Content uploaded by Misha Pavel
Author content
All content in this area was uploaded by Misha Pavel
Content may be subject to copyright.
Abstract— The ability to assess the neurological state of
patients with neurodegenerative diseases on a continuous basis
is an important component of future care for these chronically
ill patients. In this paper we describe a set of algorithms to
infer gait velocity and its variability using data from an
unobtrusive sensor network by incorporating a simple dynamic
description of a patient’s movements within his or her
residence. The sensors include a combination of passive motion
detectors and active radio frequency identification tags. The
dynamic model is a simple 4 state hidden Markov model. We
investigated the ability of this model to assess gait velocity and
its variability using data from a six month pilot study of several
patients with early stage Parkinson’s disease.
I. INTRODUCTION
HE ability to assess the sensory, motor and cognitive
functionality of an individual is an important problem in
caring for individuals with chronic diseases, as well as for
the healthy elders. Current methods relying on observations,
e.g., Unified Parkinson’s Disease Rating Scale (UPDRS),
Clinical Dementia Rating (CDR) as well as those relying on
neuropsychological testing are inherently highly variable,
costly, and to some extent depend on the testers’ capabilities
and training. As such, they are administered infrequently,
require multiple administrations to assess long-term change,
and are generally difficult to use to assess the instantaneous
state of the patient.
Yet, in many situations it would be desirable to assess the
functions related to the neurological state of the patient on a
continuous basis. For example, in the care of patients with
Parkinson’s disease (PD), the instantaneous aspects of gait
velocity may be a useful measure for the determination of
the administration of drugs, such as Levodopa. Also,
continuous measurement of motor function would be a large
improvement over current techniques of estimating impact
This work was supported in part by grants from NIH, NSF, and the
Kinetics Foundations to Oregon Health and Science University.
M. Pavel is in Biomedical Engineering and Medical Informatics at
Oregon Health and Science University (OHSU) Tamara Hayes is Assistant
Professor in Biomedical Engineering at OHSU,
Deniz Erdogmus is Assistant Professor in Biomedical Engineering and
Computer Science and Electrical Engineering at OHSU
Ann Tsay is Research Associate in Biomedical Engineering at OHSU,
Nicole Larimer Research Assistant in Biomedical Engineering at OHSU,
Holly Jimison is Associate Professor in Medical Informatics and
Biomedical Engineering at OHSU
John Nutt is a Professor in Neurology at OHSU.
of symptomatic therapies on motor function throughout the
day using diaries filled out by study subjects. Further,
continuous longitudinal measurements would be mare
accurate way to estimate the effects of treatments that are
hypothesized to alter the progression of a motor disorder.
Our approach to the problem of assessment is based on a
combination of neural informatics and sensor technology.
Neural Informatics is a collection of computer-based
methods that address issues at the intersection of neural
engineering, neuroscience and clinical practice. The
methods include modeling clinically relevant aspects of
neurological states of both individuals and populations. In
our preliminary studies as well as those from other
laboratories, the variability of the various metrics appears to
be as important as the average values. For example,
variability in mobility measures, such as gait velocity or
stride length, have been shown to correlate with age, and
with dementias, such as those associated with Alzheimer’s
Disease [1]. Motor pattern generating mechanisms and gait
velocity appear to be useful predictors of future cognitive
decline [2, 3]. The goal of our approach is to replace the
occasional sampling of cognitive and motor function using
clinic-based control tests with continuous observations in
the patients’ natural environments. A system that can assess
aspects of mobility on a continuous basis can, of course, be
used to assess the variability of these measures.
In this paper, we describe an approach to the unobtrusive
measurement of gait velocity and its variability based on
continuous monitoring of the PD patients in their homes. In
particular we extend our prior work [6] by incorporating a
simple dynamic model of the movements of the individuals
in their dwellings.
II. UNOBTRUSIVE MEASUREMENT OF GAIT VELOCITY
The system for the unobtrusive measurement of gait
comprises two components that will be described in the
following two sections: A) a sensor system for sensing and
collecting the raw data, and B) a set of algorithms that
estimate the parameters of interests from the raw data.
A. Sensor System
One of the most important requirements of the sensor and
assessment system is its unobtrusive or minimally intrusive
nature [6] as well as its economical feasibility. In order to
develop a system that is minimally intrusive, we have been
Continuous Assessment of Gait Velocity in Parkinson’s Disease from
Unobtrusive Measurements
Misha Pavel, Member, IEEE, Tamara Hayes, Member, IEEE, Ishan Tsay,
Deniz Erdogmus, Member, IEEE, Anindya Paul, Member, IEEE,
Nicole Larimer, Holly Jimison, Member, IEEE, John Nutt
T
Proceedings of the 3rd International
IEEE EMBS Conference on Neural Engineering
Kohala Coast, Hawaii, USA, May 2-5, 2007
SaE1.6
1-4244-0792-3/07/$20.00©2007 IEEE. 700
investigating approaches that would be affordable by a large
number of patients and their families. For example, we have
been investigating systems based on passive infrared sensors
(PIR) and contact switches that would be deployed in a
similar manner as are the components used in many security
systems.
The general architecture installed in a typical home is
shown in Fig. 1. When a sensor detects the presence of a
moving human body at the normal body temperature and the
motion signal exceeds a fixed threshold, it fires. There are
various details that control the firing, such as the refractory
period of the sensor following a firing, but a discussion of
these is beyond the scope of this paper. In general, the
inference of the gait velocity is based on the time that it
takes to traverse from one part of the patient’s home to
another – this approach, in conjunction with semi-Markov
models was described previously in an article by Pavel et al.
in 2006 [7].
In some dwellings it is possible to make the
measurements of speed of walking more directly by taking
advantage of the layout. In particular, if the residence has a
hall or a corridor, we would place three modified PIR
detectors the hall in a row as shown in Fig. 1. The PIR
motion detectors placed in the “test area”, i.e., the hall, were
modified to restrict their field of view to 4o. The PIR
monitoring system is using a simple wireless communication
network in order to collect the data. The main data relevant
to monitoring mobility consist of records of PIR motion
detectors firing events.
In dwellings with multiple residents, it is necessary to
identify the particular individual associated with different
events recorded from the PIR motion detectors. In this
study, we used a commercially available RFID location
tracking system for this purpose (HomeFree, Inc). Each
individual residing in the same residence wore a small RFID
device in the form of a watch that emits periodically – every
4 seconds – a signal received by three or more base stations.
B. Inference Algorithms
The unobtrusive nature of the sensor system is inherently
plagued with considerable uncertainties arising from the fact
that the measured phenomenon, e.g. gait velocity is only
indirectly related to the sensed data. In order to compensate
for this, we base our inference on a fusion of information
from a set of passive infrared detectors, contact switches and
active radio frequency identification (RFID) system.
As noted above, the incorporation of the RFID system is
essential for inference in dwellings with multiple residents.
Such was the case in a recent pilot study of a small number
of patients with Parkinson’s disease and their spouses that
served as control subjects.
The measurement of speed of walking in the test area is a
trivial problem whenever an observation consists of the
three detectors firing in one of two temporal orders
consistent with a particular direction of the movement. In
those cases, the time between the first and the third detector
events is taken as the time to walk the distance between
detectors. The difficult problem is to infer the identity of the
walking individual.
Our original notion was that the received signal strength
indicator (RSSI) would provide sufficient information to
determine the locations of each individual. However, this
was not possible due to the variability of the RSSI signal
and its lack of monotonicity with the distance. In our prior
work [6], this problem was addressed by static modeling of
the RSSI distribution over the dwelling, and the use of the
expectation-maximization (EM) algorithm to best estimate
the individual associated with the motion detector events
[9,10]. Despite its fairly successful deployment, further
improvements could be achieved by adding constraints due
to the sequential dependencies – dynamics – of the moving
individuals.
The general approach is to determine the likelihood of
each individual resident being in the test area – a similar
approach to [6] – combined with a dynamic model based on
the assumption that the movements of an individual could be
described by a simple hidden Markov model (HMM) shown
in Fig. 2.
In particular, the HMM developed for the inference was
assumed to consist of 4 unobservable states described by
Fig. 1. Example of a part of a residence with a
number of PIR sensors, RFID receiver stations and
the test area.
Else
where
Test
Area
Left Right
Else
where
Test
Area
Left Right
Fig. 2 Hidden Markov model representing the
dynamics of the patient in his residence.
701
locations in relation to the test area, namely:
1. Left of the test area
2. Right of test area
3. At the test area
4. Other – Elsewhere
The transition probabilities jk
p
between each pair of states
are constrained by the direction of the patient’s movement as
sensed by the PIR motion detector sequences – left to right
and vice versa and the general topology of the HMM is
shown in Fig. 2. The transition probabilities were
determined using the occupancy of the different parts of the
dwelling assessed by the cumulative number of motion
detector events in each part of the house. Table 1 lists an
example of the transition probabilities used in conjunction
with the house diagram shown in Fig. 1.
TABLE 1
EXAMPLE OF TRANSITION PROBABILITIES
Left Test Right Other
Left 0.20 0.60 0.00 0.20
Test 0.30 0.20 0.50 0.00
Right 0.00 0.33 0.33 0.33
Other 0.25 0.00 0.25 0.50
The notation to describe the inference algorithm is similar
to that used in [6]. The observable outputs, R in each state
are vectors of the RSSI values recorded by each receiver or
base station in the residence. The probability density of the
observable RSSI are given by
()
|
i
f
Rq, where q is the state
in the HMM and i is the individual. These probability
density functions were estimated using a Gaussian mixture
model (GMM) with parameters estimated using the
expectation-maximization approach, from calibration data
gathered during the initial installation of the sensor system.
The calibration RSSI data were obtained by collecting RSSI
data from each “significant” location in the patient’s
dwelling, as identified by the patient or the spouse. For the
purpose of this study, each resident is associated with a
particular HMM and these individual HMMs are treated as
statistically independent.
III. RESULTS
The data collected in the pilot study from 6 houses were
used to examine the applicability of this approach. In this
presentation we illustrate the approach on the data from one
Fig. 3. Example of results from one house with a Parkinson’s patient (lower graph) and a control subject (upper
graph). The abscissa is the date of each observation and the ordinate is the time it takes to walk along the test area.
The three different sets of data correspond to the 20, 50, and 80 percentiles of walking times. The missing data
indicate that the corresponding individual was not in the residence. The data are computed using 7 day moving
id
702
of the houses. The data from the patients’ dwellings were
processed whereby the events form the motion detectors in
the test area were used to identify all the instances when the
three detectors fired in one of the temporal orders
corresponding to rightward or leftward motion (ignoring
partial sequences). For each time t associated with the first
detector event, the sequence of RSSI samples was used
()()()
() ( ) ( ) ( )
{3, 2, ,
,,2,3}
Rt Rt Rt
Rt Rt Rt Rt
=−Δ −Δ −Δ
+Δ + Δ + Δ
R
to determine the most likely sequence of states consistent
with the observed RSSI. The likelihood of any particular
sequence of states
{}
12 7
,,... ,
iqq q=Q for the i-th individual
is given by
() { }
()
()
()
1
6
31 ,1 1
3
1
Pr |
||,
ii
qi t kk i k
tk
k
fR q p fR q
λ
π
−+ =
+Δ −
=
=
=∏
RRQ
where q
π
is the prior probability of the state 1
qand ,1kk
p
+
is
the transition probability from state k to state k+1. Using
Viterbi search, the inference algorithm found, for each
individual, the most likely sequence of states *
i
Q given the
RSSI observations. Since the identity of the walking
individual is not known with certainty, the overall
distribution )(tg of the observed times to walk is a mixture
of the distributions associated with each individual,
() ()
1
,
N
ii
i
g
tgt
α
=
=∑
where i
α
is the probability that the observation is associated
with the i-th individual. Conversely, given the probability
that a particular measurement is associated with the i-th
individual, it is possible to estimate the expected value and
the variance of the time to walk distribution for that
individual using the probabilities derived from the HMM. In
particular, using a Bayesian estimation procedure with
uniform priors, the estimate of the expected value for the
time to walk for the i-th individual is given by:
[]
()
,iitit
t
ET w T=∑R
Where ,it
T is the walking time measurement, wi is a weight
determined as follows: wi is zero if *
i
Qdoes not contain test
q
and
()
()
it
i
j
t
j
w
λ
λ
=∑
R
R
for all sequences that contain the state test
q, corresponding
to the test area. In other words, the weight is determined by
the relative likelihood of each individual walking in the test
area.
An example of the results from one house occupied by a
Parkinson’s patient and his spouse is shown in Fig. 3. The
distribution of walking times for the i-th individual was
computed using the probability estimates that a given
measurement came from the i-th individual. The triangles
correspond to the median estimates. The data indicate higher
variability of the walking times of the Parkinson’s patient..
IV. CONCLUSION
We have developed a technique to reduce the uncertainty
associated with unobtrusive measurement of mobility with
multiple individuals in a single dwelling. We have
demonstrated that a simple 4 state HMM with minimal
training can incorporate the dynamic aspects of the
sequential data. Although the application of this approach to
the data from our pilot study appears to be promising, the
real evaluation of this approach would require a data set
with known ground truth. Alternatively, there are additional
enhancements of this approach that would improve the
inference; for example, including a coupling between the
HMMs corresponding to each individual would impose
additional constraints and probably improve the inference.
ACKNOWLEDGMENT
We thank Jeff Kaye for his leadership in the development
of related approaches in the monitoring of elders and John
Hunt for his engineering and technical expertise and advice.
REFERENCES
[1] P. L. Sheridan, J. Solomont, N. Kowall, and J. M. Hausdorff,
"Influence of executive function on locomotor function: divided
attention increases gait variability in Alzheimer's disease," J Am
Geriatr Soc, vol. 51, pp. 1633-7, 2003.
[2] R. Camicioli, D. Howieson, B. Oken, G. Sexton, and J. Kaye, "Motor
slowing precedes cognitive impairment in the oldest old," Neurology,
vol. 50, pp. 1496-8, 1998.
[3] R. Camicioli, M. M. Moore, G. Sexton, D. B. Howieson, and J. A.
Kaye, "Age-related brain changes associated with motor function in
healthy older people," J Am Geriatr Soc, vol. 47, pp. 330-4, 1999
[4] T. L. Hayes, M. Pavel, and J. A. Kaye, "An unobtrusive in-home
monitoring system for detection of key motor changes preceding
cognitive decline," presented at 26th Annual International Conference
of the IEEE Engineering in Medicine and Biology Society, San
Francisco, CA, 2004.
[5] T. Hayes, M. Pavel, P. K. Schallau, and A. M. Adami, "Unobtrusive
monitoring of health status in an aging population.," presented at 5th
International Conference on Ubiquitous Computing, Seattle,
Washington., 2003.
[6] T.L. Hayes, M. Pavel, A. Adami, N. Larimer, A. Tsay and J. Nut,
“Distributed healthcare: Simultaneous assessment of multiple
individuals, Pervasive Computing”, 6 (1), 36-43, 2007.
[7] M. Pavel, T. L. Hayes, A. Adami, H. B. Jimison, and J. Kaye,
"Unobtrusive assessment of mobility," presented at 28th Annual
International Conference of the IEEE Engineering in Medicine and
Biology Society, New York, NY, 2006.
[8] R. B. Postuma, , T. Arenovich, A. E. Lang, Does severity of
Parkinson's disease vary according to season? Movement Disorders,
Volume 20, Issue 4 ,2004,pp 476 – 479
[9] A. Dempster, N. Laird, and D. Rubin, "Maximum likelihood from
incomplete data via the EM algorithm," Journal of the Royal
Statistical Society, vol. 39, pp. 1-39, 1977.
[10] M. Jordan and R. Jacobs, "Hierarchical mixtures of experts and EM
algorithm," Neural Computation, vol. 5, pp. 181-214, 1994.
703