Content uploaded by Paolo Bonato
Author content
All content in this area was uploaded by Paolo Bonato
Content may be subject to copyright.
Using Wearable Sensors to Measure Motor Abilities following Stroke
Todd Hester
1
, Richard Hughes
1
, Delsey M. Sherrill
1
, Bethany Knorr
2
,
Metin Akay
3
, Joel Stein
1
, and Paolo Bonato
1,4
1
Dept of PM&R, Harvard Medical School, Spaulding Rehabilitation Hospital, Boston MA
2
Thayer School of Engineering, Dartmouth College, Hanover NH
3
Dept of Bioengineering, Arizona State University, Tempe AZ
4
The Harvard-MIT Division of Health Science and Technology, Cambridge MA
tahester@partners.org, rhughes1@partners.org, delsey.sherrill@gmail.com,
bethany.r.knorr@dartmouth.edu
, metin.akay@asu.edu, jstein@partners.org,
pbonato@partners.org
Abstract
Motor abilities of stroke survivors are often
severely affected. Post-stroke rehabilitation is guided
by the use of clinical assessments of motor abilities.
Clinical assessment scores can be predicted by models
based on features extracted from the wearable sensor
data. Wearable sensors would allow monitoring of
subjects in the home and provide accurate assessments
to guide the rehabilitation process. We propose the use
of a wearable sensor system to assess the motor
abilities of stroke victims. Preliminary results from
twelve subjects show the ability of this system to
predict clinical scores of motor abilities.
Keywords: Wearable Sensors, Stroke, Clinical
Assessment
1. Introduction
Approximately 700,000 people are affected by
stroke each year in the United States and about
275,000 die from stroke each year [1]. Strokes affect a
person’s cognitive, language, perceptual, sensory, and
motor abilities [2]. More than 1,100,000 Americans
have reported difficulties with functional limitations
following stroke [3]. Recovery from stroke is a long
process that continues beyond the hospital stay and
into the home setting. The rehabilitation process is
guided by clinical assessments of motor abilities.
Accurate assessment of motor abilities is important
in selecting the best therapies for stroke survivors.
These assessments are based on observations of
subjects’ motor behavior using standardized clinical
rating scales. The accuracy and consistency of
observational assessments may vary greatly across
clinicians [4]. Wearable sensors could be used to
provide more accurate measures or could be used in
addition to observational clinical tools. Wearable
systems have the ability to measure motor behavior at
home and for longer periods than could be observed in
a clinical setting. Accelerometers can capture specific
patterns of movement relating to motor disabilities. We
propose that wearable systems can be used to predict
clinical scores of motor abilities and we present an
initial analysis of data demonstrating an association
between accelerometer data and clinical scores.
2. Methods
Twelve subjects who had a stroke within the past 2
to 24 months were recruited for the study. Each subject
was evaluated by a clinician using standardized clinical
motor performance scales, including the Fugl-Meyer
Assessment of Sensorimotor Recovery after Stroke,
Chedoke-McMaster Stroke Assessment, Wolf Motor
Performance Test, and the Reaching Performance
Scale. These scales measure dimensions of upper limb
motor behavior including movement quality, stage of
motor recovery, use of compensatory movement
strategies, and the ability to perform functional tasks.
All testing was performed at Spaulding Rehabilitation
Hospital. Subjects provided informed consent
approved by the hospital’s research review board.
Accelerometers were attached to the affected arm and
the trunk (Figure 1).
Sensor data was recorded using the Vitaport 3
(Temec BV, The Netherlands) ambulatory recorder,
which was worn on the waist. Subjects performed
multiple repetitions of tasks requiring reaching and
prehension, selected from the clinical scales. The tasks
included reaching to close and distant objects, placing
the hand or forearm from lap to a table, pushing and
pulling a weight across a table, drinking from a
beverage can, lifting a pencil, flipping a card, and
turning a key.
The accelerometer data was digitally low-pass
filtered in Matlab with a cutoff frequency of 15 Hz to
remove high frequency noise. Both this low-pass
filtered and a high-pass filtered version of the data
were utilized in the analysis. The high-pass filtered
version of the data was derived in an attempt to isolate
actual acceleration components from gross postural
adjustments. A 1 Hz cut-off frequency was used.
The signals were marked manually during testing
using a 5 V pulse and marks were checked via visual
inspection of the data. The marks were used to
segment the data by task through an automated
software procedure based on threshold detection of the
manual markings. Manipulation tasks such as card
flipping were also segmented within the task using a
touch sensor. They were segmented into a reaching
epoch, a manipulation epoch, and a release/return
epoch. Subjects performed between 10 and 20
repetitions of each task, resulting in an average of 109
segments for each subject. The epochs ranged from
0.30 s to 23.3 s, with a mean length of 2.54 s and a
standard deviation of 1.90 s. The following features
were extracted from each epoch of accelerometer data
for later analysis:
• Mean value of the low-pass filtered data
• Root-Mean-Square value (this feature and the
following ones were all derived from the
high-pass filtered version of the accelerometer
data)
• Dominant frequency
• Ratio of energy in 0.2 Hz bin around the
dominant frequency to total energy (measure
of periodicity)
• Range of autocovariance
• Root-Mean-Square value of the derivative of
acceleration (i.e. jerk time series)
• Dominant frequency of the jerk time series
• Ratio of energy in 0.2 Hz bin around the
dominant frequency of jerk to total energy
(measure of periodicity)
• Peak velocity
• Jerk metric (i.e. the RMS jerk normalized by
the peak velocity
• Approximate entropy (nonlinear measure of
complexity)
• Correlation at zero lag between selected pairs
of accelerometer time series
• Peak correlation within a 1 s window between
selected pairs of accelerometer time series
• Lag time of the peak correlation between
selected pairs of accelerometer time series
Features from each task were imported into the
Waikato Environment for Knowledge Analysis
(WEKA) for exploratory analysis [5]. Initially, we
used WEKA to look at scatterplots of features and
clinical scores to assess the suitability of the current
sensors, tasks, and features to predict the subjects’
clinical scores. Next, we built linear regression models
in WEKA to explore their ability to predict the clinical
scores as well as to examine which feature sets were
useful in predicting the scores. Features for the linear
regression models were selected by the M5 method,
which performs a backward stepwise regression using
the Akaike information criterion [6]. The models
provided feature sets for each clinical score. Then,
linear regression models were built in Matlab to
predict clinical scores using a leave-one-subject-out
method (i.e. clinical scores for each subject were
predicted based on a model built with data from all
other subjects). All of the features were normalized to
have a mean of zero and a standard deviation of one.
Table 1 lists the clinical scores predicted by the model.
Scores were predicted based on analysis of features
from 16 different segments from 8 tasks. We looked at
the forearm to table, hand to table, pushing a weight,
and retrieving a weight tasks in their entirety and at the
Figure 1 Sensor setup and orientation of the axes
of the accelerometers.
can lifting, pencil lifting, card-flipping, and key-
turning tasks in three segments each.
3. Results
Figure 2 shows a scatterplot of the peak correlation
within a 1 s window between the index finger and
hand accelerometers and the peak velocity of the
thumb accelerometer during the manipulation epoch of
the can lifting task. The colors represent the scores on
the Chedoke-McMaster Hand Stage. The figure shows
that higher peak velocities for the thumb accelerometer
data correlate to higher clinical scores, while higher
correlations between the index finger and hand data
correlate with lower scores.
Models predicting all seven clinical scores with
features from all 16 different task segments were built.
Table 1 shows the root mean square error of the best
model for each clinical score along with the range for
each clinical score. The models were most successful
in predicting the Chedoke-McMaster Hand Stage and
the shoulder and elbow portion of the Fugl-Meyer
scale, with relative errors close to 10%. The models
were less successful in predicting other clinical scores.
Table 2 shows the RMS error of the prediction of
the shoulder and elbow portion of the Fugl-Meyer
scale and the Chedoke-McMaster Hand Stage using
each task. The best predictors of the Hand Stage were
models built with features extracted from the forearm
to table task and the manipulation segments of the can
lifting and card flipping tasks. The best predictors of
the shoulder and elbow portion of the Fugl-Meyer
were the “reaching” segments of the manipulation
tasks. The worst predictor of both clinical scores was
the “releasing” segment of the pencil lifting task.
Table 1. Errors in predicting clinical scores
Clinical Score
RMS
Error
Score
Range
Range for
Subjects
Tested
Chedoke-
McMaster Hand
Stage
0.42 1-7 3-5
Chedoke-
McMaster Arm
Stage
1.27 1-7 3-7
Wolf Test –
Median Time
1.30 0-120 1.57-9.66
Fugl-Meyer
Shoulder-Elbow
2.35 0-30 19-30
Fugl-Meyer
Shoulder-Elbow
3.32 0-24 4-22
Fugl-Meyer Total
Score
10.01 0-66 29-63
Table 2. RMS Errors in predicting clinical scores for
different tasks.
Task
Chedoke-
McMaster
Hand Stage
Fugl-Meyer
Shoulder-
Elbow
Can - Segment 1 0.58 3.66
Can - Segment 2 0.50 6.02
Can - Segment 3 0.67 5.13
Card - Segment 1 0.58 2.35
Card - Segment 2 0.50 3.74
Card - Segment 3 0.67 5.23
Forearm To Table 0.67 4.90
Hand To Table 0.42 6.23
Key - Segment 1 0.86 2.99
Key - Segment 2 0.90 3.75
Key - Segment 3 0.67 4.80
Pencil - Segment 1 0.62 3.10
Pencil - Segment 2 0.67 5.22
Pencil - Segment 3 0.93 7.76
Push Weight 0.71 6.19
Retrieve Weight 0.74 5.16
Figure 2 Scatterplot of peak correlation between
accelerometer data from index finger and hand and
peak velocity derived from thumb accelerometer
data in comparison to Chedoke-McMaster Hand
Stage scores. A line separates well samples
associated with a score of 3 from samples
associated with a score of 5. Samples for a score o
f
4 are in between and overlap with the rest of the
data.
Table 3 shows the linear regression model used to
predict the shoulder and elbow portion of the Fugl-
Meyer score using features from the reaching segment
of the card flipping task. Table 4 shows the actual
scores, predicted scores, and the standard deviation of
the predicted scores for the shoulder and elbow portion
of the Fugl-Meyer scale using the model shown in
Table 3. Five of the predicted scores were within 1
point of the actual score, and the closest was within
0.02. The worst prediction was 35.81 for a subject with
a clinical score of 30.
Table 3. Linear regression model for prediction of the
shoulder and elbow portion of the Fugl-Meyer score
based on the reaching segment of the card flipping task
Coefficient Feature
1.18 * Mean of Forearm X acc
1.81 * Mean of Forearm Y acc
0.65 * Mean of Upper Arm X acc
1.50 * Mean of Upper Arm Y acc
-0.86 * Mean of sternum acc
0.99 * RMS of forearm X acc
1.66 * RMS of forearm Y acc
0.31 * RMS of Upper Arm X acc
-1.31 * RMS of Upper Arm Y acc
-0.16 * RMS of sternum acc
-0.28 * Peak Corr. of Thumb and Hand
Table 4. Scores for the shoulder and elbow portion of
the Fugl-Meyer score
Subject
Actual
Score
Predicted
Score
STD of
Prediction
A 23 22.15 0.86
B 26 26.50 1.40
C 19 20.66 0.40
D 26 22.00 0.63
E 30 35.81 2.47
F 30 29.98 2.83
G 24 24.97 0.18
H 24 22.63 0.99
I 30 26.04 1.57
J 30 28.79 0.75
K 20 21.16 1.11
L 27 27.40 1.38
4. Discussion and Conclusion
The results of the linear regression models have
been promising so far. Our models predicted two
clinical scores within 10% of the average score. Our
sensor system showed the ability to detect specific
movement patterns related to clinical scores of
movement ability. For example, a negative coefficient
was associated with the root mean square value of the
sternum accelerometer channel in most of the linear
regression models, indicating that the method was able
to detect compensatory trunk movements related to
lower clinical scores. The jerk metric and the root
mean square of the jerk time series usually had a high
coefficient in the models, showing that the system
determined that smooth movement was significantly
related to the clinical scores. Scores such as the median
time on the Wolf Motor Function Test may be difficult
to predict because of non-linear relationships between
post-stroke motor ability and performance scores on
this test. To predict these scores more accurately, it
may be necessary to include non-linear parameters or
create a non-linear model to predict the scores. We are
currently collecting data from more subjects, which
will allow us to improve the linear regression models
and explore the use of nonlinear models. The small
size of the current dataset limits the number of features
we can use in the models. Including more features is
expected to improve the model. Collecting data from
subjects with a wider range of motor abilities and
clinical scores will allow us to develop a more accurate
linear regression model.
Acknowledgments
This study was supported by the grant entitled
"Field Measures of Functional Tasks for CIT
Intervention", #R21HD045873-01, NIH-NICHD.
References
[1] American Heart Association, Heart Disease and Stroke
Statistics – 2005 Update, American Heart Association,
Dallas, TX, 2004.
[2] National Institute of Neurological Disorders and Stroke
(NINDS), Stroke: Hope Through Research, NINDS,
Washington, DC, 2004.
[3] Center for Disease Control, Morbidity and Mortality
Weekly Report, CDC, Atlanta, GA, 2001.
[4] V. Pomeroy, A. Pramanik, L. Sykes, J. Richards, and E.
Hill, “Agreement between physiotherapists on quality
of movement rated via videotape”, Clinical
Rehabilitation, May 2003, pp. 264-272.
[5] IH Witten and E. Frank, Data Mining: Practical
machine learning tools and techniques. Morgan
Kaufman, San Francisco, 2005.
[6] H Akaike, “A New Look at Statistical Model
Identification." IEEE Transactions on Automatic
Control, IEEE, 1974, pp. 716-722.