A. Participants
This study was conducted as a randomized controlled trial to determine the effect of wearing an activity tracking and feedback device on affected arm use during a 6-week intervention at home (20). Forty-two participants were recruited, all of whom had a unilateral ischemic or hemorrhagic stroke with residual hemiparesis in the upper limb. Participants were included if they were older than 18 years, had no known intolerance to the material of the activity tracker, had no severely impaired sensation of the wrist, depression, major cognitive impairment, major comorbidities, or comprehensive aphasia. The study was approved by the Cantonal Ethics Committee Zurich and Cantonal Ethics Committee Northwest and Central Switzerland (BASEC-number 2017 − 00948) and was registered at https://clinicaltrials.gov, unique identifier NCT03294187, before recruitment. For more information on the study protocol, please refer to (20).
B. Experimental Procedure
After obtaining a baseline assessment with a battery of clinical tests to assess the initial upper limb impairment, participants were randomly allocated to the intervention (n = 19) or control group (n = 23) (20). Randomization was stratified based on the impairment level of the upper limb (< 32 and ≥ 32 of Fugl-Meyer Assessment – Upper Extremity Subscale (FMA-UE) score). To continuously monitor arm activity, participants in both groups were equipped with an "ARYS™ pro|tracker" (Tyromotion GmbH, Austria, formerly yband therapy AG, Switzerland) for their unaffected wrist and the "ARYS™ me|tracker" on the affected wrist, which could trigger vibrotactile reminders (20) (Fig. 1). Participants were asked to wear both devices for as long as possible every day of the 6-week intervention period. Additionally, the intervention group received two forms of feedback: participants received a phone application with visual feedback about arm use, allowing them to track their arm activity levels and whether arm activity deviated from the trajectory needed to achieve their daily goal by the end of the day (Fig. 1). Whenever participants were inactive and not on track to reach their arm activity goal, smart reminders such as vibrotactile (vibrating pulses) and visual cues (LED lights on the tracker) were triggered on the affected side. The control group did not receive any feedback or reminders (20, 29).
C. Outcome measures
The primary outcome to evaluate the effectiveness of the intervention was the patient-reported paretic upper limb use in daily life, measured with the Motor Activity Log-14 Item Version, Amount of Use subscale (MAL-14 AOU) (39, 40). Secondary outcomes were arm-use metrics derived from the sensor data collected during the 6-week intervention and the following clinical assessments: MAL-14 Quality of Movement subscale (MAL-14 QOM) (39, 40), FMA-UE (41), Action Research Arm Test (ARAT) (42), EuroQol Five Dimensions Five Levels questionnaire (EQ-5D-5L) (43), modified Rankin Scale (mRS) (44). Clinical assessments were recorded at baseline, after the 6-week post-intervention, and after the 6-week follow-up. The Global Rating of Perceived Changes (GRPC) assesses how the affected arm use and physical activity changes are perceived in everyday life at post-intervention and follow-up. For the EQ-5D-5L, the index value and visual analog scale (VAS) results were extracted (45). A more detailed description of the clinical outcome parameters and their properties can be found in the published clinical study protocol (20).
Raw sensor data were processed into arm activity counts from the affected and unaffected sides using a threshold-based approach. Arm activity was considered true for each minute where the aggregated raw three-axis accelerometer data exceeded a threshold of 0.1 g acceleration after gravity subtraction (20, 29). Three arm-use metrics were derived from the arm activity counts to characterize different aspects of affected arm activity. Affected Arm Use (AAU) represents the magnitude of total daily activity counts. It was calculated as the cumulative activity counts of the affected arm. Arm Use Ratio (AUR) represents the utilization ratio of the affected in relation to the unaffected side. It was calculated as the ratio of cumulants between the affected and the unaffected sides. Percentage of Time Active (PTA) represents the percentage of active time throughout the day. It was calculated as the cumulative duration that the affected arm activity was considered true within one day. These three metrics were computed for each day between 6:00 a.m. and 11:59 p.m., as this interval aligns with the assumed waking hours of participants, during which they were expected to be actively engaged in daily activities. An average of AAU, AUR, and PTA was taken each week to evaluate the change over the six-week study duration. A linear regression model was fitted on the available data to account for missing data and applied to intra- and extrapolate arm activity counts. Participants were excluded from the sensor data analysis if more than 90% of the dataset of either the affected or unaffected side was missing, if no data from the first week of data collection were available, or if activity data on each side were collected on less than 70% of the study duration days, reducing the accuracy of intra- and extrapolation of activity counts.
D. Statistical analysis
All available clinical data from enrolled participants was analyzed using an intention-to-treat approach. Data was tested for normality using the Shapiro-Wilk test. Mixed linear models (MLM), with a random intercept, were used to identify differences between the intervention and control group and evaluate changes in clinical scores from baseline to post-intervention to follow-up. The scores of the clinical scales (MAL-14 AOM, MAL-14 QOM, FMA-UE, ARAT, EQ-5D-5L, mRS) were used as input for the dependent variable. Timepoints (baseline, post-intervention, and follow-up) and group allocations (intervention and control) were set as the covariates.
Subgroup analysis (per protocol analysis) was performed for the sensor data to account for incomplete datasets. The impact of the intervention on the amount of arm activity behavior over the six-week study duration was investigated with MLM. The MLM included (1) AAU, (2) AUR, (3) PTA, (4) MAL-14 AOM, and (5) MAL-14 QOM as the dependent variable. Group allocation (intervention and control), week of the study duration, and impairment (mild, moderate, and severe) were set as covariates for each model. Based on the baseline FMA-UE assessment, three impairment levels were defined: mild impairment with a score ranging from 43 to 66, moderate impairment ranging from 29 to 42, and severe impairment ranging from 0 to 28 (46). A logarithmic transformation was applied in the case of inadequate model fits. The variance inflation factors were calculated to test for the multicollinearity of the predictor variables. Model performance was evaluated by calculating the R2. The Tukey honest significant difference (HSD) posthoc test was used to explore if there were differences concerning the metrics and the covariates. The distribution of the residuals was compared to the normal distribution using a kernel density estimate (KDE) and a Q-Q plot. Additionally, homoskedasticity was checked with a residual versus fitted values (RVF) plot.
Finally, the Pearson correlation coefficient was applied to examine the association between upper limb impairment level and daily AAU, AUR, and PTA averaged over the first week for the intervention and control group separately. A weak correlation was defined with a rho between 0.2–0.39, a moderate correlation between 0.40-0-59, a strong correlation between 0.6–0.79, and a very strong correlation between 0.80-1 (47). Fisher's z-transformation was applied to investigate differences in the correlation coefficients of the two groups. All analyses (post-processing and statistical) were performed using Python (Version 3.8, Python Software Foundation, packages: scipy.stats, sklearn, statsmodels). A two-sided significance level of alpha = 0.05 was used.