Participants
In
the present study, we aimed to include at least 27 participants since
this sample size would enable us to detect a medium effect (Cohen’s
d = 0.50) in the change in arm use with an alpha of 0.05 and a power of
0.80. We included people entering Rijndam Rehabilitation (Rotterdam, The
Netherlands) after an ischemic or hemorrhagic stroke that suffered from
a paretic arm or leg (defined as National Institutes of Health Stroke
Scale (NIHSS) 5 A/B or 6 A/B 4 ≥ score > 0). They had to be (1) 18
years or older, (2) had a Mini-Mental State Examination (MMSE) > 19,
and (3) were able to sit at least 30 min with back support. We excluded
patients who were more than 3 weeks poststroke when admitted to Rijndam
Rehabilitation. Participants were screened by a researcher between
September 2016 and September 2018. All participants gave their written
informed consent and the study was approved by the Medical Ethics
Committee of Erasmus MC University Medical Center Rotterdam, The
Netherlands (MEC-2015-687).
Procedures
At
the start of the study (week 3 poststroke), all participants were
inpatient at Rijndam Rehabilitation, where they received usual care for
people after stroke. The usual care program for arm rehabilitation at
Rijndam Rehabilitation is based on the principles of the Concise Arm and
Hand Rehabilitation Approach in Stroke (CARAS) [3, 8]. The amount and the content of the rehabilitation program were not adapted for this study.
A
researcher performed arm use and arm function assessments at 3 weeks,
12 weeks, and 26 weeks poststroke. In addition, the same researcher
evaluated stroke severity (National Institutes of Health Stroke Scale
(NIHSS) [8, 9])
and collected demographic data including age, gender, affected
bodyside, dominant bodyside, admission to the rehabilitation clinic in
weeks poststroke, discharge from the rehabilitation clinic in weeks
poststroke. Due to individual differences in the usual care, some
participants were still at the rehabilitation center at week 12, while
at week 26 all participants were at home and were visited by the same
researcher for the assessments.
Arm recovery assessments
The
Fugl-Meyer Upper Exterimity assessment (FMUE) was used to measure arm
impairments (where impairments refers to a loss of body function and
structure) [10, 11].
The FMUE consists of nine components examining voluntary movements and
the ability to execute arm movements outside of synergies. The score of
the FMUE ranges from 0 to 66, with higher scores indicating a better
motor function.
To define arm recovery clusters, we used the classification model of Van der Vliet et al. [12].
This recent study found that different arm recovery clusters exist
during the first 6 months poststroke - each with a specific recovery
profile - and that recovery cluster belonging can be well-predicted
early poststroke. In the present study, we used an online available
application (https://emcbiostatistics.shinyapps.io/LongitudinalMixtureModelFMUE/) that implements the model developed by Van der Vliet et al. [12]
for the prediction of arm recovery after stroke. For each individual
patient, we entered the FMUE data available from week 3, 12 and 26 in
the application to identify arm recovery cluster belonging. The model
identifies arm recovery cluster belonging as poor, moderate, and
excellent based on the initial FMUE after stroke and the amount and rate
of recovery in FMUE score poststroke.
Arm use assessments
We applied an arm use monitor that was developed and validated for the measurement of arm use in stroke survivors [7, 13].
The system is based on the assumption that voluntary arm use is related
to arm movement during sitting and standing rather than during
whole-body movements such as walking. Arm use is measured by recording
arm movement intensity during sitting and standing, thereby avoiding the
influence of whole-body movements. The arm use monitor consists of
three accelerometers (Activ8 Activity Monitor, Activ8): one attached to
the front of the nonaffected thigh to detect body postures/movements
(lying/sitting, standing, walking, cycling, running), and one attached
to each wrist to measure arm movement intensity (Fig. 1).
Each Activ8 accelerometer (30 × 32 × 10 mm, 20 g) measures raw
acceleration data with a sample frequency of 12.5 Hz, filters the
acceleration data with an exponential moving average filter, and
converts these data with a resolution of 1.6 Hz to body
postures/movements and movement counts (a commonly used measure of
movement intensity) where 1 count is equal to 0.01 g (1 g = 9.81 m/s2).
The device stores the data in epochs of 30 s—with 48 samples per
epoch—and sums the movement counts per epoch. In the present study we
examined the number of samples per epoch to ensure that the samples are
equal across epochs. A previous study showed that an Activ8 sensor on
the upper thigh provides an accurate detection of body
postures/movements in stroke survivors (82–100 % accuracy) [13].
Participants were asked to wear the arm use monitor
for 1 week (seven days) at three timepoints: week 3, week 12, and week
26 poststroke. The sensors on the wrists were attached with watch-type
wristbands and were taken off during the night and during water
activities such as showering. The sensor on the nonaffected leg was
attached with water-resistant, anti-allergic skin tape, and was worn
seven days continuously. During each one week measurement period, the
data were stored locally on the sensor devices. After the 1 week
measurement, a researcher downloaded the data of the three Activ8
sensors on a PC for data processing and analysis.
All data analysis was performed in R [14]
using RStudio (version 1.2.50001, RStudio, Inc.) and a custom-made
script based on the study of Fanchamps et al. (2018) that developed and
validated the arm use monitor [7].
The first step in the algorithm was to synchronize the Activ8 sensors
based on the timestamps within the data files. Then, the measurement
period was selected. Only waking hours were analyzed, for which we
selected 7 am to 10 pm. Within this period, nonwear of the wrist sensors
was detected when at least one device measured zero movement counts for
at least one hour. Data were used for analysis when participants had at
least two valid days in a measurement week, with a valid day defined as
at least ten hours of data without nonwear. In the next step, 30-second
epochs were selected in which the posture was sitting or standing
according to data of the leg sensor. An epoch was classified as
sitting/standing when at least 90 % of the 48 samples were classified as
sitting or standing. For each 30-second epoch classified as
sitting/standing, arm use was estimated by calculating the total
movement counts per wrist-worn sensor. Next, the following arm use
outcome measures were calculated per valid day: (1) the total daily
movement counts of the affected arm—a measure of the amount of arm
movement—during sitting and standing, (2) the total daily movement
counts of the nonaffected arm during sitting and standing, (3) the ratio
between the total daily use of both arms, calculated as the total daily
movement counts of the affected arm during sitting and standing divided
by the total daily movement counts of the nonaffected arm during
sitting and standing, (4) the mean movement counts of the affected arm
per sit/stand hour, and (5) the mean movement counts of the nonaffected
arm per sit/stand hour. Finally, per measurement week, a mean daily
value was calculated for each arm use outcome measure by averaging
across valid days.
Statistical analysis
All statistical analyses were performed in R [14]
using RStudio (version 1.2.50001, RStudio, Inc.). Characteristics of
the study participants are described as mean ± SD with minimum and
maximum values. Before conducting the statistical analyses, we
determined the distribution of the data based on visualizations of the
data and normality tests. We used generalized estimating equation (GEE)
to investigate how arm use changes over time and to compare the change
in arm use between arm recovery clusters. GEE takes into account the
dependence between repeated measurements within subjects and can deal
with missing data as well as nonnormal distributed data [15].
We developed GEE models for different dependent variables: total daily
affected arm use, total daily nonaffected arm use, daily ratio between
arms, affected arm use per sitting and standing hour, nonaffected arm
use per sitting and standing hour, daily duration of sitting and
standing, daily walking duration, daily wearing time of the arm use
monitor. To investigate how arm use changes over time, we only included
time as factor (three levels: 3, 12 and 26 weeks). To compare the change
in arm use between arm recovery clusters, we included time, recovery
cluster (two levels: poor/moderate and excellent), and the interaction
time × recovery cluster as factors. For the development of GEE models,
we used the Generalized Estimating Equation package (‘geepack’ package) [16]
and set the distribution of the data at ‘gaussian’ and the correlation
structure at ‘exchangeable’. A p-value below 0.05 was considered
statistically significant. For significant effects in the GEE models, we
performed posthoc comparisons with a Bonferroni correction using the
Estimated Marginal Means package (‘emmeans’ package) [17].
Change percentages between time points (3, 12 and 26 weeks poststroke)
were calculated as: (new value – previous value) / previous value ×
100 %.
Peer-Reviewed
