Content uploaded by Daniel Austin
Author content
All content in this area was uploaded by Daniel Austin
Content may be subject to copyright.
Unobtrusive and Ubiquitous In-Home Monitoring: A
Methodology for Continuous Assessment of Gait Velocity in
Elders
Stuart Hagler,
Biomedical Engineering division, Oregon Health and Science University, Portland, OR 97239
USA, (haglers@bme.ogi.edu)
Daniel Austin [Member, IEEE],
Biomedical Engineering division, Oregon Health and Science University, Portland, OR 97239
USA, (austidan@bme.ogi.edu)
Tamara L. Hayes [Member, IEEE],
Biomedical Engineering division, Oregon Health and Science University, Portland, OR 97239
USA, (hayest@bme.ogi.edu)
Jeffrey Kaye [Member, IEEE], and
Department of Neurology, Oregon Health and Science University, Portland, OR 97239 USA
(kaye@ohsu.edu)
Misha Pavel [Member, IEEE]
Biomedical Engineering division, Oregon Health and Science University, Portland, OR 97239
USA, (pavel@bme.ogi.edu)
Abstract
Gait velocity has been shown to quantitatively estimate risk of future hospitalization, has been
shown to be a predictor of disability, and has been shown to slow prior to cognitive decline. In this
paper, we describe a system for continuous and unobtrusive in-home assessment of gait velocity, a
critical metric of function. This system is based on estimating walking speed from noisy time and
location data collected by a “sensor line” of restricted view passive infrared (PIR) motion
detectors. We demonstrate the validity of our system by comparing with measurements from the
commercially available GAITRite® Walkway System gait mat. We present the data from 882
walks from 27 subjects walking at three different subject-paced speeds (encouraged to walk
slowly, normal speed, or fast) in two directions through a sensor line. The experimental results
show that the uncalibrated system accuracy (average error) of estimated velocity was 7.1cm/s (SD
= 11.3cm/s), which improved to 1.1cm/s (SD = 9.1cm/s) after a simple calibration procedure.
Based on the average measured walking speed of 102 cm/s our system had an average error of less
than 7% without calibration and 1.1% with calibration.
Index Terms
Eldercare; unobtrusive monitoring; ubiquitous computing; gait; walking speed; passive infrared
(PIR) motion detectors
Correspondence to: Daniel Austin.
NIH Public Access
Author Manuscript
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
Published in final edited form as:
IEEE Trans Biomed Eng
. 2010 April ; 57(4): 813–820. doi:10.1109/TBME.2009.2036732.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
I. Introduction
Improving quality of life and providing adequate medical care for the rising number of
elderly while keeping health care costs under control has in recent years become a major
problem [1]. Several methodologies have been proposed to address this problem that include
the use of technology to develop systems that promote aging in place [2] and the use of
pervasive healthcare [3] to help alleviate the burden placed on health care providers. One of
the underlying themes of these approaches is to employ technology such as wireless
networks combined with novel sensing systems to gather and interpret data in non-health
care settings such as the home environment.
Many systems have been proposed that use these methodologies to assist the elderly. For
example, passive infrared sensors have been used in-home for the estimation of amount [4]
and type [5] of daily activity and in-hospital for classification of patient movements [6].
Other systems have been proposed that detect falls in elders [7], infer activities of daily
living (ADLs) [8], use computer interactions to detect cognitive changes [9], and for
continuous and unobtrusive in-home behavioral monitoring [10]. Other recent applications
of pervasive healthcare and wireless sensor networks for supporting elder healthcare for
aging in place include multimodal sensing and computer vision [11,12] while systems for
supporting independence in assisted living are described in [13,14]. For completeness we
mention that a comprehensive literature review of pervasive computing in health care from
2002 to 2006 is available [15] including applications to eldercare.
One specific measure of particular interest for unobtrusive assessment for health monitoring
is walking speed. Walking speed has been shown to be a quantitative estimate of risk of
future hospitalization [16]. Slower walking speed has been demonstrated in dementia
patients compared to controls [17] and has been shown to precede cognitive impairment [18]
and dementia [19], and timed walk has been used as a partial characterization of lower
extremity function which has been shown to predict disability [20,21]. Other studies have
shown a relationship between walking speed and cognition [22,23]. Current evaluation of
walking speed is typically done both infrequently and in the clinic setting which suffers
from at least five shortcomings. First, frequent assessment visits are impractical and cost
prohibitive since either it is difficult for patients to make frequent trips to a doctor's office or
other clinical settings or in the case of research assessments, inconvenient for the research
team to visit homes frequently. Second, for longitudinal study each testing session is
typically scheduled in increments of six months or a year after a baseline visit making it
difficult both to evaluate the validity or stability of baseline measurements and to detect
short and long term variability [24]. Third, there may be an intentional [16] change in
walking speed in the clinical setting or an unintentional [24] change in abilities during a
single assessment. These pacing considerations themselves may have important implications
for predicting outcomes [23]. Fourth, infrequent measurements report only the net change
between measurement times and cannot distinguish between functional changes occurring
slowly over time and abrupt functional changes, which may have different causes. Fifth,
infrequent measurements do not detect changes when they happen which may reduce the
ability of a clinician to provide intervention or reduce the effect of an intervention. By
shifting to continuous in-home monitoring of walking speed from the current paradigm, the
effect of all of these short comings can be significantly reduced or removed.
There have been several systems proposed for monitoring walking speed and other gait
features outside of the clinical setting [25-27]. These systems typically consist of some
wearable combination of gyroscopes and/or accelerometers and have demonstrated accuracy
and precision in the field. However, these systems suffer from several limitations such as
short battery life, the need to download the data or introduce additional hardware and
Hagler et al. Page 2
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
complexity for wireless data collection, and the inconvenience of both a wearable device
and having to remember to wear a device. For these reasons the wearable devices are
currently inadequate for long-term, in-home, unobtrusive monitoring.
In order to address these concerns and to improve diagnostic ability for clinicians and
researchers, we propose a methodology for continuous in-home monitoring of walking
speed using passive infrared motion sensors. Specifically, we describe the hardware
preparation and deployment, the techniques for data collection, and the data processing
algorithms for continuous in-home assessment of walking speed in elders. Finally, we
validate our approach by comparing the results of our method for walking speed estimation
with the commercially available GAITRite® Walkway System gait mat.
II. System Description and Data Collection
In this section, we describe the hardware and methodology used to deploy the walking speed
measurement system in a residence. A partial description of this system has been described
elsewhere [4,10] in the more general context of total activity monitoring, as has a simpler
version of the proposed approach [28]. Here we specialize and describe in more detail the
specific nature of the walking speed measurement system. We begin by describing the
sensors and how they are placed in a residence and follow with a description of the wireless
network based data collection.
To detect motion we used the X10 model MS16A (X10.com) passive infrared motion sensor
which emits a unique programmable bit code at 310 MHz when motion is detected. We
restricted the field of view of each motion sensor to ±4 degrees and installed four sensors
sequentially on the ceiling (average height of 2.54 m) approximately 61cm apart in a
confined area such as a hallway or other corridor. This combination tends to force a resident
to walk linearly through each sensor pair in the sensor line and ensures that each sensor will
only fire when someone passes directly below. Limiting the field of view precisely and
placing sensors in exact locations is not possible, and therefore there is some variability in
the physical locations which cause the sensors to fire, as will be discussed shortly. Fig.1
shows how these sensors look from a resident's point of view when entering the sensor line.
To collect the wirelessly transmitted sensor firings, we use a WGL 800 wireless transceiver
connected to a desktop computer installed in the residence. Simultaneous sensor firings or
other interfering sources can result in lost data due to collisions at the wireless transceiver.
However, these have been shown to be minor, with a less than 2% overall data loss [4]. The
computer timestamps the sensor firings and the data pair is both stored locally and sent via a
secure Internet connection to a central database for analysis.
Our experience with the described technology comes from the deployment and monitoring
of approximately 250 Portland (OR, USA) metropolitan area homes and retirement
community dwellings from between 6 months to over 2 years in ongoing studies. We have
instrumented both single and multi person dwellings and have collected data from over
1,200,000 walking events from single person homes with minimal technical challenge or
sophistication needed for setup. Installation of the complete system (including additional
technologies described elsewhere [4]) takes an average of 1.5 to 2 hours with 2 people.
Deploying only the equipment necessary to measure walking speed (computer, sensors,
wireless transceiver, and internet) is estimated to take 1 person approximately 1 hour – if the
home already contains an Internet-connected computer, this could be done in 20 minutes.
The technologies are managed remotely using custom systems management software that
supports data viewing, remote software updates, and remote computer reboots if needed.
Other issues, such as replacing motion sensor batteries (battery life is about 1 year) or
Hagler et al. Page 3
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
changing sensors if they become unreliable or defective can typically be handled in a very
short visit to the residence, typically 10-15 minutes. Overall, we have found the system has
been simple to install, unobtrusive in the sense of both passive sensor technology and
minimal outside intervention, and easy to maintain.
III. Data Modeling and Analysis
In this section we introduce and discuss both the proposed linear model and the estimator for
determining the gait velocity from noisy motion sensor data. We start by describing how to
determine the precise spatial separation of the sensors from the sensor firings since they will
not, as mentioned, be the same as the measured values due to a combination of installation
variability and differences between individual sensors. We then use this information to
model the walking speed as a linear function of the measured data degraded by two sources
of additive noise. The first source of noise is in units of distance and is due to the sensor
firing in slightly different locations during each pass through the sensor line. This error is
based on the field of view and sensitivity of an individual sensor. The second source of error
is in units of time and is due to the discrepancy between when a sensor fires and when the
computer timestamps the firing, which generally causes positive time errors. We conclude
the section by proposing a walking speed estimator that minimizes the combination of these
errors followed by a discussion of model calibration in the presence of ground-truth data
versus estimating the calibration factor when ground-truth data is unavailable.
A. The Linear Model
We start by assuming the sensors are placed at physical positions {x̃i} in some spatial
coordinate system. For a particular walking event the sensors fire at times where k
indexes the particular sensor line walking event and i indexes the particular sensor which
fired. We then define {xi} to be the average position at which the walker is when the ith
sensor fires. In other words, for a particular walking event the ith sensor fires when the
walker is at some random location {xi+εi} with the errors {εi} being independent random
variables with zero expected value. Fig. 2 illustrates this arrangement.
The differences {xi – x̃i} represent likely biases due to the field of view of the sensor and the
direction of movement as shown in fig. 2. For the sake of simplicity we restrict the present
discussion to the analysis of movement in one direction.
Now assuming a walker moves with some known velocity ν through a sensor line and we
have some absolute reference time, {τi} can be defined as the time at which the walker is
expected to be at location {xi} and trigger the ith sensor. If we now include the errors in
detection location {εi} explicitly in the measured time we find that the measured times
should be .
By further assuming that there is some random delay {ηi} between when a sensor fires at
location {xi + εi} and when the computer time-stamps the sensor firing data, the measured
time can be written as where the {ηi} are independent random variables
with some common non-negative expected value and the explicit dependence on the
measured time from both position errors and time errors is shown.
Hagler et al. Page 4
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
Now, consider an event k to comprise a person walking past the line of sensors with a
constant, but unknown velocity νk. Then for any pair of sensors, i,j we have
, from which we find:
(1)
Using (1) we can solve for the velocity νk, for any three sensors i,j,m:
(2)
We now economize the notation and define and for i,j,m.
Rewriting (2) yields:
(3)
Taking the expectation of both sides and using the facts that: , , and
for i ≠ j, results in:
(4)
which simplifies to:
(5)
From this we may conclude that:
(6)
The expected value of the random variable can therefore be used to estimate the spatial
separation of the sensors up to a scale factor by computing the average values over a large
number of events.
By explicitly writing (6) with the proportionality constant c we have:
Hagler et al. Page 5
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
(7)
Here c has a ready interpretation as the speed a person would have to walk in order for the
sensors to register time differences equal to the average time differences calculated from the
training data set.
Let us look more closely at the estimated spacings (xj – xi). Considering fig. 2 again, we see
that the sensor line is effectively hovering at some height between the ceiling and the floor.
In the figure it has been drawn at the top of the head. Let us imagine that we knew the actual
mean firing position for each sensor as a function of the height above the floor, so that if the
sensor is triggered by motion at a height h it will typically be triggered at a position xi (h). In
addition, the body itself does not move at a single constant velocity ν during gait, but rather
different segments move with various velocities over the course of the gait cycle. The effect
of this is that the values {τi} (which are the actual measurements) reflect triggering at
various heights due to different body segments. In effect, all these factors are averaged over
to produce effective sensor spacing based on the subject's height and style of walking
together with the sensor characteristics. We expect that for people of similar heights, as well
as reasonably similar styles of gait, that the estimated effective sensor spacing (xj – xi)
should be close in value.
B. Estimation of Gait Velocity
When we estimate the gait velocity we must consider two sources of measurement error.
First, there is the combined error for sensors i,j resulting from detecting the walker at
positions away from the mean detection locations xi, xj which we denote by εij Second, there
is the combined error for sensors i,j resulting from errors in time-stamping the moments at
which each sensor fired. This is represented by ηij. In general, these two types of errors
should be given different weights in accordance with the relative variability captured by the
spatial and temporal covariance matrices of the error terms εij and ηij [29]. The relative
weighting of the temporal error term relative to the spatial error term is represented by the
parameter ρ. We proceed initially by assuming that the calibration factor c and the weighting
factor ρ are known values and derive an estimator for the walking velocity, ν
̂
through the
sensor line. With this estimator, we then proceed to consider situations in which we know
the actual walking velocity, ν and consider the estimator now as a function, ν
̂
(c,ρ), which
arises when one calibrates the line using a set of data where the velocities are known. Finally
we consider the case where c,ρ are not known and use information from the physical set up
of the sensors to estimate a value of c. In this last case we assume that weighting of the
errors is a general value across all reasonably similar sensor lines and so take the value for ρ
obtained from our calibrated experiment described in the next section.
In a sensor line of four sensors i,j,l,m that fire sequentially when a subject walks along the
sensor line, we can estimate the walking speed by minimizing the overall error in the
dependent and independent variables using the method of total least squares [30] applied to
the model:
(8)
Hagler et al. Page 6
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
where we have rewritten the linear equations in matrix notation, used the fact that
to keep the noise term in (8) in the standard additive form, and used the estimate of the
spatial sensor separation as in (7). We proceed assuming that c,ρ are fixed and known
constants for the sensor line.
To compute the velocity estimates, we now construct a matrix containing as column vectors
the distances between adjacent sensors, and the time differences between adjacent sensor
firings:
(9)
This matrix may be factored per the singular value decomposition into a trio of matrices Ak
(ρ), Bk (ρ), Σk (ρ) so that the equation Mk (ρ) = Ak (ρ)Σk (ρ)Bk* (ρ) is satisfied. Letting
, denote the appropriate elements of the matrix Bk (ρ), the estimated velocity is
given by:
(10)
C. The Calibrated Sensor Line
Taking the weighting factor ρ still to be fixed and known, we now treat the calibration factor
c as a variable whose value we may estimate using known velocity data. We are given a set
of training data consisting of sensor firing times and true gait velocities {νk} for a
sample set of walks through the sensor line. We assume the actual walking speeds and the
estimated walking speeds satisfy the linear model:
(11)
where the {ωk}are independent random variables with zero expected value. By collecting all
the measurements into vectors ν,b(ρ), we can estimate the calibration factor
with linear least squares:
(12)
If we knew the actual value ρ we would be done at this point as we could estimate the
calibration factor c given the measured time and velocity values. However, where we do not
know the actual value of ρ, we would like to choose a value which yields the best
Hagler et al. Page 7
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
performance of the method for estimating velocities. In particular we would like to find a
value of ρ which gives the best performance across many sensor-lines. Let us consider the
calibration factor for the nth sensor-line as a function of ρ, that is:
(13)
We may consider the set of velocity estimates also as functions of ρ:
(14)
This expresses the estimated velocity for the kth walk along the nth sensor-line (ρ)
entirely in terms of the measured time, the measured velocity, and the unknown parameter ρ.
The weighting factor may now be found as a value which gives the best sensor-line
performance on average.
In practical situations where calibration data are not available we use an average value of ρ
determined from existing data. In particular, we found that the average value that minimized
the estimation error in our controlled experiments, described in the next section, was ρ =
0.75.
D. The Uncalibrated Sensor Line
If the velocity is not known, then the training data set will contain only the sensor firing
times . In this case we must estimate the value for c using the values for the physical
sensor positions {x̃i}. Choosing any pair of sensors i,j we may make the estimate:
(15)
We would like to choose our pair of sensors so that the expected value x̃j – x̃i is as near in
value as possible to the distance between mean detection positions xj – xi. In general the
expected error will be minimized by choosing the pair consisting of the outermost sensors of
the line.
We do note that care should be taken when using the uncalibrated sensor line cross-
sectionally, especially over small samples, as the individual instantiations of a sensor line
can have sizable differences between ĉ and c as shown in fig. 5.
IV. Experimental Verification
A. Experimental Description
27 subjects (9 male and 18 female, aged 75 to 95 years, mean age 85.2 years, 145cm to
185cm in height, average height 164.8 cm) participated in the experiment; all provided
informed consent. The experiment was conducted in a common use room at the facility
where the participants live. A single sensor line of eight (8) restricted-field PIR motion
Hagler et al. Page 8
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
sensors was placed on the ceiling with sensors physically spaced at 61cm (2ft) intervals. The
ceiling height was 240cm (7.8ft). Beneath the sensor line an 854cm (14ft) long GAITRite®
Walkway System gait mat was placed so that the ends of the mat aligned with the outermost
sensors. Participants were instructed to walk at self-determined “slow”, “normal”, and “fast”
walking speeds. A total of 30 walks were recorded for most participants such that each
participant walked five times at each of the three speeds in the two directions available along
the sensor line. Five participants did a larger number of walks (36, 42, 42, 44, and 46) but
their larger group of walks included the basic 30 which all participants did. Their precise
walking speed for each trial was calculated using the gait mat data. Firing times were
collected for each PIR sensor during each trial and used to determine the accuracy of the
PIR sensors for measuring walking speed.
Based on our experience with several hundred homes, a reasonable sensor line configuration
which may be installed under the space constraints of typical small homes consists of four
sensors placed in a line at approximately 61cm (2ft) intervals. The choice of four sensors
was influenced by a few different factors. First, since the sensors are known to be noisy, we
wanted multiple measurements of the walking speed to use in our estimator to reduce
estimator variance. Experiments showed, for example, that moving from three sensors to
four sensors reduced variance by a factor of approximately 3.8. Second, due to space
constraints in the homes and retirement communities we found that four sensors are all that
would reliably fit in most homes. Third, the probability of an individual sensor firing is
approximately 0.937. With four sensors in place and assuming we use walks where either
three or four sensors fire, we can capture almost 98% of walking events in the home.
Finally, we note that using two sensors in not sufficient as this causes equation (8) to reduce
to a single equation with a single unknown which can be solved exactly, and therefore does
not allow mitigation for known noise effects. In accordance with this we have considered
sensor data in groups of four adjacent sensors. Thus our line of eight sensors is treated as
five individual sensor lines. Furthermore as there is no reason to suppose that the effective
sensor spacing is the same in the two directions along which the line may be walked, – in the
“forward” or “return” direction through the line (with respect to the experimenter) – we
evaluated each direction independently.
For each sensor line of four sensors in each direction only those walking events in which all
4 sensors fired were considered for the purposes of calculating the effective sensor spacing
and calibration factor. However, to estimate the velocity for walking events we used all the
sensor line data in which 3 or 4 of the 4 sensors fired.
B. Experimental Results
A total of 882 walks from the 27 participants (mean age 85.2 years) were recorded during
this experiment with 441 in the “forward” direction and 441 in the “return” direction (as
referenced by the experimenter). The numbers of “slow”, “normal”, and “fast” speed walks
were the same in either direction for each given participant. The 8 ceiling-mounted PIR
sensors were divided into 5 sensor lines of 4 sensors with a regular 61cm (2ft) physical
spacing for analysis. In the “forward” direction the sensor lines had all four sensors fire 350
± 36 times, and in the “return” direction all 4 sensors fired 330 ± 29 times. The effective
sensor spacing for each sensor line was calculated and normalized as in the foregoing using
only the events where all 4 sensors fired.
Participants walked in the “forward” direction with a speed of 104 ± 30.6cm/s, and in the
“return” direction with a speed of 100 ± 29.3cm/s as measured by the gait mat. We
estimated velocity using a sensor line only in those cases where 3 or 4 of the 4 sensors in the
line fired. Fig. 3 shows the directional walking speed estimates versus the measured values
for the combined sensor line data after calibration using velocity data from the gait mat. Fig.
Hagler et al. Page 9
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
4 shows the directional walking speed estimates using estimated calibration factors. In both
figures the “return” direction is differentiated from the “forward direction” by introducing a
negative sign on all the velocity estimates. Additionally, in both figures the line x = y
(corresponding to perfect estimation) has been plotted as a dashed line to demonstrate that
both calibrated and uncalibrated estimates are distributed around the correct values. Further,
the distribution of points in fig.4 is wider than in fig. 3 demonstrating that the calibration
procedure does improve estimation. Also, of particular note is the fact that distribution of
estimates in both figures is more centered and densely packed around the true value in the
“return” direction, indicating that velocity estimates in the “return” direction are better than
in the “forward” direction (i.e., the same sensors performed better in one direction than in
the other).
To be more precise about the discussion of the walking speed estimates, we denote the
estimated speed for the kth walk through sensor line i for both directions by and the actual
speed by νk. The accuracy of the system was evaluated by computing the average difference
. For the case of the calibrated sensor lines the mean of this difference is 1.1cm/s
and the standard deviation is 9.1cm/s. In the case of the uncalibrated sensor lines the mean is
7.1cm/s, and the standard deviation is 11.3cm/s.
Figure 5 shows the relationship of the estimated calibration factors to the true (measured)
calibration factors, with the line x = y drawn for comparison. This shows that the
uncalibrated sensor line with ĉ as in (15) tends to underestimate the true calibration factor,
this making the velocity estimates slightly higher than in the calibrated case. This also
demonstrates the need to be careful when comparing uncalibrated sensor line estimates
cross-sectionally.
V. Discussion
The proposed system for unobtrusive and continuous monitoring of in home walking speeds
has been shown to accurately estimate velocity when compared to the GAITRite® Walkway
System gait mat standard. The mean estimation errors of 7.1cm/s and 1.1cm/s for the
uncalibrated and calibrated sensor lines when compared to the average speed of 102cm/s
result in average errors of 6.96% and 1.08%, respectively. Further, the standard deviations
of the error distributions for the uncalibrated and calibrated sensor lines are 11.1% and
8.92% when compared to the average speed of walking. This shows that each individual
estimate is accurate, and local averaging and other statistical techniques can be used to
increase precision (reduce the error variance further).
These positive results demonstrate the feasibility of the proposed method and address
several deficits in the current paradigm of assessing gait episodically or in clinic settings.
First, with this system the variability of walking speed can now be monitored continuously
over the short term (e.g., walk-to-walk variability) in addition to longer time scales (e.g.,
month-to-month yearly) without expensive and inconvenient clinic visits. Second, subjects
in high-risk groups can be monitored more closely and rapidly than is currently feasible.
Third, researchers can have access to more frequent measurements of walking speed, which
facilitates the refinement and better understanding of walking speed as it relates to health
outcomes and correlations presently in the literature that are based on single or infrequent
measurements. Fourth, wide scale analysis of multiple subjects can be performed relatively
easily which we anticipate will open further areas of population-based research and
diagnostic ability not discussed here. We do note that while our proposed system is less
expensive than repeated clinical visits, the cost of the sensors, computer, internet service,
and transceiver may currently be cost prohibitive for studies that might involve thousands of
Hagler et al. Page 10
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
subjects who are widely dispersed. However, since historically the cost of equipment and
services has continued to drop as better and faster technology becomes available in the
marketplace, it is likely that deployment of these kinds of systems to larger cohorts will be
facilitated. In addition, less expensive motion sensors may work adequately and simple
application specific computers may be built cheaper than off the shelf models which could
be deployed today. Further work is needed to identify the most cost efficient approaches to
maximize scalability of in-home assessment platforms.
One of the largest challenges to the broad use of our approach to continuous monitoring of
walking speed, and to in-home monitoring in general, is the differentiation of multiple
residents. This problem is typically addressed by requiring the participants to wear or carry
some type of radio frequency identification (RFID) tag. We are currently working on both
pattern recognition and model-based approaches to distinguish between multiple residents
based on the walking events. This will allow the expansion of this methodology from single
resident homes to multiple resident dwellings without the need for additional equipment or
hardware.
Future work will address comparisons of the in-home continuous method with standard
clinical tests of walking, mobility, and physical performance such as the Short Physical
Performance Battery, Unified Parkinson's Disease Rating Scale, and various other timed
walks of different durations (e.g., 4-meter, 10-meter, 400 meter) thus facilitating
interpretation of these established clinical metrics with our new framework. Other future
work will include relaxing the assumption that velocity is fixed over a walking event in
order to measure the step-to-step variability in each walking event. In this case the velocity
of the kth walking event becomes some function of time ν(t). Retaining our definition of the
error {εi} above we find that the time at which the sensor fires may be expressed as {τi + υi},
where {υi} satisfies:
(16)
The values {ηi} are still defined as above which gives a relation to the measured time values
of . Finally, for any pair of sensors we have the relation:
(17)
which generalizes (1). We anticipate that adjusting the model along the lines of (16) will
allow us to derive additional gait parameters from the current and future data.
VI. Conclusion
In this paper we have proposed a new system for continuous in home assessment of walking
speed based on PIR sensors and a wireless network for data collection. We have shown that
this method is both accurate and precise when compared to the standard of the GAITRite®
Walkway System gait mat. This method allows the convenient in home collection of a large
number of walking events otherwise gathered infrequently in a clinical setting. Since
walking speed has been shown to be an indicator or predictor of many diseases and other
Hagler et al. Page 11
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
health issues such as cognitive decline and hospitalization, we feel that the continuous
monitoring of this measure and its applications is an important and useful area of future
research.
Acknowledgments
The authors would like to thank the volunteer subjects who participated in this research and the staff from the
Oregon Center for Aging & Technology who assisted in this study.
This work was supported in part by the National Institute of Health and the National Institute on Aging under
Grants R01AG024059, P30 AG024978, and P30AG008017. This work was also partially funded by Intel
Corporation.
References
1. Kaye, JA.; Zitzelberger, TA. Overview of Healthcare, Disease, and Disability. In: Bardram, JE.;
Mihailidis, A.; Wan, D., editors. Pervasive Computing in Healthcare. CRC Press; 2006.
2. Dishman E. Inventing wellness systems for aging in place. Computer 2004;37:4–34.
3. Varshney U. Pervasive Healthcare. Computer 2003;36:138–140.
4. Hayes TL, Abendroth F, Adami A, Pavel M, Zitzelberger TA, Kaye JA. Unobtrusive assessment of
activity patterns associated with mild cognitive impairment. Alzheimers Dement 2008;4:395–405.
[PubMed: 19012864]
5. Logan, B.; Healey, J.; Philipose, M.; Tapia, EM.; Intille, S. A long-term evaluation of sensing
modalities for activity recognition. Lnnsbruck, Austria: 2007.
6. Banerjee S, Steenkeste F, Couturier P, Debray M, Franco A. Telesurveillance of elderly patients by
use of passive infra-red sensors in a ‘smart’ room. J Telemed Telecare 2003;9:23–9. [PubMed:
12641889]
7. Sixsmith A, Johnson N. A smart sensor to detect the falls of the elderly. IEEE Pervasive Computing
2004;3:42–47.
8. Philipose M, Fishkin KP, Perkowitz M, Patterson DJ, Fox D, Kautz H, Hahnel D. Inferring activities
from interactions with objects. IEEE Pervasive Computing 2004;3:50–57.
9. Jimison H, Pavel M, McKanna J, Pavel J. Unobtrusive monitoring of computer interactions to detect
cognitive status in elders. IEEE Transactions on Information Technology in Biomedicine
2004;8:248–252. [PubMed: 15484429]
10. Hayes TL, Pavel M, Larimer N, Tsay IA, Nutt J, Adami AG. Distributed healthcare: Simultaneous
assessment of multiple individuals. IEEE Pervasive Computing 2007;6:36–43.
11. Perry M, Dowdall A, Lines L, Hone K. Multimodal and ubiquitous computing systems: supporting
independent-living older users. IEEE Trans Inf Technol Biomed 2004;8:258–70. [PubMed:
15484431]
12. Mihailidis A, Carmichael B, Boger J. The use of computer vision in an intelligent environment to
support aging-in-place, safety, and independence in the home. IEEE Trans Inf Technol Biomed
2004;8:238–47. [PubMed: 15484428]
13. Lymberopoulos D, Teixeira T, Savvides A. Macroscopic human behavior interpretation using
distributed imager and other sensors. Proceedings of the IEEE 2008;96:1657–1677.
14. Stanford V. Using pervasive computing to deliver elder care. IEEE Pervasive Computing
2002;1:10–13.
15. Orwat C, Graefe A, Faulwasser T. Towards pervasive computing in health care - a literature
review. BMC Med Inform Decis Mak 2008;8:26. [PubMed: 18565221]
16. Studenski S, Perera S, Wallace D, Chandler JM, Duncan PW, Rooney E, Fox M, Guralnik JM.
Physical performance measures in the clinical setting. J Am Geriatr Soc 2003;51:314–22.
[PubMed: 12588574]
17. Bramell-Risberg E, Jarnlo GB, Minthon L, Elmstahl S. Lower gait speed in older women with
dementia compared with controls. Dement Geriatr Cogn Disord 2005;20:298–305. [PubMed:
16166777]
Hagler et al. Page 12
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
18. Camicioli R, Howieson D, Oken B, Sexton G, Kaye J. Motor slowing precedes cognitive
impairment in the oldest old. Neurology 1998;50:1496–8. [PubMed: 9596020]
19. Beauchet O, Allali G, Berrut G, Hommet C, Dubost V, Assal F. Gait analysis in demented
subjects: Interests and perspectives. Neuropsychiatr Dis Treat 2008;4:155–60. [PubMed:
18728766]
20. Guralnik JM, Ferrucci L, Pieper CF, Leveille SG, Markides KS, Ostir GV, Studenski S, Berkman
LF, Wallace RB. Lower extremity function and subsequent disability: consistency across studies,
predictive models, and value of gait speed alone compared with the short physical performance
battery. J Gerontol A Biol Sci Med Sci 2000;55:M221–31. [PubMed: 10811152]
21. Guralnik JM, Ferrucci L, Simonsick EM, Salive ME, Wallace RB. Lower-extremity function in
persons over the age of 70 years as a predictor of subsequent disability. N Engl J Med
1995;332:556–61. [PubMed: 7838189]
22. Holtzer R, Verghese J, Xue X, Lipton RB. Cognitive processes related to gait velocity: results from
the Einstein Aging Study. Neuropsychology 2006;20:215–23. [PubMed: 16594782]
23. Fitzpatrick AL, Buchanan CK, Nahin RL, Dekosky ST, Atkinson HH, Carlson MC, Williamson
JD. Associations of gait speed and other measures of physical function with cognition in a healthy
cohort of elderly persons. J Gerontol A Biol Sci Med Sci 2007;62:1244–51. [PubMed: 18000144]
24. Kaye J. Home-based technologies: a new paradigm for conducting dementia prevention trials.
Alzheimers Dement 2008;4:S60–6. [PubMed: 18632003]
25. Aminian K, Robert P, Jequier E, Schutz Y. Incline, speed, and distance assessment during
unconstrained walking. Med Sci Sports Exerc 1995;27:226–34. [PubMed: 7723646]
26. Miyazaki S. Long-term unrestrained measurement of stride length and walking velocity utilizing a
piezoelectric gyroscope. IEEE Trans Biomed Eng 1997;44:753–9. [PubMed: 9254988]
27. Sabatini AM, Martelloni C, Scapellato S, Cavallo F. Assessment of walking features from foot
inertial sensing. IEEE Trans Biomed Eng 2005;52:486–94. [PubMed: 15759579]
28. Hayes, TL.; Hagler, S.; Austin, D.; Kaye, J.; Pavel, M. Unobtrusive assessment of walking speed
in the home using inexpensive PIR sensors. presented at 31st Annual International Conference of
the IEEE Engineering in Medicine and Biology Society; Minneapolis, MN. 2009;
29. Bjorck, A. Numerical Methods for Least Squares Problems. SIAM; 1996.
30. Golub, GH.; Loan, CFV. Matrix Computations. third. The John Hopkins University Press; 1996.
Biographies
Stuart Hagler received the degrees of B.A. degree in philosophy, B.S. degree in physics,
and M.S. degree in physics from the University of Washington, Seattle, Washington, USA.
He is currently working as a Biomedical Researcher at the Oregon Health and Science
University, Portland, OR, USA.
Daniel Austin (M'05) received the B.S. degree in electronic engineering technology from
the Oregon Institute of Technology, Portland, OR, USA in 2006, the M.S. degree in
electrical engineering from the University of Southern California, Los Angeles, CA, USA in
2008, and is currently working toward the Ph.D. degree in biomedical engineering at the
Oregon Health and Science University, Portland, OR, USA.
Hagler et al. Page 13
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
He is currently working as a Biomedical Researcher for the Point of Care Laboratory at the
Oregon Health and Science University, Portland, OR. His research interests include signal
processing and modeling with applications to biosignal analysis and early detection of
physical and cognitive decline.
Tamara L. Hayes (M '01) received the M.S. in electrical engineering from the University of
Toronto, Toronto, Canada and the Ph.D. degree in behavioral neuroscience from the
University of Pittsburgh, Pittsburgh, PA, USA.
She is an assistant professor in the Division of Biomedical Engineering at Oregon Health
and Science University, Portland, OR. Her research interests include using ubiquitous
computing to deliver healthcare in the home, with the goal of changing the current paradigm
of clinic-centered healthcare to a less costly, more effective model, which lets individuals
participate more fully in their health care.
Dr. Hayes is also a member of the Sleep Research Society.
Hagler et al. Page 14
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
Jeffrey Kaye (M '05) received his M.D. degree at New York Medical College, New York,
NY, USA in 1980 and trained in neurology at Boston University, Boston, MA USA from
1981-1984.
He was a fellow in movement disorders at Boston University and then a Medical Staff
Fellow in brain aging at the National Institute on Aging, National Institutes of Health from
1984 to 1988. He is currently Professor of Neurology and Biomedical Engineering, director
of the Oregon Center for Aging & Technology (ORCATECH), and director of the Layton
Aging & Alzheimer's Disease Center, at the Oregon Health and Science University,
Portland, OR, USA. His research interests include the design, study and application of
Hagler et al. Page 15
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
ubiquitous unobtrusive technologies and systems for home-based health research and the
identification of behavioral and biological biomarkers of healthy aging.
Dr. Kaye chairs the Working Group on Technology of the US National Alzheimer's
Association and is a Commissioner for the Center for Aging Services and Technology,
Washington, DC, USA.
Misha Pavel (M'69) received the B.S. in electrical engineering from Polytechnic Institute of
Brooklyn, New York, NY, USA, the M.S. degree in electrical engineering from Stanford
Hagler et al. Page 16
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
University, Stanford, CA, USA, and the Ph.D. degree in experimental/mathematical
psychology at New York University, New York, NY, USA.
He is a Professor and Head of the Division of Biomedical Engineering with a joint
appointment in Medical Informatics and Biomedical Computer Science at the Oregon Health
and Science University, Portland, OR, USA. He is also the director of the Point of Care
Laboratory which focuses on unobtrusive monitoring and neurobehavioral assessment, and
computational modeling with applications to care for the aging population. Previously he
was a Technology Leader at AT&T Laboratories developing networked, wireless, and
mobile applications for information access and context aware interactions. Other past
positions include Faculty at New York University and Stanford where he worked on sensor
fusion, modeling of pattern recognition in sensory motor systems and human computer
communications systems and Member of Technical Staff at Bell Labs where he was
developing new approaches to network analysis and modeling by incorporating
characteristics of human behavior.
Hagler et al. Page 17
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
Fig. 1.
A motion sensor line for measuring walking speed where the four sensors are placed 0.61 m
apart and are installed on a ceiling typically 2.54 m high.
Hagler et al. Page 18
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
Fig. 2.
Schematic of a person walking through a sensor line containing four sensors with the fields
of view and the locations of the x̃i and xi shown.
Hagler et al. Page 19
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
Fig. 3.
Combined walking speed data for all subjects for the 5 sensor lines of 4 sensors (the various
shapes indicate data from different sensor lines), using a calibration factor calibrated to the
walking speed measured by the gait mat. The sign of the estimated speed differentiates
“forward” walks (positive values) versus “return” walks (negative values).
Hagler et al. Page 20
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
Fig. 4.
Combined walking speed data for all subjects for the 5 sensor lines of 4 sensors (the various
shapes indicate data from different sensor lines), using estimated calibration factors. The
sign of the estimated speed differentiates “forward” walks (positive values) versus “return”
walks (negative values).
Hagler et al. Page 21
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
Fig. 5.
Combined calibration factor data (c value) for the 10 possible sensor lines (5 “forward” and
5 “return”). The direction the triangle is pointing indicates whether the direction is
“forward” (up) or “return” (down).
Hagler et al. Page 22
IEEE Trans Biomed Eng. Author manuscript; available in PMC 2011 April 1.
NIH-PA Author Manuscript NIH-PA Author Manuscript NIH-PA Author Manuscript
