- Email: [email protected]
Estimation of Energy Expenditure for Wheelchair Users Using a Physical Activity Monitoring System
Accepted Manuscript Estimation of Energy Expenditure for Wheelchair Users using a Physical Activity Monitoring System Shivayogi V. Hiremath, PhD, Stephen S. Intille, PhD, Annmarie Kelleher, MS, Rory A. Cooper, PhD, Dan Ding, PhD PII:
S0003-9993(16)00155-6
DOI:
10.1016/j.apmr.2016.02.016
Reference:
YAPMR 56471
To appear in:
ARCHIVES OF PHYSICAL MEDICINE AND REHABILITATION
Received Date: 8 December 2015 Revised Date:
25 January 2016
Accepted Date: 22 February 2016
Please cite this article as: Hiremath SV, Intille SS, Kelleher A, Cooper RA, Ding D, Estimation of Energy Expenditure for Wheelchair Users using a Physical Activity Monitoring System, ARCHIVES OF PHYSICAL MEDICINE AND REHABILITATION (2016), doi: 10.1016/j.apmr.2016.02.016. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Title Page Running Head: Energy Expenditure in Wheelchair users
RI PT
Title: Estimation of Energy Expenditure for Wheelchair Users using a Physical Activity Monitoring System
Rory A. Cooper, PhD1-2,7, Dan Ding PhD,1-2,7
Affiliations:
M AN U
1-2,
SC
Authors: Shivayogi V. Hiremath, PhD,1-4*, Stephen S. Intille, PhD,5-6, Annmarie Kelleher, MS,
1) Department of Veterans Affairs, Human Engineering Research Laboratories, VA Pittsburgh Healthcare System, Pittsburgh, PA
PA
TE D
2) Department of Rehabilitation Science and Technology, University of Pittsburgh, Pittsburgh,
3) Department of Physical Medicine and Rehabilitation, University of Pittsburgh, Pittsburgh,
EP
PA
4) Department of Physical Therapy, Temple University, Philadelphia, PA
AC C
5) College of Computer and Information Science, Northeastern University, Boston, MA 6) Department of Health Sciences, Northeastern University, Boston, MA 7) Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA
* This research study was executed at the Human Engineering Research Laboratories, VA Pittsburgh Healthcare System and Department of Rehabilitation Science and Technology,
ACCEPTED MANUSCRIPT
University of Pittsburgh.
Acknowledgements:
RI PT
The work is supported by the Department of Defense (W81XWH-10-1-0816). SVH’s work in this article was funded through the Switzer Research Fellowship (H133F110032) awarded by the National Institute on Disability and Rehabilitation Research, Department of Education.
SC
The work is also supported by the Human Engineering Research Laboratories, VA Pittsburgh Healthcare System. The contents do not represent the views of the Department of Veterans
M AN U
Affairs or the United States Government. The authors thank their colleagues at the Human Engineering Research Laboratories for their input and effort during development and data collection.
TE D
Corresponding Author: Shivayogi V. Hiremath, PhD
Department of Physical Therapy
EP
College of Public Health, Temple University
3307 North Broad Street, Jones Hall - Suite 623
AC C
Philadelphia, PA 19140
Phone: +001 215-707-7283, Fax: +001 215-707-7500 Email: [email protected]
Conflicts of Interest: The authors have no conflict of interest.
ACCEPTED MANUSCRIPT
1
Title: Estimation of Energy Expenditure for Wheelchair Users using a Physical Activity
2
Monitoring System
3
Abstract
RI PT
4 5
Objective: To develop and evaluate energy expenditure (EE) estimation models for a physical
7
activity monitoring system (PAMS) in manual wheelchair users (MWUs) with spinal cord injury
8
(SCI).
9 10
Design: Cross-sectional study.
11
M AN U
SC
6
Setting: University-based laboratory environment, a semi-structured environment at the National
13
Veterans Wheelchair Games, and the participants’ home environments.
TE D
12
14 15
Participants: Volunteer sample of MWUs with SCI (N=45).
EP
16
Intervention: Participants were asked to perform 10 physical activities of various intensities
18
from a list. The PAMS consists of a gyroscope based wheel rotation monitor (G-WRM) and an
19
accelerometer device worn on the upper arm or on the wrist. Criterion EE using a portable
20
metabolic cart and raw sensor data from PAMS were collected during each of these activities.
21
AC C
17
22
Main Outcome Measures: Estimated EE using custom models for MWUs based on either the
23
G-WRM and arm accelerometer (PAMS-Arm) or the G-WRM and wrist accelerometer (PAMS-
24
Wrist). 1
ACCEPTED MANUSCRIPT
25
Results: EE estimation performance for the PAMS-Arm (average error ± SD: -9.82% ± 37.03%)
27
and PAMS-Wrist (-5.65% ± 32. 61%) on the validation dataset indicated that both PAMS-Arm
28
and PAMS-Wrist were able to estimate EE for a range of physical activities with less than 10%
29
error. Moderate to high Intraclass Correlation Coefficients (ICC) indicated that the EE estimated
30
by PAMS-Arm (ICC(3,1)=0.82, p <0.05) and PAMS-Wrist (ICC(3,1)=0.89, p<0.05) are
31
consistent with the criterion EE.
SC
RI PT
26
32
Conclusions: Availability of physical activity (PA) monitors can assist wheelchair users to track
34
PA levels leading towards a healthier lifestyle. The new models we developed can estimate PA
35
levels in MWUs with SCI in laboratory and community settings.
M AN U
33
36
Keywords: Arm ergometry test; Energy expenditure; Physical activity; Rehabilitation; Spinal
38
cord injuries; Wheelchairs; Activity monitor system; Smartphones; Estimation
39
TE D
37
List of Abbreviations
41
EE
Energy expenditure
42
PAMS
Physical activity monitoring system
43
MWUs
Manual wheelchair users
44
SCI
Spinal cord injury
45
G-WRM
Gyroscope based wheel rotation monitor
46
PAMS-Arm
Gyroscope based wheel rotation monitor and arm accelerometer
47
PAMS-Wrist Gyroscope based wheel rotation monitor and wrist accelerometer
AC C
EP
40
2
ACCEPTED MANUSCRIPT
48
ICC
Intraclass Correlation Coefficients
49
PA
Physical activity
50
XX Laboratories
51
NVWG
National Veterans Wheelchair Games
52
10-fold-CV
Ten-fold within-subject cross validation
53
MSE
Mean signed error
54
MAE
Mean absolute error
55
MET
Metabolic Equivalent of Task
AC C
EP
TE D
M AN U
SC
RI PT
XLaboratories (Blinded for review purposes)
3
ACCEPTED MANUSCRIPT
56
When engaging in regular physical activity (PA), the 3.3 million individuals in the USA who use
58
wheelchairs for mobility face numerous challenges including mobility limitations, changes in
59
physiologic conditions, lack of accessible equipment and environmental barriers (1-4). These
60
factors affect the PA and sedentary behavior of wheelchair users leading to higher obesity rates
61
and other secondary conditions (1, 5-7). Research in the general population has shown that
62
behavioral weight loss interventions can produce clinically significant weight loss among obese
63
or overweight adults (8-12). Many of these interventions rely on self-monitoring of diet, PA and
64
body weight, and reducing energy intake and increasing energy expenditure (EE) (8-12). In
65
addition, recent research has shown that a combination of behavior and self-monitoring
66
technologies lead to significantly more weight loss than the traditional behavior-based weight
67
loss programs (13, 14). To change the sedentary lifestyle of wheelchair users, we can take an
68
approach similar to the general population to develop technological interventions that support
69
self-monitoring of PA levels.
SC
M AN U
TE D
70
RI PT
57
Sensor-based PA monitors have been used to track wheelchair movement (15-19), arm or wrist
72
movements (20-23), and physiological changes (18, 21) for quantifying PAs among persons who
73
use wheelchairs. Garcia-Masso et al. and Nightingale et al. (22, 23) indicated that the EE
74
estimated by the activity counts from the GT3X worn on the wrist was highly correlated with
75
criterion EE (housework activities, arm-ergometry, and propulsion: r=0.86 (22), propulsion and
76
deskwork: r=0.93 (23)). Furthermore, Kiuchi et al. found that EE estimated, by acceleration and
77
angular velocity, from an upper arm sensor (left upper arm: R2=0.75, right upper arm R2=0.87)
78
was similar to a wrist sensor (left wrist: R2=0.86, right wrist: R2=0.68) during wheelchair
AC C
EP
71
4
ACCEPTED MANUSCRIPT
propulsion on a treadmill (20). Hiremath et al. used SenseWear, a multi-sensor based activity
80
monitor, to develop activity specific models for four PAs including resting, wheelchair
81
propulsion, arm-ergometry, and deskwork to estimate EE in wheelchair users (r=0.88) (21).
82
Even though these studies have validated the use of sensor-based activity monitors, none of these
83
devices can concurrently capture both the wheelchair and arm movement, which are essential
84
variables for real-world day-to-day tracking of PA in wheelchair users.
RI PT
79
SC
85
The primary aim of the study was to develop and validate activity-specific EE estimation models
87
for manual wheelchair users based on a physical activity monitoring system (PAMS). The
88
system consists of a gyroscope-based wheel rotation monitor (G-WRM) (19) for capturing
89
wheelchair movement and a wearable accelerometer device (24) that tracks upper arm or wrist
90
movements. We evaluated two systems based on either the G-WRM and arm accelerometer
91
(PAMS-Arm) or the G-WRM and wrist accelerometer (PAMS-Wrist). The overall EE estimation
92
performance analysis is a two-step process (21) of sequentially applying the best classification
93
algorithms, which detect the wheelchair-based PAs and then applying the corresponding activity-
94
specific EE estimation model. Hiremath et al. developed classification algorithms such as
95
support vector machines and decision trees (25), which utilized sensor data from PAMS-Arm
96
and PAMS-Wrist to detect and classify various wheelchair-based PAs in laboratory, semi-
97
structured organizational, and unstructured home environments. Hiremath et al. (25) addressed
98
detection of wheelchair-based PAs; whereas this study addresses development and validation of
99
activity-specific EE estimation models for PAMS making a new and notable contribution in
100
estimating PA levels in wheelchair users. EE is an actionable parameter that individuals may
101
understand and relate to with their meal consumption in kilocalories. PAMS can allow
AC C
EP
TE D
M AN U
86
5
ACCEPTED MANUSCRIPT
individuals to learn about their PA patterns which may lead them to performing wheelchair-
103
based PAs that are associated with higher or lower EE values to attain the daily quota of EE
104
towards a healthier lifestyle. The secondary aim of the study was to assess if the PAMS-Arm will
105
be a better PA level estimator than the PAMS-Wrist.
RI PT
102
106 107
Methods
109
The study was approved by Institutional Review Boards of the University of YY, ZZ and the AA
110
Healthcare System.
M AN U
SC
108
111
Participants
113
A total of 45 individuals with Spinal Cord Injury (SCI) took part in the study. Participants were
114
included if they were 18-65 years of age, used a manual wheelchair as their primary means of
115
mobility (>80% of their ambulation), and had a diagnosis of SCI. Participants were excluded if
116
they were unable to tolerate sitting for three hours, had active pelvic or thigh wounds, had a
117
history of cardiovascular disease or were pregnant (self-report).
118
EP
TE D
112
Procedures
120
The first part of the study was performed in a structured laboratory environment at the XX
121
Laboratories (XLaboratories), University of YY (N=25) or in the semi-structured convention
122
center environment at the National Veterans Wheelchair Games (NVWG) 2012 held in
123
Richmond, VA (N=20). A subsection of the population who took part in the XLABORATORIES
124
testing sessions (N=20) participated in the study for a second time in their home environments.
AC C
119
6
ACCEPTED MANUSCRIPT
125
Protocol at XLABORATORIES or NVWG
127
Participants provided informed consent and then answered questions on demographics,
128
wheelchair information, and health and activity history (25). Body weight was measured using a
129
MX490D wheelchair scalea. Body height was either self-reported or measured using a tapeb (21).
RI PT
126
130
Participants were asked to choose from a list of PAs and perform at least 10 PAs other than
132
resting (25). Many of the activities were performed at submaximal capacity limiting the order
133
effect on the EE measurement. The type of PAs included: propelling their wheelchair at self-
134
selected speeds on various surfaces, or up and down a ramp; being pushed in their wheelchair;
135
playing wheelchair basketball or darts; folding laundry; performing deskwork; using a resistance
136
band; and performing arm-ergometry at self-selected speeds and resistances. This list of
137
activities included a range of common everyday activities that involve different parts of the body
138
and varying levels of intensity. Participants were instructed to refrain from eating and exercising
139
at least 2 hours and 12 hours, respectively, prior to the experiment. The resting trial involved
140
collecting baseline EE for six minutes while the participants sat still in their wheelchairs.
M AN U
TE D
EP
141
SC
131
During testing, the participants wore a K4b2 portable metabolic cartc. Participants also wore a
143
PAMS-Arm and a PAMS-Wrist. All participants performed PAs in their own wheelchairs for a
144
minimum of 6 min, with at least a 3 min break between PAs. Participants rated each activity trial
145
on Borg’s modified rate of perceived exertion scale (possible scores: 6-20).
AC C
142
146 147
7
ACCEPTED MANUSCRIPT
Protocol in Home Environment
149
The follow-up session involved an activity session of 10 daily activities and a resting trial that
150
the participants were able to perform in their home environments. The participants were provided
151
with an opportunity to perform PAs that they perform on a regular basis to increase the
152
applicability of the algorithms to real-world scenarios. The PAs performed in addition to the PAs
153
mentioned in the laboratory environment were: propelling in their home or community on
154
various surfaces; watching television; preparing food or simulating cooking; simulating eating;
155
sweeping the floor or vacuuming; making a bed; washing dishes or laundry; wheelchair pushups;
156
using dumbbells or a handgrip; and other household activities. The other household activities
157
included deskwork, folding clothes, using a resistance band, and playing video games.
158
M AN U
SC
RI PT
148
Instrumentation and Data Collection
160
The criterion EE was collected from a K4b2 portable metabolic cart comprised of an analyzer
161
unit, a battery pack and a face mask covering the participant’s mouth and nose. The analyzer unit
162
measured the quantity of oxygen (O2) and carbon dioxide (CO2) to estimate EE per breath. The
163
K4b2 was calibrated for every participant or every six hours before use to ensure its accuracy.
164
The participants also wore a Polar-T31 heart rate monitord on their chest.
EP
AC C
165
TE D
159
166
The G-WRM is a self-enclosed rechargeable Bluetooth-based wireless device that contains six
167
reed switches and a two-axis gyroscope to measure angular velocity of the wheelchair wheel and
168
distance traveled (19). The G-WRM was secured to the spokes of the wheelchair wheel. The
169
accelerometer is a small Bluetooth®-based wireless accelerometer that captures tri-axial
170
acceleration of a body part using a capacitive micro-machined accelerometer (24). The
8
ACCEPTED MANUSCRIPT
accelerometers were worn on the right upper arm over the triceps muscle and wrist. One of the
172
investigators placed the accelerometers on all of the participants and made sure that the location
173
was consistent during testing sessions. Both the G-WRM and the accelerometers were calibrated
174
prior to the testing and were time synchronized by the Android cellphone that collects sensor
175
data. The Android phone was secured to the participant’s waist.
176
RI PT
171
Development of EE Estimation Models
178
First, the data were categorized into three groups including near-stationary PAs, PAs that might
179
involve wheelchair movement (1.8m/min ≥ distance travelled ≤ 12m/min), and PAs with
180
consistent wheelchair movement (25). The choice of the threshold was based on the distance
181
travelled by the participants for various PAs in this study. The near-stationary PAs and PAs with
182
consistent wheelchair movement were further subcategorized into: a) resting, arm-ergometry,
183
and other household activities, and b) wheelchair propulsion, caretaker pushing and basketball,
184
respectively. This process of grouping various PAs from the total dataset led us to develop EE
185
estimation models for seven activities. The EE estimation models used statistical measures such
186
as time and frequency domain features, which were calculated based on the data collected from
187
G-WRM and accelerometers (Supplementary Note 1). We also included the participant
188
parameters such as weight, height, gender, age, injury characteristics, wheelchair weight and
189
basal metabolic rates, in order to see if these variables were predictors of EE estimation for
190
various activities. New linear regression models were developed using ten-fold within-subject
191
cross validation (10-fold-CV) on 80% of the participants’ data (training dataset). The models
192
were then tested on the remaining 20% of the participants’ data (80-20CV) (25). Training and
193
testing datasets for the 80-20CV were prepared using a stratified approach with injury level
AC C
EP
TE D
M AN U
SC
177
9
ACCEPTED MANUSCRIPT
194
(paraplegia versus tetraplegia) and gender (male versus female) in order to allocate 80% of the
195
participants into the training dataset and 20% into the testing dataset (Supplementary Note 2).
196
Model Evaluation
198
The performance of the EE estimation models from PAMS-Arm and PAMS-Wrist were
199
compared with the criterion EE using mean signed error (MSE) and mean absolute error (MAE).
200
MSE is obtained by averaging the over and under estimated EE during a PA. Intraclass
201
correlation coefficients (ICC(3,1)) for single measure, using a two-way mixed model with
202
consistency, and Bland Altman plots (26) assessed the agreement between criterion EE and the
203
estimated EE. In addition, we evaluated the overall performance of PAMS-Arm and PAMS-
204
Wrist for the two-step process by sequentially applying the best classification algorithms (25)
205
and the regression models developed in this study. All statistical analysis was performed using
206
IBM SPSS Statistics softwaref, with a statistical significance at an alpha level of 0.05.
207 208
TE D
M AN U
SC
RI PT
197
Results
210
Table 1 shows the demographics of the participants. All participants completed the study with 37
211
of them finishing 10 activity trials, 7 finishing 9 trials, and 1 finishing 8 trials (25). Table 2
212
summarizes the metabolic costs, the Metabolic Equivalent of Task (MET) from the metabolic
213
cart, the MET-SCI (27), the heart rate, the rate of perceived exertion, and the number of
214
participants who completed each activity trial. Perceived exertion ranged from no exertion at all
215
(6.0) to somewhat hard (13.0).
AC C
EP
209
216
10
ACCEPTED MANUSCRIPT
Table 3 shows features of the seven activity-specific EE prediction models that were chosen by
218
regression analysis during 10-fold cross validation on training dataset. All the EE estimation
219
models overestimated EE using the 80% training dataset (Table 4) for PAMS-Arm and PAMS-
220
Wrist. The validation errors obtained from using the 20% testing dataset (Table 4) ranged from
221
overestimation to underestimation (positive percentages) for PAMS-Arm and PAMS-Wrist.
222
Bland Altman plots (Figures 1 and 2) for PAMS-Arm and PAMS-Wrist show that majority of
223
the EE estimated values for various wheelchair-related PAs lie within the band of mean ± 2SD.
224
The pattern also indicates that the higher EE values were underestimated compared to the
225
criterion EE. EE estimation performance (Table 5) for the PAMS-Arm and PAMS-Wrist on the
226
validation dataset for the two-step process of fist classifying the activity (25) and then applying
227
the corresponding activity-specific model indicated that both PAMS-Arm and PAMS-Wrist had
228
less than 10% error.
M AN U
SC
RI PT
217
230
TE D
229
Discussion
232
PA level measurement in wheelchair users could be used to provide self-tracking tools that could
233
lead to a healthier lifestyle. The novelty and contribution of this study to the field are: for the
234
first time we have concurrently captured both the wheelchair and arm movement towards
235
estimating EE in PAs performed in laboratory and community settings, and developed
236
classification based EE estimation models for wheelchair users that utilize sensor data collected
237
via smartphone. The EE estimation models chose demographic features such as weight of the
238
person, lean body mass, height and gender. This is not surprising because total EE during resting
239
and PAs for an individual depends on these parameters (28). However, the estimation models for
AC C
EP
231
11
ACCEPTED MANUSCRIPT
arm-ergometry activity for both PAMS-Arm and PAMS-Wrist did not use the demographic
241
characteristics for EE estimation, indicating that movement variables better explained the
242
variance in the EE for arm-ergometry. Activity-specific EE estimation errors using the training
243
dataset for PAMS-Arm and PAMS-Wrist were similar for most wheelchair-related PAs, with the
244
exception of arm-ergometry exercise. Higher EE error by PAMS-Arm compared to PAMS-Wrist
245
may result from smaller movement of the upper arm versus wrist during arm-ergometry. Also,
246
we found that the activity-specific model for both PAMS-Arm and PAMS-Wrist underestimated
247
EE for the PA that may involve wheelchair movement in the testing dataset. This error might be
248
due to variation in participants’ style of performing activities and the amount of wheelchair
249
propulsion necessary for these types of PAs.
M AN U
SC
RI PT
240
250
EE estimation error obtained by sequentially applying classification algorithms (25) and the
252
estimation models showed that the overall EE error based on MSE was lower than 10% for both
253
PAMS-Arm (-9.8%) and PAMS-Wrist (-5.7%). MSE is commonly used in sports sciences as an
254
overall performance indicator for PA monitors over a period of time (21). The low MSE
255
indicates that both PAMS-Arm and PAMS-Wrist can estimate EE with low bias (<10%) in
256
MWUs with SCI. Another important measure that needs to be considered during EE estimation
257
is the MAE, which provides information regarding the magnitude of error. The MAE values for
258
PAMS-Arm and PAMS-Wrist show that features, based on movement and demographic
259
variables, were able to estimate EE with low to moderately high error. Higher MSE (SD) and
260
MAE values in this study are due to EE errors estimated per minute for each PA compared to
261
using EE estimated per participant for each PA. Furthermore, the EE estimated for PAMS-Arm
262
and PAMS-Wrist had moderate to high ICC values for the majority of PAs, thus indicating that
AC C
EP
TE D
251
12
ACCEPTED MANUSCRIPT
the EE values estimated by PAMS-Arm and PAMS-Wrist are consistent with the EE measured
264
by the portable metabolic cart. Bland Altman plots for PAMS-Arm and PAMS-Wrist indicated
265
that the EE estimated by the new models was balanced with over and under estimation of EE for
266
up to 5kcal/min; the new models tended to underestimate EE for both PAMS-Arm and PAMS-
267
Wrist above 5kcal/min. This may indicate that movement sensing alone might not be adequate
268
for vigorous wheelchair based activities. Overall, the validation analyses indicated that both
269
PAMS-Arm and PAMS-Wrist can estimate EE with reasonable accuracy (<10% MSE) in MWUs
270
with SCI. In addition, similar EE estimation performance by PAMS-Arm and PAM-Wrist allow
271
the wheelchair users to either wear the accelerometer on the wrist or the upper arm.
M AN U
SC
RI PT
263
272
Even though direct comparison between our work and other studies (20-23) is not possible due to
274
differences in the protocols of PAs, we have compared the studies on certain aspects. The high
275
correlations for the estimated EE and the criterion EE for PAMS-Wrist (ICC=0.84) and PAMS-
276
Arm (ICC=0.69) have a similar pattern to Kiuchi et al. for upper arm (left: R2=0.75, right:
277
R2=0.87) and wrist (left: R2=0.86, right: R2=0.68) during wheelchair propulsion (20). Similarly,
278
high overall ICC values for PAMS-Wrist (0.89) are comparable to the results of Garcia-Masso et
279
al. and Nightingale et al. with a high correlation (r=0.86 (22), r= 0.93 (23)) for various
280
wheelchair related activities. Furthermore, the EE estimation errors for PAMS-Arm (MSE: -
281
32.0% for basketball to 12.1% for may be moving) during the seven PAs were much higher than
282
the EE estimation error for SenseWear (MSE: -4.3% for resting to 9.9% for Arm-ergometry)
283
during the four PAs from our previous study (21). The higher variation in our current study is
284
due to the following reasons: a much larger number of wheelchair-based PAs collected in both
285
laboratory and community settings; the merging of these numerous PAs into seven groups;
AC C
EP
TE D
273
13
ACCEPTED MANUSCRIPT
collection of data at two time points for 20 participants; and the participants performing PAs at a
287
self-selected pace or pattern.
288
Study Limitations
289
Some of the limitations of the study are a large percentage of our participants self-reported that
290
they were physically active on a regular or occasional basis, inclusion of individuals with SCI,
291
and EE models developed here were based on movement-based variables. The PA levels
292
reported by the participants of this study were much higher than reported in the community (29,
293
30). In future we plan to recruit a greater percentage of individuals with different disabilities
294
from the community so that the EE estimation models can be used for individuals with a variety
295
of disabilities who use wheelchairs. We also plan to evaluate real-time feedback provided by
296
PAMS to wheelchair users, which may assist them in pursuing behavior changes to achieve a
297
healthy and active lifestyle. Future studies should assess incorporating other forms of
298
physiologic sensing, such as galvanic skin response, skin temperature, near body temperature,
299
and heart rate, in order to quantify resistance-based PAs. Another limitation of the study is the
300
different number of participants for each activity trial due to participants performing PAs that
301
they perform on a regular basis in their home environments, and inability of the participants to
302
perform certain types of PAs, such as basketball, in their home environments. To address this
303
limitation we reduced the PAs from the total dataset into seven activities which led to decreased
304
variability of EE estimation error.
306
SC
M AN U
TE D
EP
AC C
305
RI PT
286
307
Conclusions
308
This study developed and validated new EE estimation models for MWUs based on upper arm,
309
wrist and wheelchair movements detected with the help of a physical activity monitoring system. 14
ACCEPTED MANUSCRIPT
310
The new models we developed can estimate PA levels in MWUs with SCI in laboratory and
311
community settings.
312
RI PT
313 314
Suppliers list
315
a. Befour MX490D wheelchair scale; Befour, Inc, 102 N Progress Dr, Saukville, WI 53080, USA
SC
316
b. Stanley Tape; Stanley Works, 480 Myrtle Street, New Britain, CT 06053, USA
318
c. K4b2 portable metabolic cart; COSMED srl, Via dei Piani di Mt. Savello 37, Pavona di
322 323 324 325 326
Success, NY 11042, USA
e. MATLAB® version 2013a; The Mathworks, Inc., 3 Apple Hill Drive Natick, MA, 01760, USA, USA
TE D
321
d. Polar-T31 heart rate monitor; Polar Electro Inc., 1111 Marcus Avenue, Suite M15 Lake
f. IBM SPSS Statistics software version 20.0; IBM Corporation, 1 New Orchard Road Armonk, NY, 10504, USA
EP
320
Albano, Rome 0004, Italy
AC C
319
M AN U
317
15
ACCEPTED MANUSCRIPT
References
328 329
1. Durstine JL, Painter P, Franklin BA, Morgan D, Pitetti KH, Roberts SO. Physical activity for the chronically ill and disabled. Sports Med. 2000;30(3):207-19.
330 331
2. Hoenig H, Landerman LR, Shipp KM, George L. Activity restriction among wheelchair users. Journal of American Geriatrics Society. 2003;51(9):1244-51.
332 333 334
3. Glaser RM, Janssen TWJ, Suryaprasad AG, Gupta SC, Mathews T. The Physiology of Exercise. In: Apple DF, editor. Physical Fitness: A Guide for Individuals with Spinal Cord Injury. Washington, DC: Department of Veterans Affairs; 1996.
335 336
4. Rimmer JH. Use of the ICF in identifying factors that impact participation in physical activity/rehabilitation among people with disabilities. Disabil Rehabil. 2006;28(17):1087 – 95.
337 338
5. Jacobs PL, Nash MS, Rusinowski JW. Circuit training provides cardiorespiratory and strength benefits in persons with paraplegia. Med Sci Sports Exerc. 2001;33:711-7.
339 340
6. Rimmer JH, Yamaki K, Davis BM, Wang E, Vogel LC. Obesity and Overweight Prevalence Among Adolescents With Disabilities. Preventing Chronic Disease. 2011;8(2):1-6.
341 342
7. Rimmer JH, Shenoy SS. Impact of Exercise on Targeted Secondary Conditions. Field MJ, Jette AM, Martin L, editors. Washington, DC: National Academy of Sciences; 2006.
343 344 345
8. Jakicic JM, Tate DF, Lang W, Davis KK, Polzien K, Rickman AD, et al. Effect of a stepped-care intervention approach on weight loss in adults: a randomized clinical trial. Journal of the American Medical Association. 2012;307(24):2617-26.
346 347 348
9. Coons MJ, DeMott A, Buscemi J, Duncan JM, Pellegrini CA, Steglitz J, et al. Technology Interventions to Curb Obesity: A Systematic Review of the Current Literature. Current cardiovascular risk reports. 2012;6:120-34.
349 350 351
10. Alhassan S, Kim S, Bersamin A, King A, Gardner C. Dietary adherence and weight loss success among overweight women: results from the A to Z weight loss study. Int J Obes. 2008;32:985-91.
352 353
11. Baker RC, Kirschenbaum DS. Self-monitoring may be necessary for successful weight control. Behavior Therapy. 1993;24:377–94.
AC C
EP
TE D
M AN U
SC
RI PT
327
16
ACCEPTED MANUSCRIPT
12. 55.
Wing R, Phelan S. Long-term weight loss maintenance. Am J Clin Nutr. 2005;82:2225–
356 357 358
13. Shuger SL, Barry VW, Sui X, McClain A, Hand GA, Wilcox S, et al. Electronic feedback in a diet- and physical activity-based lifestyle intervention for weight loss: a randomized controlled trial. International Journal of Behavioral Nutrition and Physical Activity. 2011;8(41).
359 360 361
14. Spring B, Duncan JM, Janke EA, Kozak AT, McFadden HG, Demott A, et al. Integrating technology into standard weight loss treatment a randomized controlled trial. JAMA Internal Medicine. 2013;173(2):105-11.
362 363 364
15. Tolerico ML, Ding D, Cooper RA, Spaeth DM, Fitzgerald SG, Cooper R, et al. Assessing mobility characteristics and activity levels of manual wheelchair users. J Rehabil R D. 2007;44(4):561-72.
365 366
16. Coulter EH, Dall PM, Rochester L, Hasler JP, Granat MH. Development and validation of a physical activity monitor for use on a wheelchair. Journal of Spinal Cord. 2011;49:445-50.
367 368
17. Sonenblum SE, Sprigle S, Caspall J, Lopez R. Validation of an accelerometer-based method to measure the use of manual wheelchairs. Med Eng Phys. 2012;34:781-86.
369 370
18. Conger SA, Scott SN, Bassett DR. Predicting energy expenditure through hand rim propulsion power output in individuals who use wheelchairs. Br J Sports Med. 2014;48:1048-53.
371 372
19. Hiremath SV, Ding D, Cooper RA. Development and evaluation of a gyroscope based wheel rotation monitor for manual wheelchair users. Spinal Cord Medicine. 2013;36(4):347-56.
373 374 375 376
20. Kiuchi K, Inayama T, Muraoka Y, Ikemoto S, Uemura O, Mizuno K. Preliminary study for the assessment of physical activity using a triaxial accelerometer with a gyro sensor on the upper limbs of subjects with paraplegia driving a wheelchair on a treadmill. Spinal Cord. 2014;52:556-63.
377 378 379
21. Hiremath SV, Ding D, Farringdon J, Cooper RA. Predicting energy expenditure of manual wheelchair users with spinal cord injury using a multi-sensor based activity monitor. Arch Phys Med Rehabil. 2012;93(11):1937-43.
AC C
EP
TE D
M AN U
SC
RI PT
354 355
17
ACCEPTED MANUSCRIPT
22. García-Massó X, Serra-Añó P, García-Raffi L, Sánchez-Pérez E, López-Pascual J, Gonzalez L. Validation of the use of Actigraph GT3X accelerometers to estimate energy expenditure in full time manual wheelchair users with spinal cord injury. Spinal Cord. 2013;51(12):898-903.
384 385
23. Nightingale TE, Walhin J-P, Thompson D, Bilzon JL. Predicting Physical Activity Energy Expenditure in Manual Wheelchair Users. Med Sci Sports Exerc. 2014;46(9):1849-58.
386 387 388 389 390
24. Intille SS, Albinali F, Mota S, Kuris B, Botana P, Haskell WL, editors. Design of a Wearable Physical Activity Monitoring System using Mobile Phones and Accelerometers. Proceedings of Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2011 Aug 30- Sep 3; Boston, MA: IEEE Engineering in Medicine and Biology Society.
391 392
25. Hiremath SV, Intille SS, Kelleher A, Cooper RA, Ding D. Detection of physical activities using a physical activity monitor system for wheelchair users. Med Eng Phys. 2015;37(1):68-76.
393 394
26. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet. 1986;1:307-10.
395 396
27. Collins EG, Gater D, Kiratli J, Butler J, Hanson K, Langbein WE. Energy cost of physical activities in persons with spinal cord injury. Med Sci Sports Exerc. 2010;42(4):691-700.
397 398 399
28. American College of Sports Medicine, Mitchell H. Whaley, Peter H. Brubaker, Robert Michael Otto, Lawrence E. Armstrong ACSM's guidelines for exercise testing and prescription. 7, illustrated ed: Lippincott Williams & Wilkins; 2005.
400 401
29. Washburn R, Hedrick BN. Descriptive epidemiology of physical activity in university graduates with locomotor disabilities. Int J Rehabil Res. 1997;20(3):275-87.
402 403 404
30. Tasiemski T, Kennedy P, Gardner BP, Taylor N. The association of sports and physical recreation with life satisfaction in a community sample of people with spinal cord injuries. NeuroRehabilitation. 2005;20:253-65.
AC C
EP
TE D
M AN U
SC
RI PT
380 381 382 383
405 406 407 408 409
18
ACCEPTED MANUSCRIPT
410
Figure Legends
411
Figure 1: Bland Altman plot of EE estimated using activity-specific models for PAMS-Arm and
413
EE measured for the various wheelchair-related PAs in the validation dataset. The x and y axes
414
represents the mean and difference of the EE estimated and the EE measured, respectively.
RI PT
412
415
Figure 2: Bland Altman plot of EE estimated using activity-specific models for PAMS-Wrist and
417
EE measured for the various wheelchair-related PAs in the validation dataset. The x and y axes
418
represents the mean and difference of the EE estimated and the EE measured, respectively.
M AN U
SC
416
419 420
AC C
EP
TE D
421
19
ACCEPTED MANUSCRIPT
Table 1: Demographic characteristics of the participants.
39 6 41.0 (12.6)* 78.1 (18.1) 1.8 (0.1) C5 to L5 13 32
36 5 4 31 14
AC C
EP
TE D
Note: * Values are n or mean (SD).
SC
22 23 12.6 (8.6)
RI PT
Values 45
M AN U
Variables Number of participants Sex Male Female Age (years) Weight (kg) Height (m) Injury level (range) Injury between C5 and C8 Injury between T3 and L5 Injury completeness Complete Incomplete Manual wheelchair usage (years) Self-reported PA Regular Occasional No regular PA Self-reported smokers Non-smokers Smokers
ACCEPTED MANUSCRIPT
Table 2: Metabolic costs in terms of EE, METs, heart rate, rate of perceived exertion, and number of participants and minutes per
RI PT
activity for various wheelchair-based physical activities. The MET-SCI was calculated based on previous research in individuals with SCI (27) by using a reference metabolic rate (VO2/Kg) of 2.7ml/min/kg for SCI population. No. of Participants
No. of Min
Heart Rate in beats/min Mean
SD
EE from K4b2 in kcal/min Mean SD
MET
Mean
MET-SCI
SC
Activity Trial
Rate of Perceived Exertion Mean SD
SD
Mean
SD
0.2
1.1
0.3
6.0
0.1
1.0
3.0
1.3
10.9
2.2
43
363
69.9
16.2
1.1
0.3
0.9
Arm-Ergometry
43
500
96.0
19.0
3.0
1.2
2.3
Darts
33
214
91.1
15.3
2.7
0.8
2.1
0.6
2.7
0.8
8.7
2.0
Deskwork
43
574
79.1
14.2
1.5
0.6
1.2
0.5
1.6
0.6
7.5
2.1
Folding Clothes
42
343
92.5
17.6
2.3
0.6
1.8
0.4
2.3
0.6
8.6
2.3
Propulsion
43
901
101.2
20.5
3.5
1.5
2.7
1.1
3.5
1.5
11.0
3.1
Caretaker Pushing
42
341
75.3
14.2
1.3
0.4
1.0
0.3
1.3
0.4
6.4
1.3
Resistance
43
367
86.2
15.2
2.0
0.7
1.6
0.5
2.0
0.7
10.0
2.3
Basketball
19
112
110.2
19.7
4.5
1.7
3.7
1.2
4.8
1.6
12.6
2.7
Eating
17
17
73.0
13.2
2.0
0.4
1.6
0.6
2.1
0.8
7.5
2.1
Sweeping Floor
14
90
96.2
15.8
3.0
0.9
2.4
0.8
3.1
1.1
10.9
3.1
Preparing Food
11
68
87.6
17.9
2.3
0.6
1.8
0.5
2.3
0.7
7.7
1.6
Making Bed
1
6
90.6
7.3
2.7
0.6
2.3
0.5
3.0
0.6
7
0.0
Cleaning Room
4
26
97.8
27.7
2.3
0.6
2.5
0.6
3.2
0.7
8.7
2.9
Filing papers
2
12
91.8
16.1
1.1
0.2
1.5
0.2
1.9
0.3
7.5
2.1
Check mail
2
8
89.9
12.9
2.3
0.6
2.1
0.4
2.7
0.6
6.0
0.0
Laundry
2
16
89.2
7.0
2.7
0.4
2.7
0.3
3.5
0.4
9.0
2.6
Video Game
1
6
72.3
2.8
1.9
0.2
1.2
0.1
1.6
0.2
13.0
0.0
Cleaning Car Wheelchair Pushups
1
6
76.6
7.0
2.7
0.6
1.9
0.4
2.5
0.5
7.0
0.0
1
6
97.0
15.4
2.6
0.7
2.9
0.7
3.7
1.0
13.0
0.0
AC C
EP
TE D
M AN U
Resting
ACCEPTED MANUSCRIPT
Table 3: Features of the activity-specific EE estimation models that were chosen by regression analysis during 10-fold cross validation on training dataset (80% of participants’ data).
Caretaker Pushing Basketball May be moving Resting Armergometry
PAMSWrist OA not moving
Propulsion Caretaker Pushing Basketball May be moving
RI PT
SC
Propulsion
M AN U
PAMSArm
OA not moving
TE D
Armergometry
EP
Resting
Features Lean body mass, Mean of velocity from G-WRM, Entropy without the DC component of z axis acceleration (acc.) from arm accelerometer (Arm), Dominant frequency content (range: 0.3-15 Hz) of x axis acc. from Arm, Root mean square of z axis acc. from Arm Back trend of resultant acc. from Arm, Ratio of dominant frequency's power with total power for resultant acc. from Arm, Entropy without the DC component of y axis acc. from Arm, Number of peaks for z axis acc. from Arm., Energy content of velocity from GWRM Mean absolute deviation of z axis acc. from Arm multiplied by square root of height, Gender, Standard deviation of z axis acc. from Arm, Standard deviation of six minutes or less for z axis acc. from Arm, Dominant frequency content’s power of x axis acc. from Arm Ratio of dominant frequency's power with total power for resultant acc. from Arm, Second dominant frequency content's power of resultant acc. from Arm, Mass to the power of 0.75, Ratio of dominant frequency for the current and the past minute of velocity from GWRM, Dominant frequency content’s power of velocity from G-WRM Lean body mass, Energy without the DC component of z axis acc. from Arm, Entropy without the DC component of x axis acc. from Arm, Mean cross rate of z axis acc. from Arm, Entropy without the DC component of y axis acc. from Arm Back trend of velocity from G-WRM, Lean body mass, Mean absolute deviation with respect to median of resultant acc. from Arm, Second dominant frequency content's power of x axis acc. from Arm, Entropy without the DC component of resultant acc. from Arm Mean absolute deviation with respect to mean of x axis acc. from Arm, Lean body mass, Standard deviation of six minutes or less for x axis acc. from Arm, Second dominant frequency content's power of y axis acc. from Arm, Mean absolute deviation with respect to median of x axis acc. from Arm Lean body mass, Difference between mean values of acc. from x and y axes from wrist accelerometer (Wrist), Ratio of dominant frequency for the current and the past minute of resultant acc. from Wrist, Mean cross rate of resultant acc. from Wrist, Zero cross rate of x axis acc. from Wrist Sum of mean absolute deviation in X and Y axes acc. from Wrist, mean of velocity from G-WRM, Number of peaks for x axis acc. from Wrist, Ratio of dominant frequency for the current and the past minute of x axis acc. from Wrist, Energy without the DC component of x axis acc. from Wrist Number of peaks for y axis acc. from Wrist, Gender, Total power for frequencies (range: 0.3-15 Hz) of y axis acc. from Wrist, Standard deviation of six minutes or less for resultant acc. from Wrist, Entropy content of y axis acc. from Wrist Third dominant frequency content's power for velocity from G-WRM, Back trend of velocity from G-WRM, Mass to the power of 0.75, Correlation between y axis acc. and resultant acc. from Wrist, Total power for frequencies for resultant acc. from Wrist Lean body mass, Second dominant frequency content's power of z axis acc. from Wrist, Standard deviation of six minutes or less for velocity from G-WRM, Correlation between x axis acc. and resultant acc. from Wrist, Dominant frequency content of z axis acc. from Wrist Back trend of velocity from G-WRM, Number of peaks for resultant acc. from Wrist, Number of peaks for velocity from G-WRM, Mufflin basal metabolic rate, Second dominant frequency content's power of x axis acc. from wrist Lean body mass, amplitude of y axis acc. from Wrist, Standard deviation of six minutes or less for z axis acc. from Wrist, World Health Organization resting metabolic rate divided by lean body mass, World Health Organization resting metabolic rate
AC C
Device Activity
ACCEPTED MANUSCRIPT
Table 4: Mean signed error (%) of EE estimation for each activity-specific model using the features obtained by 10-fold cross validation on 80% of participants’ data (training dataset) and tested on the remaining 20% of participants’ data not used for training (testing data).
-14.44
49.94
-1.66
42.32
-7.69
33.29
-10.14
38.38
-11.86
33.70
-9.72
36.79
-7.50
27.27
-11.65
12.84
-5.15
-6.66
29.28
-0.96
39.33
-7.45
-1.85
35.97
-16.21
16.79
-2.14
-6.64
16.68
29.82
37.71
-7.20
-7.66
33.75
-1.00
13.10
-6.54
RI PT
Resting Armergometry OA Not moving Propulsion Caretaker pushing Basketball May be moving Overall
PAMS-Wrist Train Test Mean SD Mean SD -6.82 34.36 -6.94 20.33 5.34
33.00
-12.61
32.78
SC
Activity
PAMS-Arm Train Test Mean SD Mean SD -7.05 35.36 -5.32 19.75
24.21
-7.81
31.35
36.23
-4.77
17.20
19.80
-11.45
30.97
28.53
22.38
13.61
30.28
-0.77
11.28
M AN U
Device
Note: Other PAs (OA) while being stationary; Overall error or combined error (Overall); Negative percentages indicate overestimation where the models estimated an EE value higher
AC C
EP
TE D
than the EE measured by K4b2; positive percentages indicate underestimation.
ACCEPTED MANUSCRIPT
Table 5: EE estimation performance for activity-specific (AS) models on the validation dataset (20% of participants’ data not used for training) for the two-step process of fist classifying the
Wrist. MAE Mean
MSE Mean
SD
ICC
ICC(3,1) 95% CI LB UB
P
1.09
0.27
1.21
0.21
19.41
-14.02 19.76
0.77
0.48
0.90
<0.05
3.25
1.27
2.98
0.66
31.83
-2.13
42.47
0.47
0.19
0.66
<0.05
1.86
0.85
1.93
0.78
30.47
-15.04 35.13
0.73
0.64
0.80
<0.05
3.65
1.99
3.47
0.93
31.69
-11.56 39.64
0.69
0.57
0.78
<0.05
1.3
0.39
1.25
18.46
0.82
0.68
0.90
<0.05
3.57
1.31
4.61
3.66
1.02
2.90
2.65 1.22
1.63 0.57
2.58 1.17
3.15
1.27
2.87
1.91
SC
Resting Armergometry OA not moving Propulsion Caretaker Pushing Basketball May be moving Overall Resting Armergometry OA not moving Propulsion Caretaker Pushing Basketball May be moving Overall
0.28
14.10
0.25
1.62
35.18
-31.96 38.30
0.84
0.56
0.94
<0.05
0.34
27.69
12.14
37.10
0.40
-0.47
0.76
0.13
1.19 0.26
29.04 21.20
-9.82 -4.00
37.03 24.82
0.82 0.62
0.79 0.30
0.85 0.80
<0.05 <0.05
1.05
23.29
3.85
32.40
0.85
0.77
0.90
<0.05
0.86
1.91
0.39
28.73
-12.32 33.50
0.63
0.48
0.73
<0.05
2.01
3.50
1.21
26.09
-8.39
33.37
0.84
0.78
0.89
<0.05
0.64
1.31
0.27
16.28
-1.51
20.81
0.68
0.43
0.82
<0.05
1.30
4.30
2.02
28.64
-14.77 29.98
0.91
0.75
0.97
<0.05
3.66
1.02
2.97
0.32
29.69
9.26
45.86
0.61
0.05
0.84
0.02
2.65
1.63
2.53
1.26
25.19
-5.65
32.61
0.89
0.87
0.91
<0.05
3.64 1.4 3.61
AC C
EP
PAMSWrist
EE from K4b2 EE AS in kcal/min Mean SD Mean SD
M AN U
PAMSArm
Activity
TE D
Device
RI PT
activity and then applying the corresponding activity-specific model for PAMS-Arm and PAMS-
Note: EE Activity Specific (AS); Mean Absolute Error in percentage (MAE); Mean Signed Error (MSE); Intraclass correlation coefficients (ICC); ICC with 95% confidence interval (CI) – lower bound (LB) and upper bound (UB); ICC significance values (p); Overall error or combined error (Overall).
AC C
EP
TE D
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT
Supplementary Note 1: The time domain features, such as mean, mean absolute deviation, and peaks were simple to extract and can be used to estimate EE of PAs that are considerably different. The frequency domain features, such as total power between a band of frequencies,
RI PT
energy, and entropy can be used to estimate EE based on the key frequency of movement
(wheelchair propulsion and arm-ergometry). The features were extracted using custom programs written in MATLAB® (version 2013a) for a 1-min window size to be consistent with the EE
SC
estimation.
M AN U
Supplementary Note 2: The 10-fold-CV is a statistical technique used as part of the development process to select variables and optimize the regression model. The process involves generating the regression models iteratively, by selecting and adding features, to arrive at a final set of variables that had the least EE estimation error for the 80% of the data used in training. Then
TE D
once we developed and saved the final model on the training dataset we tested the final model on
AC C
EP
the testing dataset, which was not used in building the model.
S0003-9993(16)00155-6
DOI:
10.1016/j.apmr.2016.02.016
Reference:
YAPMR 56471
To appear in:
ARCHIVES OF PHYSICAL MEDICINE AND REHABILITATION
Received Date: 8 December 2015 Revised Date:
25 January 2016
Accepted Date: 22 February 2016
Please cite this article as: Hiremath SV, Intille SS, Kelleher A, Cooper RA, Ding D, Estimation of Energy Expenditure for Wheelchair Users using a Physical Activity Monitoring System, ARCHIVES OF PHYSICAL MEDICINE AND REHABILITATION (2016), doi: 10.1016/j.apmr.2016.02.016. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
Title Page Running Head: Energy Expenditure in Wheelchair users
RI PT
Title: Estimation of Energy Expenditure for Wheelchair Users using a Physical Activity Monitoring System
Rory A. Cooper, PhD1-2,7, Dan Ding PhD,1-2,7
Affiliations:
M AN U
1-2,
SC
Authors: Shivayogi V. Hiremath, PhD,1-4*, Stephen S. Intille, PhD,5-6, Annmarie Kelleher, MS,
1) Department of Veterans Affairs, Human Engineering Research Laboratories, VA Pittsburgh Healthcare System, Pittsburgh, PA
PA
TE D
2) Department of Rehabilitation Science and Technology, University of Pittsburgh, Pittsburgh,
3) Department of Physical Medicine and Rehabilitation, University of Pittsburgh, Pittsburgh,
EP
PA
4) Department of Physical Therapy, Temple University, Philadelphia, PA
AC C
5) College of Computer and Information Science, Northeastern University, Boston, MA 6) Department of Health Sciences, Northeastern University, Boston, MA 7) Department of Bioengineering, University of Pittsburgh, Pittsburgh, PA
* This research study was executed at the Human Engineering Research Laboratories, VA Pittsburgh Healthcare System and Department of Rehabilitation Science and Technology,
ACCEPTED MANUSCRIPT
University of Pittsburgh.
Acknowledgements:
RI PT
The work is supported by the Department of Defense (W81XWH-10-1-0816). SVH’s work in this article was funded through the Switzer Research Fellowship (H133F110032) awarded by the National Institute on Disability and Rehabilitation Research, Department of Education.
SC
The work is also supported by the Human Engineering Research Laboratories, VA Pittsburgh Healthcare System. The contents do not represent the views of the Department of Veterans
M AN U
Affairs or the United States Government. The authors thank their colleagues at the Human Engineering Research Laboratories for their input and effort during development and data collection.
TE D
Corresponding Author: Shivayogi V. Hiremath, PhD
Department of Physical Therapy
EP
College of Public Health, Temple University
3307 North Broad Street, Jones Hall - Suite 623
AC C
Philadelphia, PA 19140
Phone: +001 215-707-7283, Fax: +001 215-707-7500 Email: [email protected]
Conflicts of Interest: The authors have no conflict of interest.
ACCEPTED MANUSCRIPT
1
Title: Estimation of Energy Expenditure for Wheelchair Users using a Physical Activity
2
Monitoring System
3
Abstract
RI PT
4 5
Objective: To develop and evaluate energy expenditure (EE) estimation models for a physical
7
activity monitoring system (PAMS) in manual wheelchair users (MWUs) with spinal cord injury
8
(SCI).
9 10
Design: Cross-sectional study.
11
M AN U
SC
6
Setting: University-based laboratory environment, a semi-structured environment at the National
13
Veterans Wheelchair Games, and the participants’ home environments.
TE D
12
14 15
Participants: Volunteer sample of MWUs with SCI (N=45).
EP
16
Intervention: Participants were asked to perform 10 physical activities of various intensities
18
from a list. The PAMS consists of a gyroscope based wheel rotation monitor (G-WRM) and an
19
accelerometer device worn on the upper arm or on the wrist. Criterion EE using a portable
20
metabolic cart and raw sensor data from PAMS were collected during each of these activities.
21
AC C
17
22
Main Outcome Measures: Estimated EE using custom models for MWUs based on either the
23
G-WRM and arm accelerometer (PAMS-Arm) or the G-WRM and wrist accelerometer (PAMS-
24
Wrist). 1
ACCEPTED MANUSCRIPT
25
Results: EE estimation performance for the PAMS-Arm (average error ± SD: -9.82% ± 37.03%)
27
and PAMS-Wrist (-5.65% ± 32. 61%) on the validation dataset indicated that both PAMS-Arm
28
and PAMS-Wrist were able to estimate EE for a range of physical activities with less than 10%
29
error. Moderate to high Intraclass Correlation Coefficients (ICC) indicated that the EE estimated
30
by PAMS-Arm (ICC(3,1)=0.82, p <0.05) and PAMS-Wrist (ICC(3,1)=0.89, p<0.05) are
31
consistent with the criterion EE.
SC
RI PT
26
32
Conclusions: Availability of physical activity (PA) monitors can assist wheelchair users to track
34
PA levels leading towards a healthier lifestyle. The new models we developed can estimate PA
35
levels in MWUs with SCI in laboratory and community settings.
M AN U
33
36
Keywords: Arm ergometry test; Energy expenditure; Physical activity; Rehabilitation; Spinal
38
cord injuries; Wheelchairs; Activity monitor system; Smartphones; Estimation
39
TE D
37
List of Abbreviations
41
EE
Energy expenditure
42
PAMS
Physical activity monitoring system
43
MWUs
Manual wheelchair users
44
SCI
Spinal cord injury
45
G-WRM
Gyroscope based wheel rotation monitor
46
PAMS-Arm
Gyroscope based wheel rotation monitor and arm accelerometer
47
PAMS-Wrist Gyroscope based wheel rotation monitor and wrist accelerometer
AC C
EP
40
2
ACCEPTED MANUSCRIPT
48
ICC
Intraclass Correlation Coefficients
49
PA
Physical activity
50
XX Laboratories
51
NVWG
National Veterans Wheelchair Games
52
10-fold-CV
Ten-fold within-subject cross validation
53
MSE
Mean signed error
54
MAE
Mean absolute error
55
MET
Metabolic Equivalent of Task
AC C
EP
TE D
M AN U
SC
RI PT
XLaboratories (Blinded for review purposes)
3
ACCEPTED MANUSCRIPT
56
When engaging in regular physical activity (PA), the 3.3 million individuals in the USA who use
58
wheelchairs for mobility face numerous challenges including mobility limitations, changes in
59
physiologic conditions, lack of accessible equipment and environmental barriers (1-4). These
60
factors affect the PA and sedentary behavior of wheelchair users leading to higher obesity rates
61
and other secondary conditions (1, 5-7). Research in the general population has shown that
62
behavioral weight loss interventions can produce clinically significant weight loss among obese
63
or overweight adults (8-12). Many of these interventions rely on self-monitoring of diet, PA and
64
body weight, and reducing energy intake and increasing energy expenditure (EE) (8-12). In
65
addition, recent research has shown that a combination of behavior and self-monitoring
66
technologies lead to significantly more weight loss than the traditional behavior-based weight
67
loss programs (13, 14). To change the sedentary lifestyle of wheelchair users, we can take an
68
approach similar to the general population to develop technological interventions that support
69
self-monitoring of PA levels.
SC
M AN U
TE D
70
RI PT
57
Sensor-based PA monitors have been used to track wheelchair movement (15-19), arm or wrist
72
movements (20-23), and physiological changes (18, 21) for quantifying PAs among persons who
73
use wheelchairs. Garcia-Masso et al. and Nightingale et al. (22, 23) indicated that the EE
74
estimated by the activity counts from the GT3X worn on the wrist was highly correlated with
75
criterion EE (housework activities, arm-ergometry, and propulsion: r=0.86 (22), propulsion and
76
deskwork: r=0.93 (23)). Furthermore, Kiuchi et al. found that EE estimated, by acceleration and
77
angular velocity, from an upper arm sensor (left upper arm: R2=0.75, right upper arm R2=0.87)
78
was similar to a wrist sensor (left wrist: R2=0.86, right wrist: R2=0.68) during wheelchair
AC C
EP
71
4
ACCEPTED MANUSCRIPT
propulsion on a treadmill (20). Hiremath et al. used SenseWear, a multi-sensor based activity
80
monitor, to develop activity specific models for four PAs including resting, wheelchair
81
propulsion, arm-ergometry, and deskwork to estimate EE in wheelchair users (r=0.88) (21).
82
Even though these studies have validated the use of sensor-based activity monitors, none of these
83
devices can concurrently capture both the wheelchair and arm movement, which are essential
84
variables for real-world day-to-day tracking of PA in wheelchair users.
RI PT
79
SC
85
The primary aim of the study was to develop and validate activity-specific EE estimation models
87
for manual wheelchair users based on a physical activity monitoring system (PAMS). The
88
system consists of a gyroscope-based wheel rotation monitor (G-WRM) (19) for capturing
89
wheelchair movement and a wearable accelerometer device (24) that tracks upper arm or wrist
90
movements. We evaluated two systems based on either the G-WRM and arm accelerometer
91
(PAMS-Arm) or the G-WRM and wrist accelerometer (PAMS-Wrist). The overall EE estimation
92
performance analysis is a two-step process (21) of sequentially applying the best classification
93
algorithms, which detect the wheelchair-based PAs and then applying the corresponding activity-
94
specific EE estimation model. Hiremath et al. developed classification algorithms such as
95
support vector machines and decision trees (25), which utilized sensor data from PAMS-Arm
96
and PAMS-Wrist to detect and classify various wheelchair-based PAs in laboratory, semi-
97
structured organizational, and unstructured home environments. Hiremath et al. (25) addressed
98
detection of wheelchair-based PAs; whereas this study addresses development and validation of
99
activity-specific EE estimation models for PAMS making a new and notable contribution in
100
estimating PA levels in wheelchair users. EE is an actionable parameter that individuals may
101
understand and relate to with their meal consumption in kilocalories. PAMS can allow
AC C
EP
TE D
M AN U
86
5
ACCEPTED MANUSCRIPT
individuals to learn about their PA patterns which may lead them to performing wheelchair-
103
based PAs that are associated with higher or lower EE values to attain the daily quota of EE
104
towards a healthier lifestyle. The secondary aim of the study was to assess if the PAMS-Arm will
105
be a better PA level estimator than the PAMS-Wrist.
RI PT
102
106 107
Methods
109
The study was approved by Institutional Review Boards of the University of YY, ZZ and the AA
110
Healthcare System.
M AN U
SC
108
111
Participants
113
A total of 45 individuals with Spinal Cord Injury (SCI) took part in the study. Participants were
114
included if they were 18-65 years of age, used a manual wheelchair as their primary means of
115
mobility (>80% of their ambulation), and had a diagnosis of SCI. Participants were excluded if
116
they were unable to tolerate sitting for three hours, had active pelvic or thigh wounds, had a
117
history of cardiovascular disease or were pregnant (self-report).
118
EP
TE D
112
Procedures
120
The first part of the study was performed in a structured laboratory environment at the XX
121
Laboratories (XLaboratories), University of YY (N=25) or in the semi-structured convention
122
center environment at the National Veterans Wheelchair Games (NVWG) 2012 held in
123
Richmond, VA (N=20). A subsection of the population who took part in the XLABORATORIES
124
testing sessions (N=20) participated in the study for a second time in their home environments.
AC C
119
6
ACCEPTED MANUSCRIPT
125
Protocol at XLABORATORIES or NVWG
127
Participants provided informed consent and then answered questions on demographics,
128
wheelchair information, and health and activity history (25). Body weight was measured using a
129
MX490D wheelchair scalea. Body height was either self-reported or measured using a tapeb (21).
RI PT
126
130
Participants were asked to choose from a list of PAs and perform at least 10 PAs other than
132
resting (25). Many of the activities were performed at submaximal capacity limiting the order
133
effect on the EE measurement. The type of PAs included: propelling their wheelchair at self-
134
selected speeds on various surfaces, or up and down a ramp; being pushed in their wheelchair;
135
playing wheelchair basketball or darts; folding laundry; performing deskwork; using a resistance
136
band; and performing arm-ergometry at self-selected speeds and resistances. This list of
137
activities included a range of common everyday activities that involve different parts of the body
138
and varying levels of intensity. Participants were instructed to refrain from eating and exercising
139
at least 2 hours and 12 hours, respectively, prior to the experiment. The resting trial involved
140
collecting baseline EE for six minutes while the participants sat still in their wheelchairs.
M AN U
TE D
EP
141
SC
131
During testing, the participants wore a K4b2 portable metabolic cartc. Participants also wore a
143
PAMS-Arm and a PAMS-Wrist. All participants performed PAs in their own wheelchairs for a
144
minimum of 6 min, with at least a 3 min break between PAs. Participants rated each activity trial
145
on Borg’s modified rate of perceived exertion scale (possible scores: 6-20).
AC C
142
146 147
7
ACCEPTED MANUSCRIPT
Protocol in Home Environment
149
The follow-up session involved an activity session of 10 daily activities and a resting trial that
150
the participants were able to perform in their home environments. The participants were provided
151
with an opportunity to perform PAs that they perform on a regular basis to increase the
152
applicability of the algorithms to real-world scenarios. The PAs performed in addition to the PAs
153
mentioned in the laboratory environment were: propelling in their home or community on
154
various surfaces; watching television; preparing food or simulating cooking; simulating eating;
155
sweeping the floor or vacuuming; making a bed; washing dishes or laundry; wheelchair pushups;
156
using dumbbells or a handgrip; and other household activities. The other household activities
157
included deskwork, folding clothes, using a resistance band, and playing video games.
158
M AN U
SC
RI PT
148
Instrumentation and Data Collection
160
The criterion EE was collected from a K4b2 portable metabolic cart comprised of an analyzer
161
unit, a battery pack and a face mask covering the participant’s mouth and nose. The analyzer unit
162
measured the quantity of oxygen (O2) and carbon dioxide (CO2) to estimate EE per breath. The
163
K4b2 was calibrated for every participant or every six hours before use to ensure its accuracy.
164
The participants also wore a Polar-T31 heart rate monitord on their chest.
EP
AC C
165
TE D
159
166
The G-WRM is a self-enclosed rechargeable Bluetooth-based wireless device that contains six
167
reed switches and a two-axis gyroscope to measure angular velocity of the wheelchair wheel and
168
distance traveled (19). The G-WRM was secured to the spokes of the wheelchair wheel. The
169
accelerometer is a small Bluetooth®-based wireless accelerometer that captures tri-axial
170
acceleration of a body part using a capacitive micro-machined accelerometer (24). The
8
ACCEPTED MANUSCRIPT
accelerometers were worn on the right upper arm over the triceps muscle and wrist. One of the
172
investigators placed the accelerometers on all of the participants and made sure that the location
173
was consistent during testing sessions. Both the G-WRM and the accelerometers were calibrated
174
prior to the testing and were time synchronized by the Android cellphone that collects sensor
175
data. The Android phone was secured to the participant’s waist.
176
RI PT
171
Development of EE Estimation Models
178
First, the data were categorized into three groups including near-stationary PAs, PAs that might
179
involve wheelchair movement (1.8m/min ≥ distance travelled ≤ 12m/min), and PAs with
180
consistent wheelchair movement (25). The choice of the threshold was based on the distance
181
travelled by the participants for various PAs in this study. The near-stationary PAs and PAs with
182
consistent wheelchair movement were further subcategorized into: a) resting, arm-ergometry,
183
and other household activities, and b) wheelchair propulsion, caretaker pushing and basketball,
184
respectively. This process of grouping various PAs from the total dataset led us to develop EE
185
estimation models for seven activities. The EE estimation models used statistical measures such
186
as time and frequency domain features, which were calculated based on the data collected from
187
G-WRM and accelerometers (Supplementary Note 1). We also included the participant
188
parameters such as weight, height, gender, age, injury characteristics, wheelchair weight and
189
basal metabolic rates, in order to see if these variables were predictors of EE estimation for
190
various activities. New linear regression models were developed using ten-fold within-subject
191
cross validation (10-fold-CV) on 80% of the participants’ data (training dataset). The models
192
were then tested on the remaining 20% of the participants’ data (80-20CV) (25). Training and
193
testing datasets for the 80-20CV were prepared using a stratified approach with injury level
AC C
EP
TE D
M AN U
SC
177
9
ACCEPTED MANUSCRIPT
194
(paraplegia versus tetraplegia) and gender (male versus female) in order to allocate 80% of the
195
participants into the training dataset and 20% into the testing dataset (Supplementary Note 2).
196
Model Evaluation
198
The performance of the EE estimation models from PAMS-Arm and PAMS-Wrist were
199
compared with the criterion EE using mean signed error (MSE) and mean absolute error (MAE).
200
MSE is obtained by averaging the over and under estimated EE during a PA. Intraclass
201
correlation coefficients (ICC(3,1)) for single measure, using a two-way mixed model with
202
consistency, and Bland Altman plots (26) assessed the agreement between criterion EE and the
203
estimated EE. In addition, we evaluated the overall performance of PAMS-Arm and PAMS-
204
Wrist for the two-step process by sequentially applying the best classification algorithms (25)
205
and the regression models developed in this study. All statistical analysis was performed using
206
IBM SPSS Statistics softwaref, with a statistical significance at an alpha level of 0.05.
207 208
TE D
M AN U
SC
RI PT
197
Results
210
Table 1 shows the demographics of the participants. All participants completed the study with 37
211
of them finishing 10 activity trials, 7 finishing 9 trials, and 1 finishing 8 trials (25). Table 2
212
summarizes the metabolic costs, the Metabolic Equivalent of Task (MET) from the metabolic
213
cart, the MET-SCI (27), the heart rate, the rate of perceived exertion, and the number of
214
participants who completed each activity trial. Perceived exertion ranged from no exertion at all
215
(6.0) to somewhat hard (13.0).
AC C
EP
209
216
10
ACCEPTED MANUSCRIPT
Table 3 shows features of the seven activity-specific EE prediction models that were chosen by
218
regression analysis during 10-fold cross validation on training dataset. All the EE estimation
219
models overestimated EE using the 80% training dataset (Table 4) for PAMS-Arm and PAMS-
220
Wrist. The validation errors obtained from using the 20% testing dataset (Table 4) ranged from
221
overestimation to underestimation (positive percentages) for PAMS-Arm and PAMS-Wrist.
222
Bland Altman plots (Figures 1 and 2) for PAMS-Arm and PAMS-Wrist show that majority of
223
the EE estimated values for various wheelchair-related PAs lie within the band of mean ± 2SD.
224
The pattern also indicates that the higher EE values were underestimated compared to the
225
criterion EE. EE estimation performance (Table 5) for the PAMS-Arm and PAMS-Wrist on the
226
validation dataset for the two-step process of fist classifying the activity (25) and then applying
227
the corresponding activity-specific model indicated that both PAMS-Arm and PAMS-Wrist had
228
less than 10% error.
M AN U
SC
RI PT
217
230
TE D
229
Discussion
232
PA level measurement in wheelchair users could be used to provide self-tracking tools that could
233
lead to a healthier lifestyle. The novelty and contribution of this study to the field are: for the
234
first time we have concurrently captured both the wheelchair and arm movement towards
235
estimating EE in PAs performed in laboratory and community settings, and developed
236
classification based EE estimation models for wheelchair users that utilize sensor data collected
237
via smartphone. The EE estimation models chose demographic features such as weight of the
238
person, lean body mass, height and gender. This is not surprising because total EE during resting
239
and PAs for an individual depends on these parameters (28). However, the estimation models for
AC C
EP
231
11
ACCEPTED MANUSCRIPT
arm-ergometry activity for both PAMS-Arm and PAMS-Wrist did not use the demographic
241
characteristics for EE estimation, indicating that movement variables better explained the
242
variance in the EE for arm-ergometry. Activity-specific EE estimation errors using the training
243
dataset for PAMS-Arm and PAMS-Wrist were similar for most wheelchair-related PAs, with the
244
exception of arm-ergometry exercise. Higher EE error by PAMS-Arm compared to PAMS-Wrist
245
may result from smaller movement of the upper arm versus wrist during arm-ergometry. Also,
246
we found that the activity-specific model for both PAMS-Arm and PAMS-Wrist underestimated
247
EE for the PA that may involve wheelchair movement in the testing dataset. This error might be
248
due to variation in participants’ style of performing activities and the amount of wheelchair
249
propulsion necessary for these types of PAs.
M AN U
SC
RI PT
240
250
EE estimation error obtained by sequentially applying classification algorithms (25) and the
252
estimation models showed that the overall EE error based on MSE was lower than 10% for both
253
PAMS-Arm (-9.8%) and PAMS-Wrist (-5.7%). MSE is commonly used in sports sciences as an
254
overall performance indicator for PA monitors over a period of time (21). The low MSE
255
indicates that both PAMS-Arm and PAMS-Wrist can estimate EE with low bias (<10%) in
256
MWUs with SCI. Another important measure that needs to be considered during EE estimation
257
is the MAE, which provides information regarding the magnitude of error. The MAE values for
258
PAMS-Arm and PAMS-Wrist show that features, based on movement and demographic
259
variables, were able to estimate EE with low to moderately high error. Higher MSE (SD) and
260
MAE values in this study are due to EE errors estimated per minute for each PA compared to
261
using EE estimated per participant for each PA. Furthermore, the EE estimated for PAMS-Arm
262
and PAMS-Wrist had moderate to high ICC values for the majority of PAs, thus indicating that
AC C
EP
TE D
251
12
ACCEPTED MANUSCRIPT
the EE values estimated by PAMS-Arm and PAMS-Wrist are consistent with the EE measured
264
by the portable metabolic cart. Bland Altman plots for PAMS-Arm and PAMS-Wrist indicated
265
that the EE estimated by the new models was balanced with over and under estimation of EE for
266
up to 5kcal/min; the new models tended to underestimate EE for both PAMS-Arm and PAMS-
267
Wrist above 5kcal/min. This may indicate that movement sensing alone might not be adequate
268
for vigorous wheelchair based activities. Overall, the validation analyses indicated that both
269
PAMS-Arm and PAMS-Wrist can estimate EE with reasonable accuracy (<10% MSE) in MWUs
270
with SCI. In addition, similar EE estimation performance by PAMS-Arm and PAM-Wrist allow
271
the wheelchair users to either wear the accelerometer on the wrist or the upper arm.
M AN U
SC
RI PT
263
272
Even though direct comparison between our work and other studies (20-23) is not possible due to
274
differences in the protocols of PAs, we have compared the studies on certain aspects. The high
275
correlations for the estimated EE and the criterion EE for PAMS-Wrist (ICC=0.84) and PAMS-
276
Arm (ICC=0.69) have a similar pattern to Kiuchi et al. for upper arm (left: R2=0.75, right:
277
R2=0.87) and wrist (left: R2=0.86, right: R2=0.68) during wheelchair propulsion (20). Similarly,
278
high overall ICC values for PAMS-Wrist (0.89) are comparable to the results of Garcia-Masso et
279
al. and Nightingale et al. with a high correlation (r=0.86 (22), r= 0.93 (23)) for various
280
wheelchair related activities. Furthermore, the EE estimation errors for PAMS-Arm (MSE: -
281
32.0% for basketball to 12.1% for may be moving) during the seven PAs were much higher than
282
the EE estimation error for SenseWear (MSE: -4.3% for resting to 9.9% for Arm-ergometry)
283
during the four PAs from our previous study (21). The higher variation in our current study is
284
due to the following reasons: a much larger number of wheelchair-based PAs collected in both
285
laboratory and community settings; the merging of these numerous PAs into seven groups;
AC C
EP
TE D
273
13
ACCEPTED MANUSCRIPT
collection of data at two time points for 20 participants; and the participants performing PAs at a
287
self-selected pace or pattern.
288
Study Limitations
289
Some of the limitations of the study are a large percentage of our participants self-reported that
290
they were physically active on a regular or occasional basis, inclusion of individuals with SCI,
291
and EE models developed here were based on movement-based variables. The PA levels
292
reported by the participants of this study were much higher than reported in the community (29,
293
30). In future we plan to recruit a greater percentage of individuals with different disabilities
294
from the community so that the EE estimation models can be used for individuals with a variety
295
of disabilities who use wheelchairs. We also plan to evaluate real-time feedback provided by
296
PAMS to wheelchair users, which may assist them in pursuing behavior changes to achieve a
297
healthy and active lifestyle. Future studies should assess incorporating other forms of
298
physiologic sensing, such as galvanic skin response, skin temperature, near body temperature,
299
and heart rate, in order to quantify resistance-based PAs. Another limitation of the study is the
300
different number of participants for each activity trial due to participants performing PAs that
301
they perform on a regular basis in their home environments, and inability of the participants to
302
perform certain types of PAs, such as basketball, in their home environments. To address this
303
limitation we reduced the PAs from the total dataset into seven activities which led to decreased
304
variability of EE estimation error.
306
SC
M AN U
TE D
EP
AC C
305
RI PT
286
307
Conclusions
308
This study developed and validated new EE estimation models for MWUs based on upper arm,
309
wrist and wheelchair movements detected with the help of a physical activity monitoring system. 14
ACCEPTED MANUSCRIPT
310
The new models we developed can estimate PA levels in MWUs with SCI in laboratory and
311
community settings.
312
RI PT
313 314
Suppliers list
315
a. Befour MX490D wheelchair scale; Befour, Inc, 102 N Progress Dr, Saukville, WI 53080, USA
SC
316
b. Stanley Tape; Stanley Works, 480 Myrtle Street, New Britain, CT 06053, USA
318
c. K4b2 portable metabolic cart; COSMED srl, Via dei Piani di Mt. Savello 37, Pavona di
322 323 324 325 326
Success, NY 11042, USA
e. MATLAB® version 2013a; The Mathworks, Inc., 3 Apple Hill Drive Natick, MA, 01760, USA, USA
TE D
321
d. Polar-T31 heart rate monitor; Polar Electro Inc., 1111 Marcus Avenue, Suite M15 Lake
f. IBM SPSS Statistics software version 20.0; IBM Corporation, 1 New Orchard Road Armonk, NY, 10504, USA
EP
320
Albano, Rome 0004, Italy
AC C
319
M AN U
317
15
ACCEPTED MANUSCRIPT
References
328 329
1. Durstine JL, Painter P, Franklin BA, Morgan D, Pitetti KH, Roberts SO. Physical activity for the chronically ill and disabled. Sports Med. 2000;30(3):207-19.
330 331
2. Hoenig H, Landerman LR, Shipp KM, George L. Activity restriction among wheelchair users. Journal of American Geriatrics Society. 2003;51(9):1244-51.
332 333 334
3. Glaser RM, Janssen TWJ, Suryaprasad AG, Gupta SC, Mathews T. The Physiology of Exercise. In: Apple DF, editor. Physical Fitness: A Guide for Individuals with Spinal Cord Injury. Washington, DC: Department of Veterans Affairs; 1996.
335 336
4. Rimmer JH. Use of the ICF in identifying factors that impact participation in physical activity/rehabilitation among people with disabilities. Disabil Rehabil. 2006;28(17):1087 – 95.
337 338
5. Jacobs PL, Nash MS, Rusinowski JW. Circuit training provides cardiorespiratory and strength benefits in persons with paraplegia. Med Sci Sports Exerc. 2001;33:711-7.
339 340
6. Rimmer JH, Yamaki K, Davis BM, Wang E, Vogel LC. Obesity and Overweight Prevalence Among Adolescents With Disabilities. Preventing Chronic Disease. 2011;8(2):1-6.
341 342
7. Rimmer JH, Shenoy SS. Impact of Exercise on Targeted Secondary Conditions. Field MJ, Jette AM, Martin L, editors. Washington, DC: National Academy of Sciences; 2006.
343 344 345
8. Jakicic JM, Tate DF, Lang W, Davis KK, Polzien K, Rickman AD, et al. Effect of a stepped-care intervention approach on weight loss in adults: a randomized clinical trial. Journal of the American Medical Association. 2012;307(24):2617-26.
346 347 348
9. Coons MJ, DeMott A, Buscemi J, Duncan JM, Pellegrini CA, Steglitz J, et al. Technology Interventions to Curb Obesity: A Systematic Review of the Current Literature. Current cardiovascular risk reports. 2012;6:120-34.
349 350 351
10. Alhassan S, Kim S, Bersamin A, King A, Gardner C. Dietary adherence and weight loss success among overweight women: results from the A to Z weight loss study. Int J Obes. 2008;32:985-91.
352 353
11. Baker RC, Kirschenbaum DS. Self-monitoring may be necessary for successful weight control. Behavior Therapy. 1993;24:377–94.
AC C
EP
TE D
M AN U
SC
RI PT
327
16
ACCEPTED MANUSCRIPT
12. 55.
Wing R, Phelan S. Long-term weight loss maintenance. Am J Clin Nutr. 2005;82:2225–
356 357 358
13. Shuger SL, Barry VW, Sui X, McClain A, Hand GA, Wilcox S, et al. Electronic feedback in a diet- and physical activity-based lifestyle intervention for weight loss: a randomized controlled trial. International Journal of Behavioral Nutrition and Physical Activity. 2011;8(41).
359 360 361
14. Spring B, Duncan JM, Janke EA, Kozak AT, McFadden HG, Demott A, et al. Integrating technology into standard weight loss treatment a randomized controlled trial. JAMA Internal Medicine. 2013;173(2):105-11.
362 363 364
15. Tolerico ML, Ding D, Cooper RA, Spaeth DM, Fitzgerald SG, Cooper R, et al. Assessing mobility characteristics and activity levels of manual wheelchair users. J Rehabil R D. 2007;44(4):561-72.
365 366
16. Coulter EH, Dall PM, Rochester L, Hasler JP, Granat MH. Development and validation of a physical activity monitor for use on a wheelchair. Journal of Spinal Cord. 2011;49:445-50.
367 368
17. Sonenblum SE, Sprigle S, Caspall J, Lopez R. Validation of an accelerometer-based method to measure the use of manual wheelchairs. Med Eng Phys. 2012;34:781-86.
369 370
18. Conger SA, Scott SN, Bassett DR. Predicting energy expenditure through hand rim propulsion power output in individuals who use wheelchairs. Br J Sports Med. 2014;48:1048-53.
371 372
19. Hiremath SV, Ding D, Cooper RA. Development and evaluation of a gyroscope based wheel rotation monitor for manual wheelchair users. Spinal Cord Medicine. 2013;36(4):347-56.
373 374 375 376
20. Kiuchi K, Inayama T, Muraoka Y, Ikemoto S, Uemura O, Mizuno K. Preliminary study for the assessment of physical activity using a triaxial accelerometer with a gyro sensor on the upper limbs of subjects with paraplegia driving a wheelchair on a treadmill. Spinal Cord. 2014;52:556-63.
377 378 379
21. Hiremath SV, Ding D, Farringdon J, Cooper RA. Predicting energy expenditure of manual wheelchair users with spinal cord injury using a multi-sensor based activity monitor. Arch Phys Med Rehabil. 2012;93(11):1937-43.
AC C
EP
TE D
M AN U
SC
RI PT
354 355
17
ACCEPTED MANUSCRIPT
22. García-Massó X, Serra-Añó P, García-Raffi L, Sánchez-Pérez E, López-Pascual J, Gonzalez L. Validation of the use of Actigraph GT3X accelerometers to estimate energy expenditure in full time manual wheelchair users with spinal cord injury. Spinal Cord. 2013;51(12):898-903.
384 385
23. Nightingale TE, Walhin J-P, Thompson D, Bilzon JL. Predicting Physical Activity Energy Expenditure in Manual Wheelchair Users. Med Sci Sports Exerc. 2014;46(9):1849-58.
386 387 388 389 390
24. Intille SS, Albinali F, Mota S, Kuris B, Botana P, Haskell WL, editors. Design of a Wearable Physical Activity Monitoring System using Mobile Phones and Accelerometers. Proceedings of Annual International Conference of the IEEE Engineering in Medicine and Biology Society 2011 Aug 30- Sep 3; Boston, MA: IEEE Engineering in Medicine and Biology Society.
391 392
25. Hiremath SV, Intille SS, Kelleher A, Cooper RA, Ding D. Detection of physical activities using a physical activity monitor system for wheelchair users. Med Eng Phys. 2015;37(1):68-76.
393 394
26. Bland JM, Altman DG. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet. 1986;1:307-10.
395 396
27. Collins EG, Gater D, Kiratli J, Butler J, Hanson K, Langbein WE. Energy cost of physical activities in persons with spinal cord injury. Med Sci Sports Exerc. 2010;42(4):691-700.
397 398 399
28. American College of Sports Medicine, Mitchell H. Whaley, Peter H. Brubaker, Robert Michael Otto, Lawrence E. Armstrong ACSM's guidelines for exercise testing and prescription. 7, illustrated ed: Lippincott Williams & Wilkins; 2005.
400 401
29. Washburn R, Hedrick BN. Descriptive epidemiology of physical activity in university graduates with locomotor disabilities. Int J Rehabil Res. 1997;20(3):275-87.
402 403 404
30. Tasiemski T, Kennedy P, Gardner BP, Taylor N. The association of sports and physical recreation with life satisfaction in a community sample of people with spinal cord injuries. NeuroRehabilitation. 2005;20:253-65.
AC C
EP
TE D
M AN U
SC
RI PT
380 381 382 383
405 406 407 408 409
18
ACCEPTED MANUSCRIPT
410
Figure Legends
411
Figure 1: Bland Altman plot of EE estimated using activity-specific models for PAMS-Arm and
413
EE measured for the various wheelchair-related PAs in the validation dataset. The x and y axes
414
represents the mean and difference of the EE estimated and the EE measured, respectively.
RI PT
412
415
Figure 2: Bland Altman plot of EE estimated using activity-specific models for PAMS-Wrist and
417
EE measured for the various wheelchair-related PAs in the validation dataset. The x and y axes
418
represents the mean and difference of the EE estimated and the EE measured, respectively.
M AN U
SC
416
419 420
AC C
EP
TE D
421
19
ACCEPTED MANUSCRIPT
Table 1: Demographic characteristics of the participants.
39 6 41.0 (12.6)* 78.1 (18.1) 1.8 (0.1) C5 to L5 13 32
36 5 4 31 14
AC C
EP
TE D
Note: * Values are n or mean (SD).
SC
22 23 12.6 (8.6)
RI PT
Values 45
M AN U
Variables Number of participants Sex Male Female Age (years) Weight (kg) Height (m) Injury level (range) Injury between C5 and C8 Injury between T3 and L5 Injury completeness Complete Incomplete Manual wheelchair usage (years) Self-reported PA Regular Occasional No regular PA Self-reported smokers Non-smokers Smokers
ACCEPTED MANUSCRIPT
Table 2: Metabolic costs in terms of EE, METs, heart rate, rate of perceived exertion, and number of participants and minutes per
RI PT
activity for various wheelchair-based physical activities. The MET-SCI was calculated based on previous research in individuals with SCI (27) by using a reference metabolic rate (VO2/Kg) of 2.7ml/min/kg for SCI population. No. of Participants
No. of Min
Heart Rate in beats/min Mean
SD
EE from K4b2 in kcal/min Mean SD
MET
Mean
MET-SCI
SC
Activity Trial
Rate of Perceived Exertion Mean SD
SD
Mean
SD
0.2
1.1
0.3
6.0
0.1
1.0
3.0
1.3
10.9
2.2
43
363
69.9
16.2
1.1
0.3
0.9
Arm-Ergometry
43
500
96.0
19.0
3.0
1.2
2.3
Darts
33
214
91.1
15.3
2.7
0.8
2.1
0.6
2.7
0.8
8.7
2.0
Deskwork
43
574
79.1
14.2
1.5
0.6
1.2
0.5
1.6
0.6
7.5
2.1
Folding Clothes
42
343
92.5
17.6
2.3
0.6
1.8
0.4
2.3
0.6
8.6
2.3
Propulsion
43
901
101.2
20.5
3.5
1.5
2.7
1.1
3.5
1.5
11.0
3.1
Caretaker Pushing
42
341
75.3
14.2
1.3
0.4
1.0
0.3
1.3
0.4
6.4
1.3
Resistance
43
367
86.2
15.2
2.0
0.7
1.6
0.5
2.0
0.7
10.0
2.3
Basketball
19
112
110.2
19.7
4.5
1.7
3.7
1.2
4.8
1.6
12.6
2.7
Eating
17
17
73.0
13.2
2.0
0.4
1.6
0.6
2.1
0.8
7.5
2.1
Sweeping Floor
14
90
96.2
15.8
3.0
0.9
2.4
0.8
3.1
1.1
10.9
3.1
Preparing Food
11
68
87.6
17.9
2.3
0.6
1.8
0.5
2.3
0.7
7.7
1.6
Making Bed
1
6
90.6
7.3
2.7
0.6
2.3
0.5
3.0
0.6
7
0.0
Cleaning Room
4
26
97.8
27.7
2.3
0.6
2.5
0.6
3.2
0.7
8.7
2.9
Filing papers
2
12
91.8
16.1
1.1
0.2
1.5
0.2
1.9
0.3
7.5
2.1
Check mail
2
8
89.9
12.9
2.3
0.6
2.1
0.4
2.7
0.6
6.0
0.0
Laundry
2
16
89.2
7.0
2.7
0.4
2.7
0.3
3.5
0.4
9.0
2.6
Video Game
1
6
72.3
2.8
1.9
0.2
1.2
0.1
1.6
0.2
13.0
0.0
Cleaning Car Wheelchair Pushups
1
6
76.6
7.0
2.7
0.6
1.9
0.4
2.5
0.5
7.0
0.0
1
6
97.0
15.4
2.6
0.7
2.9
0.7
3.7
1.0
13.0
0.0
AC C
EP
TE D
M AN U
Resting
ACCEPTED MANUSCRIPT
Table 3: Features of the activity-specific EE estimation models that were chosen by regression analysis during 10-fold cross validation on training dataset (80% of participants’ data).
Caretaker Pushing Basketball May be moving Resting Armergometry
PAMSWrist OA not moving
Propulsion Caretaker Pushing Basketball May be moving
RI PT
SC
Propulsion
M AN U
PAMSArm
OA not moving
TE D
Armergometry
EP
Resting
Features Lean body mass, Mean of velocity from G-WRM, Entropy without the DC component of z axis acceleration (acc.) from arm accelerometer (Arm), Dominant frequency content (range: 0.3-15 Hz) of x axis acc. from Arm, Root mean square of z axis acc. from Arm Back trend of resultant acc. from Arm, Ratio of dominant frequency's power with total power for resultant acc. from Arm, Entropy without the DC component of y axis acc. from Arm, Number of peaks for z axis acc. from Arm., Energy content of velocity from GWRM Mean absolute deviation of z axis acc. from Arm multiplied by square root of height, Gender, Standard deviation of z axis acc. from Arm, Standard deviation of six minutes or less for z axis acc. from Arm, Dominant frequency content’s power of x axis acc. from Arm Ratio of dominant frequency's power with total power for resultant acc. from Arm, Second dominant frequency content's power of resultant acc. from Arm, Mass to the power of 0.75, Ratio of dominant frequency for the current and the past minute of velocity from GWRM, Dominant frequency content’s power of velocity from G-WRM Lean body mass, Energy without the DC component of z axis acc. from Arm, Entropy without the DC component of x axis acc. from Arm, Mean cross rate of z axis acc. from Arm, Entropy without the DC component of y axis acc. from Arm Back trend of velocity from G-WRM, Lean body mass, Mean absolute deviation with respect to median of resultant acc. from Arm, Second dominant frequency content's power of x axis acc. from Arm, Entropy without the DC component of resultant acc. from Arm Mean absolute deviation with respect to mean of x axis acc. from Arm, Lean body mass, Standard deviation of six minutes or less for x axis acc. from Arm, Second dominant frequency content's power of y axis acc. from Arm, Mean absolute deviation with respect to median of x axis acc. from Arm Lean body mass, Difference between mean values of acc. from x and y axes from wrist accelerometer (Wrist), Ratio of dominant frequency for the current and the past minute of resultant acc. from Wrist, Mean cross rate of resultant acc. from Wrist, Zero cross rate of x axis acc. from Wrist Sum of mean absolute deviation in X and Y axes acc. from Wrist, mean of velocity from G-WRM, Number of peaks for x axis acc. from Wrist, Ratio of dominant frequency for the current and the past minute of x axis acc. from Wrist, Energy without the DC component of x axis acc. from Wrist Number of peaks for y axis acc. from Wrist, Gender, Total power for frequencies (range: 0.3-15 Hz) of y axis acc. from Wrist, Standard deviation of six minutes or less for resultant acc. from Wrist, Entropy content of y axis acc. from Wrist Third dominant frequency content's power for velocity from G-WRM, Back trend of velocity from G-WRM, Mass to the power of 0.75, Correlation between y axis acc. and resultant acc. from Wrist, Total power for frequencies for resultant acc. from Wrist Lean body mass, Second dominant frequency content's power of z axis acc. from Wrist, Standard deviation of six minutes or less for velocity from G-WRM, Correlation between x axis acc. and resultant acc. from Wrist, Dominant frequency content of z axis acc. from Wrist Back trend of velocity from G-WRM, Number of peaks for resultant acc. from Wrist, Number of peaks for velocity from G-WRM, Mufflin basal metabolic rate, Second dominant frequency content's power of x axis acc. from wrist Lean body mass, amplitude of y axis acc. from Wrist, Standard deviation of six minutes or less for z axis acc. from Wrist, World Health Organization resting metabolic rate divided by lean body mass, World Health Organization resting metabolic rate
AC C
Device Activity
ACCEPTED MANUSCRIPT
Table 4: Mean signed error (%) of EE estimation for each activity-specific model using the features obtained by 10-fold cross validation on 80% of participants’ data (training dataset) and tested on the remaining 20% of participants’ data not used for training (testing data).
-14.44
49.94
-1.66
42.32
-7.69
33.29
-10.14
38.38
-11.86
33.70
-9.72
36.79
-7.50
27.27
-11.65
12.84
-5.15
-6.66
29.28
-0.96
39.33
-7.45
-1.85
35.97
-16.21
16.79
-2.14
-6.64
16.68
29.82
37.71
-7.20
-7.66
33.75
-1.00
13.10
-6.54
RI PT
Resting Armergometry OA Not moving Propulsion Caretaker pushing Basketball May be moving Overall
PAMS-Wrist Train Test Mean SD Mean SD -6.82 34.36 -6.94 20.33 5.34
33.00
-12.61
32.78
SC
Activity
PAMS-Arm Train Test Mean SD Mean SD -7.05 35.36 -5.32 19.75
24.21
-7.81
31.35
36.23
-4.77
17.20
19.80
-11.45
30.97
28.53
22.38
13.61
30.28
-0.77
11.28
M AN U
Device
Note: Other PAs (OA) while being stationary; Overall error or combined error (Overall); Negative percentages indicate overestimation where the models estimated an EE value higher
AC C
EP
TE D
than the EE measured by K4b2; positive percentages indicate underestimation.
ACCEPTED MANUSCRIPT
Table 5: EE estimation performance for activity-specific (AS) models on the validation dataset (20% of participants’ data not used for training) for the two-step process of fist classifying the
Wrist. MAE Mean
MSE Mean
SD
ICC
ICC(3,1) 95% CI LB UB
P
1.09
0.27
1.21
0.21
19.41
-14.02 19.76
0.77
0.48
0.90
<0.05
3.25
1.27
2.98
0.66
31.83
-2.13
42.47
0.47
0.19
0.66
<0.05
1.86
0.85
1.93
0.78
30.47
-15.04 35.13
0.73
0.64
0.80
<0.05
3.65
1.99
3.47
0.93
31.69
-11.56 39.64
0.69
0.57
0.78
<0.05
1.3
0.39
1.25
18.46
0.82
0.68
0.90
<0.05
3.57
1.31
4.61
3.66
1.02
2.90
2.65 1.22
1.63 0.57
2.58 1.17
3.15
1.27
2.87
1.91
SC
Resting Armergometry OA not moving Propulsion Caretaker Pushing Basketball May be moving Overall Resting Armergometry OA not moving Propulsion Caretaker Pushing Basketball May be moving Overall
0.28
14.10
0.25
1.62
35.18
-31.96 38.30
0.84
0.56
0.94
<0.05
0.34
27.69
12.14
37.10
0.40
-0.47
0.76
0.13
1.19 0.26
29.04 21.20
-9.82 -4.00
37.03 24.82
0.82 0.62
0.79 0.30
0.85 0.80
<0.05 <0.05
1.05
23.29
3.85
32.40
0.85
0.77
0.90
<0.05
0.86
1.91
0.39
28.73
-12.32 33.50
0.63
0.48
0.73
<0.05
2.01
3.50
1.21
26.09
-8.39
33.37
0.84
0.78
0.89
<0.05
0.64
1.31
0.27
16.28
-1.51
20.81
0.68
0.43
0.82
<0.05
1.30
4.30
2.02
28.64
-14.77 29.98
0.91
0.75
0.97
<0.05
3.66
1.02
2.97
0.32
29.69
9.26
45.86
0.61
0.05
0.84
0.02
2.65
1.63
2.53
1.26
25.19
-5.65
32.61
0.89
0.87
0.91
<0.05
3.64 1.4 3.61
AC C
EP
PAMSWrist
EE from K4b2 EE AS in kcal/min Mean SD Mean SD
M AN U
PAMSArm
Activity
TE D
Device
RI PT
activity and then applying the corresponding activity-specific model for PAMS-Arm and PAMS-
Note: EE Activity Specific (AS); Mean Absolute Error in percentage (MAE); Mean Signed Error (MSE); Intraclass correlation coefficients (ICC); ICC with 95% confidence interval (CI) – lower bound (LB) and upper bound (UB); ICC significance values (p); Overall error or combined error (Overall).
AC C
EP
TE D
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
AC C
EP
TE D
M AN U
SC
RI PT
ACCEPTED MANUSCRIPT
ACCEPTED MANUSCRIPT
Supplementary Note 1: The time domain features, such as mean, mean absolute deviation, and peaks were simple to extract and can be used to estimate EE of PAs that are considerably different. The frequency domain features, such as total power between a band of frequencies,
RI PT
energy, and entropy can be used to estimate EE based on the key frequency of movement
(wheelchair propulsion and arm-ergometry). The features were extracted using custom programs written in MATLAB® (version 2013a) for a 1-min window size to be consistent with the EE
SC
estimation.
M AN U
Supplementary Note 2: The 10-fold-CV is a statistical technique used as part of the development process to select variables and optimize the regression model. The process involves generating the regression models iteratively, by selecting and adding features, to arrive at a final set of variables that had the least EE estimation error for the 80% of the data used in training. Then
TE D
once we developed and saved the final model on the training dataset we tested the final model on
AC C
EP
the testing dataset, which was not used in building the model.