A body sensor network to monitor Parkinsonian symptoms: extracting features on the nodes

Abstract

The work presented in this paper concerns the development of a body sensor network to monitor changes in the severity of symptoms in patients with Parkinson's disease. We analyzed the impact of different features derived from wearable sensor (i.e. accelerometer) data on the reliability of the prediction of clinical scores that capture the severity of Parkinsonian symptoms. We implemented an off-line feature estimation procedure on the nodes to assess the computational complexity associated with estimating each feature. Our analyses revealed that good prediction performance can be achieved by extracting a subset of features that are not computationally demanding, thus making it possible to envision estimating features and saving and/or transmitting them to a base station (e.g. PDA) rather than raw accelerometer data. This strategy would avoid numerous problems with transmission of data recorded at relatively high sampling rate including power consumption and the need for dealing with transmission errors. These findings suggest that a wireless wearable system could be a viable tool for long term monitoring of patients with Parkinson's disease.

Full text

Page 1

A Body Sensor Network to Monitor Parkinsonian
Symptoms: Extracting Features on the Nodes
Shyamal Patel1,2, Konrad Lorincz3, Richard Hughes1, Nancy Huggins4, John Growdon4,
David Standaert5, Jennifer Dy2, Matt Welsh3, Paolo Bonato1
1 Department of Physical Medicine and Rehabilitation, Harvard Medical School, 125 Nashua Street, Boston, MA, USA
spatel19@partners.org
rhughes1@partners.org
2 Department of Electrical and Computer Engineering, Northeastern University, 360 Huntington Avenue, Boston, MA, USA
jdy@ece.neu.edu
3 Computer Science, Harvard University, 33 Oxford St., Cambridge, MA, USA
konrad@eecs.harvard.edu
mdw@eecs.harvard.edu
4 Department of Neurology, Harvard Medical School, 55 Fruit Street, Boston, MA, USA
nhuggins@partners.org
growdon@helix.mgh.harvard.edu
5 Department of Neurology, University of Alabama at Birmingham, 1720 7th Ave S, Birmingham, AL, USA
dstand@uab.edu

Abstract— The work presented in this paper concerns the
development of a body sensor network to monitor changes in the
severity of symptoms in patients with Parkinson’s disease. We
analyzed the impact of different features derived from wearable
sensor (i.e. accelerometer) data on the reliability of the prediction
of clinical scores that capture the severity of Parkinsonian
symptoms. We implemented an off-line feature estimation
procedure on the nodes to assess the computational complexity
associated with estimating each feature. Our analyses revealed
that good prediction performance can be achieved by extracting
a subset of features that are not computationally demanding,
thus making it possible to envision estimating features and saving
and/or transmitting them to a base station (e.g. PDA) rather than
raw accelerometer data. This strategy would avoid numerous
problems with transmission of data recorded at relatively high
sampling rate including power consumption and the need for
dealing with transmission errors. These findings suggest that a
wireless wearable system could be a viable tool for long term
monitoring of patients with Parkinson’s disease.
I. INTRODUCTION
Parkinson’s disease (PD) is the most common cause of
movement disorder, affecting about 3% of the population over
the age of 65 years and more than 500,000 US residents. The
characteristic motor features of PD include tremor,
bradykinesia (i.e. slowness of movement), rigidity (i.e.
resistance to externally imposed movements), and impaired
postural balance. Current therapy for Parkinson’s disease is
based primarily on augmentation or replacement of dopamine,
using the biosynthetic precursor levodopa or other drugs that
activate dopamine receptors. These therapies are often
successful for some time in alleviating the abnormal
movements, but most patients eventually develop motor
complications as a result of these treatments [1][2]. These
complications include wearing-off, the abrupt loss of efficacy
at the end of each dosing interval, and dyskinesia, i.e.
involuntary and sometimes violent writhing movements [3][4].
We are working towards the application of wearable
sensors to monitor motor fluctuation cycles in patients with
PD. Our goal is to develop a wireless wearable sensor system
that PD patients can wear in the home environment over a
period of several days to weeks. Two key points toward
developing the tools necessary to achieve continuous
monitoring of motor function are (1) development of a robust
and deployable wearable wireless network of sensors and
(2) the development of analysis techniques to derive clinically
relevant information from miniature sensor data.
We are working to develop a system that can efficiently
process data on the wireless nodes without the need for
transmitting raw data for off-line processing. An advantage of
doing so is the reduction in the amount of storage space
required on the nodes and power saving, which would enable
longer recordings. Also when monitoring patients for up to
one week, it is desirable to assess the quality of the recordings
beyond a few spot checks. This objective can be achieved by
estimating data features on the nodes with available
computational resources, and by wirelessly transmitting only
features, as opposed to raw data. Building this capability into
a body sensor network would allow clinical personnel to
check that data captured during the monitoring interval are
satisfactory, and carry the information they need. One
challenge to implementing this strategy is the limited
availability of computational resources on the nodes of the
body sensor network. This necessitates a trade-off between the
relevance of the clinical information captured by a given data
feature and the computational cost associated with its
estimation. In this paper, we present results obtained from
simulations to determine the amount of computational time

Page 2

required to extract features from raw sensor data. We also
estimate the importance of different features in predicting
clinical information. We aim to determine a set of features
that are both computationally inexpensive and lead to accurate
prediction of clinical information.
II. WIRELESS WEARABLE SENSOR SYSTEM
The sensor platform we are using is the Intel Digital Health
Group's Sensing Health with Intelligence, Modularity,
Mobility, and Experimental Reusability (SHIMMER).
SHIMMER consists of a TI MSP430 microprocessor; a
Chipcon CC2420 IEEE 802.15.4 2.4 GHz radio; a MicroSD
card slot; a triaxial MEMS accelerometer, the Freescale
MMA7260Q; and optionally, a Bluetooth radio which allows
streaming of sensor data at high rates, as a radio-agnostic
solution.
SHIMMER (Figure 1) is smaller than other wireless
systems, with conventional board technology and a lithium-
polymer battery for easy maintenance. Internal and external
connectors allow new sensor boards to be interfaced to the
device, expanding its capabilities. A triaxial gyroscope board
using two InvenSense IDG-300 dual-axis gyroscope chips was
designed for internal expansion. One of these gyroscopes is
mounted perpendicularly to the main sensor board.
The SHIMMER device combines computation, radio
communication, high-fidelity triaxial sensors, and a large flash
memory into a tiny, wearable rugged plastic enclosure. It
measures 1.75” x 0.8” x 0.5” and weighs just 10 g.
SHIMMER utilizes a MicroSD card slot that allows up to
2 GBytes of flash memory. This permits continuous data
recording of uncompressed, 50 Hz sampled data from 3
channels for more than 80 days. This amount of on-board
storage is unprecedented in wearable sensor design.


Figure 1. SHIMMER wearable sensor platform.





III. METHODS
A. Data Collection
Twelve individuals were recruited in the study, ranging in
age from 46 to 75 years, with a diagnosis of idiopathic
Parkinson’s disease (Hoehn & Yahr stage 2.5 to 3, i.e. mild to
moderate bilateral disease with ability to recover from sudden
postural disturbance or with some postural instability).
Subjects were assessed on the day of the experiment by a
clinician using the Unified Parkinson’s Disease Rating Scale
(UPDRS) [5].


Figure 2. PD Motor fluctuation cycle.

Figure 2 schematically represents a motor fluctuation cycle.
Subjects were asked to delay their first medication intake in
the morning so that they could be tested in a “practically-
defined OFF” state. This approach is clinically used to
observe patients in the period of most severe Parkinsonian
symptoms. Subsequently, patients took their medications and
were tested at 30-minute intervals thereafter until the
following medication intake, in order to gather sensor data
during an entire motor fluctuation cycle.
Accelerometer sensors were used to gather biomechanical
signals during standardized motor tasks utilized for clinical
assessment including sitting, finger-to-nose movements,
finger tapping, alternating hand movements, leg agility, sit-to-
stand, walking, and stand-to-sit. For each task, 30 s of sensor
data were recorded. Accelerometers were placed on the right
and left upper arm, right and left forearm, right and left thigh,
right shin, and left shin. The sensors were connected to an
ambulatory system (Vitaport 3, Temec BV, The Netherlands)
equipped with data acquisition hardware and software to
collect and store the signals. Subjects were videotaped
throughout the experiment, and motor UPDRS and dyskinesia
scores were assigned for each task by review of the videotapes.
B. Feature Extraction
Raw data were high-pass filtered with a cutoff frequency of
1 Hz to remove gross orientation changes, and low-pass
filtered with a cutoff frequency of 15 Hz to remove high
frequency noise. Sets of thirty 5-second epochs were
randomly selected from sensor data corresponding to each
task, for each testing interval. Six different types of features
were extracted from different body segments from each epoch.
The features were chosen to represent characteristics such as
intensity, modulation, rate, periodicity, and coordination of
movement. Intensity was measured as the root-mean-square
(RMS) value of the detrended accelerometer signal. The
modulation of the output of each sensor was used to represent

Page 3

dynamic characteristics of the tasks, and was calculated as the
range of the auto-covariance of each channel. Rate of
movement was represented by the dominant frequency
component below 10 Hz. Periodicity was measured by
computing the ratio of energy in the dominant frequency
component to the total energy below 10 Hz. Coordination
between body segments on the left and right side and proximal
and distal segments was captured in three aspects: magnitude
(obtained by calculating the correlation coefficient), delay
(estimated as the time lag corresponding to the peak of the
cross-correlation function) and similarity (measured by the
value of the peak of the cross-correlation function).
Approximate entropy was used as a measure of signal
complexity.
C. Feature Benchmarking
To estimate the actual computational cost of extracting
each feature on the SHIMMER platform, the time needed to
compute each feature for a 5-second epoch of simulated
accelerometer data was measured. Data were synthesized as
sampled at 100 Hz. Estimates were derived for 100 epochs
and average values were derived to compare the
computational cost associated with each feature. The overhead
associated with measuring the start and end times was
negligible because start and end times were measured by
reading the value of a time register incremented by a 32 kHz
clock, converted later into milliseconds. The processor
currently utilized by the SHIMMER nodes (MSP430) does not
have a floating-point unit. Floating-point operations were
therefore simulated via software, thus significantly increasing
the computation time required for the operation compared to
integer operations.
D. Support Vector Machines
The SVM [6] method was selected due to its success in
many classification problems, including gene classification,
speaker identification; face image detection, and text
categorization. SVM success can be attributed to several
properties. SVM optimize an objective function that is convex,
hence guaranteed to find an optimal solution; many other
classification algorithms only guarantee that local optima be
reached. SVM have the ability to generate nonlinear decision
boundaries, by mapping the feature space into a higher
dimensional space (using kernels) where classes are linearly
separable. The kernel “trick” allows SVM to project data to a
high dimensional space without added computational cost by
replacing inner products in the higher dimensional space with
kernels in the lower dimensional input space. SVM
demonstrate good generalization performance because they
implement the structural risk minimization principle [6]. SVM
construct linear hyperplanes, in high dimensional space, that
maximize the margin (separation) between classes. The
PRTools4 toolbox was used to implement SVM [7]. The
specific SVM implementation we used relies on the one-vs.-
rest approach to tackle the multi-class classification problem.


IV. RESULTS
The results of our feature benchmarking tests showed that
frequency features appear to be the most demanding from a
computational standpoint. Cross-correlation based features
appear to require the largest power consumption, given the
need for transmitting data among nodes in order to combine
accelerometer time series recorded by different nodes. The
approximate entropy is ranked next, followed by the root
mean square value; the data range is the least computational
demanding feature to be estimated. Estimating data range
feature values required approximately 2 ms, root mean square
feature values required about 20 ms, and approximate entropy
feature values required about 1 s. Estimating Fast Fourier
Transform outputs required on average longer than 4.5 s.
These simulation results suggest that the data range and root
mean square features can be estimated on the nodes without
major interference with other operations that occur on the
node (e.g. data sampling and transferring to the SD card,
synchronization of multiple nodes), that estimating the
approximate entropy feature is more challenging but
compatible with available technology, and that frequency
features should be only utilized for off-line analysis of the
data. Furthermore, our simulations indicated that cross-
correlation based features can be estimated at a relatively low
cost from a computational standpoint. Estimation of these
features required about 80 ms for each 5 s segment of
accelerometer data. However, estimation of these features
required transmitting at least one of the accelerometer time
series between nodes, leading to significant power
consumption.

TABLE 1
PREDICTION ERROR (%) FOR FEATURE COMBINATIONS
(A) Tremor, (B) Bradykinesia and (C) Dyskinesia
1 feature 2 features 3 features

A
6.6 (±8.0) 3.1 (±2.9) 2.5 (±3.3)
4 features

5 features

2.7 (±3.3) 2.8 (±3.6)

B

2.2 (±3.3)

1.6 (±2.7)

1.4 (±2.2)

1.4 (±3.0)

1.7 (±3.7)

C

3.7 (±3.0)

1.9 (±2.7)

1.8 (±1.8)

1.6 (±2.1)

1.2 (±0.9)

Table 1 summarizes the results of the simulations
performed to assess the effect of individual data feature types
and combinations of feature types on the reliability of the
estimates of clinical scores. The values shown in this table
were obtained by applying SVM with a third order polynomial
kernel to the feature sets. Results are shown for simulations
performed using a single feature type, two feature types, etc.
up to all five feature types. Numerical values shown in
Table 1 are the minimum average prediction error values
(standard deviation values are shown in parenthesis) for each
symptom and motor complication when the estimates of
clinical scores were derived using a different number of
feature types.

Page 4

Results derived when estimating clinical scores of tremor
show that the use of a single feature type can lead to an
average prediction error value as low as 6.6 %. This result was
obtained using the approximate entropy feature. A further
decrease in prediction error value is shown when two and
three feature types were utilized. With two feature types, the
lowest average error (3.1 %) was achieved by utilizing root
mean square and data range features. A prediction error value
slightly higher (i.e. 3.5 %) was achieved by using root mean
square and approximate entropy features. When three feature
types were utilized, the lowest average prediction error (2.5 %)
was achieved by using the root mean square, the data range,
and the approximate entropy feature types. Interestingly, none
of these combinations of feature types includes the frequency
feature type and the cross-correlation based feature type.
These feature types are not particularly desirable in the
context of using the SHIMMER platform because they are
either very demanding from a computational standpoint (the
frequency features) or they require wireless transmission of
data (the cross-correlation based features). It should also be
noted that approximate entropy appears to be a key feature in
estimating the severity of tremor. Our results appear to
confirm previous observations by Vaillancourt et al. [8] that
approximate entropy is relevant to studying Parkinsonian and
other types of tremor.
Results for bradykinesia show that, in general, clinical
scores for this symptom can be predicted with higher
reliability (i.e. lower prediction error) than scores for tremor.
When using only one feature type, prediction error values as
low as 2.2 % can be achieved by utilizing cross-correlation
based feature types. The lowest value of prediction error that
can be achieved without using cross-correlation based or
frequency feature types is 7.1 %. This prediction error value
can be achieved by using the approximate entropy feature.
When two feature types were considered, the lowest
prediction error achievable without using cross-correlation
based or frequency feature types was 5.1 %, which was
obtained combining root mean square and approximate
entropy feature types. When three feature types were used, the
best result was achieved by combining root mean square,
cross-correlation based, and approximate entropy feature
types. The only combination not using cross-correlation based
or frequency feature types (i.e. the one utilizing root mean
square, data range, and approximate entropy feature types) led
to a prediction error equal to 3.4 %. Combining four and all
five feature types led to a further reduction in prediction error
values for clinical scores of bradykinesia. Overall, the results
for bradykinesia showed that the use of cross-correlation
based or frequency feature types was key to minimize
prediction error values for a given number of feature types.
However, not utilizing these feature types resulted in only a
moderate increase in prediction error values of a few
percentage points (i.e. 2 % to 5 % according to the number of
utilized feature types). The use of the approximate entropy
feature appeared to be key to achieve these results.
Results for dyskinesia showed prediction error values
similar to those obtained for bradykinesia. When a single
feature type was utilized, a minimum prediction error value of
3.7 % was achieved using the approximate entropy feature.
When using two feature types, an average prediction error
value of 1.9 % was obtained by using cross-correlation based
features and the approximate entropy as input features to the
SVM. Similar prediction error values were achieved by using
root mean square value and approximate entropy (1.9 %) and
data range and approximate entropy (2.0 %), thus suggesting
that the approximate entropy feature might play a key role in
predicting the severity of dyskinesia. The best result achieved
when combining three feature types was a prediction error of
1.8 % (achieved when using frequency feature types, cross-
correlation based types, and approximate entropy). A slight
increase in prediction error values (i.e. a value of 5.6 %) was
observed when using a combination of feature types, not
including frequency feature types and cross-correlation based
types. Combinations of four feature types led to a prediction
error as low as 1.6 %, whereas using all five feature types led
to a prediction error value of 1.2 %.
Overall, these results demonstrate that the severity of
tremor, bradykinesia, and dyskinesia can be reliably estimated
using the procedures proposed in this paper with as little as
two features. Furthermore, the outcomes of our analysis
suggest that such results can be achieved without including
frequency and cross-correlation based feature types. Finally,
the results demonstrate that an average prediction error not
exceeding a few percentage points can be achieved for the
symptoms and motor complications discussed above by solely
utilizing root mean square value and approximate entropy
features.
ACKNOWLEDGMENT
This work was supported by the National Institute of
Neurological Disorders and Stroke, National Institutes of
Health under the grant #R21NS045401-02.
REFERENCES
[1] TN Chase, “Levodopa therapy: consequences of the non physiologic
replacement of dopamine”, Neurology, 50(Suppl5): S17-S25, 1998
JA Obeso, CW Olanow, JG Nutt, “Levodopa motor complications in
Parkinson's disease”, Trends Neurosci, 23: 2-7, 2000
AE Lang, AM Lozano, “Parkinson's disease. First of two parts”, N
Engl J Med, 339(16): 1044-1053, 1998
AE Lang, AM Lozano, “Parkinson's disease. Second of two parts”, N
Engl J Med, 339(16): 1130-1143, 1998
S Fahn, RL Elton, “Unified Parkinson’s Disease Rating Scale”. In Fahn
S (Ed), Recent Developments in Parkinson’s Disease, MacMillan
Healthcare Information, 153-163, 1987
V. Vapnik, The Nature of Statistical Learning Theory. New York, NY:
Springer-Verlag, 1995.
R. P. W. Duin, P. Juszczak, P. Paclik, E. Pekalska, D. de Ridder, and D.
M. J. Tax, "PRTools4, A Matlab Toolbox for Pattern Recognition," in
Delft University of Technology, 2004.
D. E. Vaillancourt, M. M. Sturman, L. Verhagen Metman, R. A.
Bakay, and D. M. Corcos,"Deep brain stimulation of the VIM thalamic
nucleus modifies several features ofessential tremor," Neurology, vol.
61, pp. 919-25, 2003.
[2]
[3]
[4]
[5]
[6]
[7]
[8]