the sample. Agreement between the experimental measurements and the theoretical model was quantified using variance accounted for (VAF), which was 0.999, 0.994, and 0.993 for the three samples. (C) Normalized sensor values versus bend angle and the quadratic fit. The data shown are the same experimental data as the black curve in (B). (D) Absolute angular error box plots for 13 different sensor samples grouped into five sensor-length categories. A one-way ANOVA did not detect a statistically significant effect of sensor length on mean absolute error (F(4, 8) = 1.41, p = 0.31, ω² = 0.11). (E) Schematic illustrating the reapplication process. Numbers indicate the application numbers: (1) apply sensor for the first time, and train Model 1, (2–3) apply sensor for the second and third times and train “Self-Trained Model i” models, and apply self-trained models and Model 1. (F) Reapplication absolute angular error box plots for three different samples in the sensor reapplication experiments. The split blue/orange boxes correspond to application (1) in (E). Blue-only boxes correspond to errors from self-trained models in reapplication steps 2–3, and orange-only boxes correspond to errors obtained by applying the first-application calibration model (Model 1). When MAEs were calculated by averaging within each sample across reapplications, Model 1 produced significantly higher errors than the corresponding self-trained models (paired t-test, t(2) = 6.72, p = 0.02). The estimated mean paired difference was 0.64° (95% CI [0.23°, 1.06°], Cohen’s d = 3.83, Hedges’ g = 2.19). No statistically significant difference was detected between the second and third reapplications when using Model 1 (paired t-test, t(2) = −0.82, p = 0.50). The estimated mean difference was −0.43° (95% CI [−2.66°, 1.81°], Cohen’s d = −0.47, Hedges’ g = −0.27). (G) Schematic of misalignment angular error testing on the autobender. The same sample was intentionally misaligned by 10 deg and 20 deg to evaluate the resulting angular error. (H) Misalignment absolute angular error box plots for three different samples. Split blue/orange boxes correspond to the 0° aligned condition. Blue-only boxes represent errors from self-trained models, while orange-only boxes represent errors obtained by applying the 0° calibration model (“Model 1”) to the misaligned data. No statistically significant difference was detected between errors derived from self-trained models and those obtained using Model 1 (paired t-test, t(2) = 1.29, p = 0.33). The estimated mean difference was 1.67° (95% CI [−3.79°, 7.13°], Cohen’s d = 0.74, Hedges’ g = 0.42). Partially created in BioRender. Oppenheim, T. (2026). https://biorender.com/86pe6ky (accessed on 19 June 2026).CC BY · PubMed Central
Figure 3 Strain sensor error analysis on the hand. (A) Camera and stabilizing fixture for user study to investigate the on-hand sensor accuracy of measuring index MCP flexion/extension. (B) Example calibration data from the first sensor application on the third participant (P3). (C) Example time-series comparison of the predicted bend angles from strain sensors and the corresponding camera-measured bend angle for the index MCP during the slowest bend rate (30 BPM). (D) Absolute angular error box plot computed from the time-series comparison shown in (C). (E) Schematic showing placement of the strain sensor over the index and thumb MCP joints. (F,G) Similar to (C,D), except the time-series data/absolute angular error box plot correspond to the 240 BPM MCP bend rate. (H) Summary angular error box plots from the first sensor application for each participant and each MCP bend rate. The bar plot on the far right shows the MAE across all participants for each MCP bend rate. (I) Second-application summary. Blue boxes represent errors derived from a calibration curve fit using second-application data (“self-trained model”), while orange boxes represent errors obtained by applying the first-application calibration to the second-application data (“Model 1”). When MAEs were averaged across speeds within each participant, no statistically significant difference was detected between the two calibration approaches (paired t-test, t(2) = 1.20, p = 0.35). The estimated mean difference was −0.61° (95% CI [−2.79°, 1.57°], Cohen’s d = −0.70, Hedges’ g = −0.40). A linear mixed-effects model with speed as a fixed effect and participant as a random effect on intercept, pooled across participants and applications, revealed no significant effect of speed on the MAE (slope = 0.096, CI [−0.252, 0.443], z = 0.539, N = 45, p = 0.59, R2 marginal = 0.007, R2 conditional = 0.007). Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/t2ywslt (accessed on 19 June 2026).
Figure 3 Strain sensor error analysis on the hand. (A) Camera and stabilizing fixture for user study to investigate the on-hand sensor accuracy of measuring index MCP flexion/extension. (B) Example calibration data from the first sensor application on the third participant (P3). (C) Example time-series comparison of the predicted bend angles from strain sensors and the corresponding camera-measured bend angle for the index MCP during the slowest bend rate (30 BPM). (D) Absolute angular error box plot computed from the time-series comparison shown in (C). (E) Schematic showing placement of the strain sensor over the index and thumb MCP joints. (F,G) Similar to (C,D), except the time-series data/absolute angular error box plot correspond to the 240 BPM MCP bend rate. (H) Summary angular error box plots from the first sensor application for each participant and each MCP bend rate. The bar plot on the far right shows the MAE across all participants for each MCP bend rate. (I) Second-application summary. Blue boxes represent errors derived from a calibration curve fit using second-application data (“self-trained model”), while orange boxes represent errors obtained by applying the first-application calibration to the second-application data (“Model 1”). When MAEs were averaged across speeds within each participant, no statistically significant difference was detected between the two calibration approaches (paired t-test, t(2) = 1.20, p = 0.35). The estimated mean difference was −0.61° (95% CI [−2.79°, 1.57°], Cohen’s d = −0.70, Hedges’ g = −0.40). A linear mixed-effects model with speed as a fixed effect and participant as a random effect on intercept, pooled across participants and applications, revealed no significant effect of speed on the MAE (slope = 0.096, CI [−0.252, 0.443], z = 0.539, N = 45, p = 0.59, R2 marginal = 0.007, R2 conditional = 0.007). Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/t2ywslt (accessed on 19 June 2026).CC BY · PubMed Central
Figure 4 IMU error analysis on the hand. (A) Schematic illustrating the plane-method calculation used to measure the palm flexion angle from two IMUs. (B) Photograph of the customized breakout board designed for the BNO055 IMU. (C) Example time-series comparison of predicted bend angles from the IMU (orange) and the corresponding camera-measured bend angle (blue) for wrist flexion/extension during the slowest bend rate (30 BPM). (D) Absolute angular error box plot computed from the time-series comparison shown in (C). (E) Schematic showing IMU placement within a watch case and on the dorsal surface of the palm, with both IMUs oriented in the same direction. (F,G) Similar to (C,D), except the time-series data/absolute angular error box plot correspond to the 240 BPM bend rate. (H) Summary angular error box plots from the first application for each participant and each wrist bend rate. The bar plot on the far right shows the mean angular error across all participants for each wrist bend rate. (I) Same as (H) but for the second application. After averaging the MAE across the first and second applications within each participant and within each movement speed, a linear mixed-effects model with speed as a fixed effect and participant as a random intercept revealed a significant effect of speed on the mean IMU MAE (slope = 0.585 deg per speed category, z = 2.018, N = 19, p = 0.04356, 95% CI [0.017, 1.152], R2 marginal = 0.114, R2 conditional = 0.508). Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/cxnk361 (accessed on 19 June 2026).
Figure 4 IMU error analysis on the hand. (A) Schematic illustrating the plane-method calculation used to measure the palm flexion angle from two IMUs. (B) Photograph of the customized breakout board designed for the BNO055 IMU. (C) Example time-series comparison of predicted bend angles from the IMU (orange) and the corresponding camera-measured bend angle (blue) for wrist flexion/extension during the slowest bend rate (30 BPM). (D) Absolute angular error box plot computed from the time-series comparison shown in (C). (E) Schematic showing IMU placement within a watch case and on the dorsal surface of the palm, with both IMUs oriented in the same direction. (F,G) Similar to (C,D), except the time-series data/absolute angular error box plot correspond to the 240 BPM bend rate. (H) Summary angular error box plots from the first application for each participant and each wrist bend rate. The bar plot on the far right shows the mean angular error across all participants for each wrist bend rate. (I) Same as (H) but for the second application. After averaging the MAE across the first and second applications within each participant and within each movement speed, a linear mixed-effects model with speed as a fixed effect and participant as a random intercept revealed a significant effect of speed on the mean IMU MAE (slope = 0.585 deg per speed category, z = 2.018, N = 19, p = 0.04356, 95% CI [0.017, 1.152], R2 marginal = 0.114, R2 conditional = 0.508). Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/cxnk361 (accessed on 19 June 2026).CC BY · PubMed Central
Figure 5 The full wireless kinematics system. (A) Full wireless kinematics system on the hand. (B) Exploded view of watch case. (C) Donning/doffing strategy for the on-hand kinematics system. Participants first measure and mark joint centers (X), and mount the watch case (mag connectors outward) and hand-back unit (HBU). The application flow then branches by adhesive: Toupee tape (TT) in orange uses a dry-alignment step that allows adjustment before adhesion, followed by pressing over TT pad zones and connecting the magnetic lead. Derma-Tac (DT) in blue is painted onto the pad guides (maintaining a consistent adhered surface area with the TT methodology), cured for at least 2 min, and then aligned and adhered in a single, non-adjustable press. Both application processes are the same for the final steps: re-press the pads to secure the sensors, attach the IMU, and settle the wires. Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/kg390w6 (accessed on 19 June 2026).
Figure 5 The full wireless kinematics system. (A) Full wireless kinematics system on the hand. (B) Exploded view of watch case. (C) Donning/doffing strategy for the on-hand kinematics system. Participants first measure and mark joint centers (X), and mount the watch case (mag connectors outward) and hand-back unit (HBU). The application flow then branches by adhesive: Toupee tape (TT) in orange uses a dry-alignment step that allows adjustment before adhesion, followed by pressing over TT pad zones and connecting the magnetic lead. Derma-Tac (DT) in blue is painted onto the pad guides (maintaining a consistent adhered surface area with the TT methodology), cured for at least 2 min, and then aligned and adhered in a single, non-adjustable press. Both application processes are the same for the final steps: re-press the pads to secure the sensors, attach the IMU, and settle the wires. Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/kg390w6 (accessed on 19 June 2026).CC BY · PubMed Central
Figure 6 User comfort study protocol and comparative performance of two sensor-attachment methods (DT, blue vs. TT, orange). (A) Study timeline. Participants completed an allergy check, experimenter-guided APP0/REM0, then three timed cycles of application and removal without any experimenter feedback (APP1–3, REM1–3), indicated by the green clock icon. Fine motor tasks (pinch/grasp, writing, and typing) and daily-living tasks (eating, phone use, hand sanitizer application, and jacket on/off) followed APP2; APP2/REM2 were performed with finger splints to limit finger range of motion. A ~60 min free-time block preceded REM3. Cycle-specific questions followed each APP/REM cycle (indicated by the clipboard icon); final open-ended questions were asked prior to concluding the session. (B) Application time by method and cycle with participant traces in gray and mixed-effects learning lines in black. DT and TT both get significantly faster with each application (linear mixed-effects model, random intercept by participant, DT: slope = −39.5 s per application, 95% CI [−74.2, −4.78], t = −2.23, p = 0.026, R2 marginal = 0.092, R2 conditional = 0.796, n = 12, TT: slope = −60.8 s per application, 95% CI [−99.5, −22.0], t = −3.07, p = 0.0021, R2 marginal = 0.143, R2 conditional = 0.833, n = 12). (C) Pooled over APP1-3, TT applications were significantly faster than DT applications (difference between mean application times with DT and TT was 130.6 s, 95% CI [−180.06, −81.11], paired, two-sided t-test, t = −5.810, p = 0.0001, n = 12 paired observations from 4 participants × 3 APP, Cohen’s d = −1.677, Hedges’ g = −1.560). (D) Removal time by method and cycle with participant traces in gray and mixed-effects learning lines in black. DT did not significantly change in removal time but TT significantly improved in speed (linear mixed-effects model, random intercept by participant, DT: 95% CI on slope [−60.48, 7.98], t = −1.503, p = 0.133, R2 marginal = 0.036, R2 conditional = 0.824, n = 12, TT slope: −44.4 s per removal, 95% CI [−87.23, −1.52], t = −2.030, p = 0.0424, R2 marginal = 0.070, R2 conditional = 0.812, n = 12). (E) Pooled over REM1–3, TT removals were significantly faster than DT removals (difference between mean removal times between DT and TT of 157.9 s, 95% CI [−226.60, −89.23], paired, two-sided t-test, t = −5.060, p = 0.0004; n = 12 paired observations = 4 participants × 3 REM, Cohen’s d = −1.461, Hedges’ g = −1.359). (F) Survey ratings (1–10; paired, two-sided t-tests; n = 12 pairs/metric) favor TT on ease (+0.75, 95% CI [0.33, 1.17], t = 3.954, p = 0.0023, Cohen’s d = 1.141, Hedges’ g = 1.062), stability (+1.33, 95% CI [0.79, 1.88], t = 5.360, p = 0.00023, Cohen’s d = 1.547, Hedges’ g = 1.439), and adhesion (+1.33, 95% CI [0.60, 2.07], t = 4.000, p = 0.0021, Cohen’s d = 1.155, Hedges’ g = 1.074). (G) Wear failure summary (counts across WEAR1–3): total failure (fully off) and partial failure (lifting) for DT and TT; y-axis shows counts (0–24). The 24 count arises from 4 participants × 3 APP × 2 sensors/application. For full figure, * indicates p < 0.05, ** indicates p < 0.01, *** indicates p < 0.001. Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/cndws0y and https://biorender.com/xn0edbu (accessed on 19 June 2026).
Figure 6 User comfort study protocol and comparative performance of two sensor-attachment methods (DT, blue vs. TT, orange). (A) Study timeline. Participants completed an allergy check, experimenter-guided APP0/REM0, then three timed cycles of application and removal without any experimenter feedback (APP1–3, REM1–3), indicated by the green clock icon. Fine motor tasks (pinch/grasp, writing, and typing) and daily-living tasks (eating, phone use, hand sanitizer application, and jacket on/off) followed APP2; APP2/REM2 were performed with finger splints to limit finger range of motion. A ~60 min free-time block preceded REM3. Cycle-specific questions followed each APP/REM cycle (indicated by the clipboard icon); final open-ended questions were asked prior to concluding the session. (B) Application time by method and cycle with participant traces in gray and mixed-effects learning lines in black. DT and TT both get significantly faster with each application (linear mixed-effects model, random intercept by participant, DT: slope = −39.5 s per application, 95% CI [−74.2, −4.78], t = −2.23, p = 0.026, R2 marginal = 0.092, R2 conditional = 0.796, n = 12, TT: slope = −60.8 s per application, 95% CI [−99.5, −22.0], t = −3.07, p = 0.0021, R2 marginal = 0.143, R2 conditional = 0.833, n = 12). (C) Pooled over APP1-3, TT applications were significantly faster than DT applications (difference between mean application times with DT and TT was 130.6 s, 95% CI [−180.06, −81.11], paired, two-sided t-test, t = −5.810, p = 0.0001, n = 12 paired observations from 4 participants × 3 APP, Cohen’s d = −1.677, Hedges’ g = −1.560). (D) Removal time by method and cycle with participant traces in gray and mixed-effects learning lines in black. DT did not significantly change in removal time but TT significantly improved in speed (linear mixed-effects model, random intercept by participant, DT: 95% CI on slope [−60.48, 7.98], t = −1.503, p = 0.133, R2 marginal = 0.036, R2 conditional = 0.824, n = 12, TT slope: −44.4 s per removal, 95% CI [−87.23, −1.52], t = −2.030, p = 0.0424, R2 marginal = 0.070, R2 conditional = 0.812, n = 12). (E) Pooled over REM1–3, TT removals were significantly faster than DT removals (difference between mean removal times between DT and TT of 157.9 s, 95% CI [−226.60, −89.23], paired, two-sided t-test, t = −5.060, p = 0.0004; n = 12 paired observations = 4 participants × 3 REM, Cohen’s d = −1.461, Hedges’ g = −1.359). (F) Survey ratings (1–10; paired, two-sided t-tests; n = 12 pairs/metric) favor TT on ease (+0.75, 95% CI [0.33, 1.17], t = 3.954, p = 0.0023, Cohen’s d = 1.141, Hedges’ g = 1.062), stability (+1.33, 95% CI [0.79, 1.88], t = 5.360, p = 0.00023, Cohen’s d = 1.547, Hedges’ g = 1.439), and adhesion (+1.33, 95% CI [0.60, 2.07], t = 4.000, p = 0.0021, Cohen’s d = 1.155, Hedges’ g = 1.074). (G) Wear failure summary (counts across WEAR1–3): total failure (fully off) and partial failure (lifting) for DT and TT; y-axis shows counts (0–24). The 24 count arises from 4 participants × 3 APP × 2 sensors/application. For full figure, * indicates p < 0.05, ** indicates p < 0.01, *** indicates p < 0.001. Partially created in BioRender. Schlegel, H. (2026). https://biorender.com/cndws0y and https://biorender.com/xn0edbu (accessed on 19 June 2026).CC BY · PubMed Central
Figure 7 Long-wear recording of the hand kinematics system during daily activities. (A) Representative segment of the wear session, with shaded regions indicating labeled activity periods and calibration blocks. (B–G) Representative 20 s windows are shown for walking, texting, typing, resting, piano, and eating, with the wrist flexion/extension (top), index MCP bend angle (middle), and thumb MP bend angle (bottom). (B) During walking, nothing was held in the hand. (C) During texting, the phone was held between the index finger and thumb while the wrist and arm remained relatively stationary. (D) Typing was performed using a low-profile keyboard. (E) During resting, the hand remained stationary on the desk. (F) The piano segment reflects a beginner pianist, and (G) the eating segment includes a mixture of typical eating-related movements. For this participant, the plotted y-axis ranges were selected to reflect a comfortable range of motion during natural activities: approximately −60 to +70° for wrist flexion/extension, 0 to 90° for index MCP, and 0 to 20° for thumb MP. Index MCP and thumb MP bend angles were estimated from sensor ADC values using the first calibration block collected immediately after device application.
Figure 7 Long-wear recording of the hand kinematics system during daily activities. (A) Representative segment of the wear session, with shaded regions indicating labeled activity periods and calibration blocks. (B–G) Representative 20 s windows are shown for walking, texting, typing, resting, piano, and eating, with the wrist flexion/extension (top), index MCP bend angle (middle), and thumb MP bend angle (bottom). (B) During walking, nothing was held in the hand. (C) During texting, the phone was held between the index finger and thumb while the wrist and arm remained relatively stationary. (D) Typing was performed using a low-profile keyboard. (E) During resting, the hand remained stationary on the desk. (F) The piano segment reflects a beginner pianist, and (G) the eating segment includes a mixture of typical eating-related movements. For this participant, the plotted y-axis ranges were selected to reflect a comfortable range of motion during natural activities: approximately −60 to +70° for wrist flexion/extension, 0 to 90° for index MCP, and 0 to 20° for thumb MP. Index MCP and thumb MP bend angles were estimated from sensor ADC values using the first calibration block collected immediately after device application.CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 2 Not valid · 28 Not checked · 144 Own results and interpretations

63 more claims

Create a free account to see all 174 claims and to verify the 50 references still open.

See all claims — free account
  • It may state two separate claims as one. 71
  • The quote only partly states this claim. 2
  • Probably not a substantial claim. 51

Create a free account to see these claims.

Statistics recomputed

Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.

  • ConsistentF(4, 8) = 1.41reported p = .31 · recomputed p = .314

    “F(4, 8) = 1.41, p = .31”

  • Consistentt(2) = 6.72reported p = .02 · recomputed p = .021

    “t(2) = 6.72, p = .02”

  • Consistentt(2) = -0.82reported p = .5 · recomputed p = .498

    “t(2) = −0.82, p = .50”

  • Consistentt(2) = 1.29reported p = .33 · recomputed p = .326

    “t(2) = 1.29, p = .33”

  • Consistentt(2) = 1.2reported p = .35 · recomputed p = .353

    “t(2) = 1.20, p = .35”

5 of 6 · the rest aren’t loaded

1 more recomputed test

Create a free account to see every test the Stats check recomputed.

See all tests — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

3 of 3

Citations

Cites

Research works printed but not linked: 1

  1. Kuipers, J.B. (2007). Quaternions and Rotation Sequences: A Primer with Applications to Orbits, Aerospace and Virtual Reality, Princeton University Press.

Cited by

Still being checked.

Metadata

Authors
Oppenheim, T., Schlegel, H., Xie, P. Y., Khokhar, Z., & Khanna, P.
Affiliations
  • California Maritime Academy (US)
  • University of California, Berkeley (US)
Corresponding author
Khanna, Preeya
Venue
Sensors
Publisher
MDPI AG
Published
2026-07-28
Volume
26
Issue
15
Pages
4795
ISSN
1424-8220 (ISSN-L)
OpenAlex
W7171515028
Cited by
0
References
48
FWCI
0.00
Type
journal article
Language
en
Open access
Locations
  • Sensors (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
  • Eunice Kennedy Shriver National Institute of Child Health & Human Development of the National Institutes of Health (P2CHD101899)
  • Weill Neurohub Next Great Ideas Program (1DP2NS142717-01)
Topics
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Context-Aware Activity Recognition Systems
  • Physical Sciences › Engineering › Biomedical Engineering › Advanced Sensor and Energy Harvesting Materials
  • Health Sciences › Medicine › Rehabilitation › Stroke Rehabilitation and Recovery
Keywords
wearable sensors, hand kinematics, electronic skin, inertial measurement unit, stroke rehabilitation, neurorehabilitation, low-cost wearable device, donning and doffing, user feedback
MeSH terms
Wearable Electronic Devices, Motion Capture, Biomechanical Phenomena, Hand, Humans, Movement

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-18, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-18 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.