Effects of liquid stimuli on dual-axis swallowing accelerometry signals in a healthy population.

Abstract

Dual-axis swallowing accelerometry has recently been proposed as a tool for non-invasive analysis of swallowing function. Although swallowing is known to be physiologically modifiable by the type of food or liquid (i.e., stimuli), the effects of stimuli on dual-axis accelerometry signals have never been thoroughly investigated. Thus, the objective of this study was to investigate stimulus effects on dual-axis accelerometry signal characteristics. Signals were acquired from 17 healthy participants while swallowing 4 different stimuli: water, nectar-thick and honey-thick apple juices, and a thin-liquid barium suspension. Two swallowing tasks were examined: discrete and sequential. A variety of features were extracted in the time and time-frequency domains after swallow segmentation and pre-processing. A separate Friedman test was conducted for each feature and for each swallowing task.
Significant main stimulus effects were found on 6 out of 30 features for the discrete task and on 5 out of 30 features for the sequential task. Analysis of the features with significant stimulus effects suggested that the changes in the signals revealed slower and more pronounced swallowing patterns with increasing bolus viscosity.
We conclude that stimulus type does affect specific characteristics of dual-axis swallowing accelerometry signals, suggesting that associated clinical screening protocols may need to be stimulus specific.

Full text

Page 1

RESEARCHOpen Access
Effects of liquid stimuli on dual-axis swallowing
accelerometry signals in a healthy population
Joon Lee1,2,3, Ervin Sejdić1,4, Catriona M Steele1,2,5, Tom Chau1,4*
Abstract
Background: Dual-axis swallowing accelerometry has recently been proposed as a tool for non-invasive analysis of
swallowing function. Although swallowing is known to be physiologically modifiable by the type of food or liquid
(i.e., stimuli), the effects of stimuli on dual-axis accelerometry signals have never been thoroughly investigated.
Thus, the objective of this study was to investigate stimulus effects on dual-axis accelerometry signal characteristics.
Signals were acquired from 17 healthy participants while swallowing 4 different stimuli: water, nectar-thick and
honey-thick apple juices, and a thin-liquid barium suspension. Two swallowing tasks were examined: discrete and
sequential. A variety of features were extracted in the time and time-frequency domains after swallow
segmentation and pre-processing. A separate Friedman test was conducted for each feature and for each
swallowing task.
Results: Significant main stimulus effects were found on 6 out of 30 features for the discrete task and on 5 out of
30 features for the sequential task. Analysis of the features with significant stimulus effects suggested that the
changes in the signals revealed slower and more pronounced swallowing patterns with increasing bolus viscosity.
Conclusions: We conclude that stimulus type does affect specific characteristics of dual-axis swallowing
accelerometry signals, suggesting that associated clinical screening protocols may need to be stimulus specific.
Background
Dysphagia refers in general to swallowing disorders [1],
and is a common consequence of neurological condi-
tions such as stroke, cerebral palsy, or Parkinson’s dis-
ease [2]. Adverse effects of dysphagia include degraded
psycho-social well-being [3], dehydration and malnutri-
tion [4,5], and compromised immune system secondary
to malnutrition [4]. Furthermore, dysphagia can jeopar-
dize airway protection during pharyngeal swallowing,
heightening the risk of entry of foreign material into the
unprotected airway during swallowing (aspiration),
which may lead to aspiration pneumonia [6]. Devastat-
ing outcomes of aspiration pneumonia range from hos-
pitalization to death [7]. The current gold standard in
dysphagia assessment is the videofluoroscopic swallow-
ing study (VFSS) [1,8]. In this imaging technique, X-ray
video of the pharyngeal region is recorded while the
patient swallows food or liquid stimuli mixed with bar-
ium. The primary objectives of VFSS are to determine
the nature and severity of dysphagia and to devise
appropriate intervention techniques. Many smaller
healthcare institutions are unable to provide VFSS,
resulting in long wait times for patients with dysphagia
[9]. In addition, VFSS is neither practical nor feasible
for long-term or day-to-day monitoring of dysphagia.
Recognizing the limitations of VFSS access, several
alternative techniques have been investigated. Examples
include pulse oximetry [10], cervical auscultation [11],
and electrophysiological methods [12]. All these techni-
ques employ non-invasive signal modalities and easy-to-
attach sensors. Among such alternatives, swallowing
accelerometry is the cervical vibration measurement
technique that utilizes an accelerometer, and has been
the focus of several recent studies (e.g. [13,14]). Notably,
a few of these studies have incorporated digital signal
processing and pattern recognition schemes in an effort
to automatically detect abnormal swallows [13,15]. Suc-
cessful diagnostic algorithms based on swallowing accel-
erometry have the potential to be implemented as a
portable device that might be used for screening or day-
to-day monitoring. Although single-axis accelerometry
in the anterior-posterior (A-P) direction has been
* Correspondence: tom.chau@utoronto.ca
1Bloorview Research Institute, 150 Kilgour Road, Toronto, Ontario, Canada
Lee et al. BioMedical Engineering OnLine 2010, 9:7
http://www.biomedical-engineering-online.com/content/9/1/7
© 2010 Lee et al; licensee BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons
Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.

Page 2

extensively studied, a recent dual-axis accelerometry
study showed that the superior-inferior (S-I) direction
contains swallowing information that is absent in the A-
P direction [14].
The physiological origins of dual-axis accelerometry
signals are yet to be completely elucidated. However,
given the cervical location of sensor placement, the
dominant source is presumably the mechanical move-
ment of the hyolaryngeal structure during the oral and
pharyngeal phases of swallowing. It is logical to expect
that the esophageal phase would generate negligible
vibrations that can be measured at the thyroid cartillage
following the descent of the hyoid and larynx at the
conclusion of the pharyngeal phase. This has recently
been supported by Zoratto et al. [16], whose study
showed that the hyoid movement during swallowing at
least partially contributes to dual-axis accelerometry sig-
nals. Also, Reddy et al. [17] reported that there exists a
significant correlation between the A-P accelerometry
signal and the the extent of laryngeal elevation during
swallowing. An acoustic source has been proposed as
well [18], but was rejected as a meaningful source of
dual-axis accelerometry [14].
Although deglutition (i.e. swallowing) involves a well-
defined pattern of neuronal activations, it can be modi-
fied according to sensory information conveyed from
the oral cavity and pharynx [2,19]. Afferent neurons
associated with mechanoreceptors and chemoreceptors
are largely responsible for such sensory feedback, and
they are primarily influenced by stimulus characteristics.
The use of thickened liquids is a well-known clinical
application of swallow modification using different sti-
muli to elicit safer and more efficient swallows [20,21].
A number of studies have investigated how stimulus
characteristics, including viscosity, density, volume,
taste, and texture, influence swallowing behaviours.
Steele et al. [20] considered thin, nectar-thick, and
honey-thick liquids and found that viscosity and density
influenced the durations of oropharyngeal transit and
downward tongue dorsum movement, measured using
electromagnetic midsagittal articulography. However,
they found no other stimulus effects on tongue beha-
viors in healthy adults. Logemann et al. [22] showed
beneficial effects of sour taste on several swallowing
measures in patients with dysphagia. In a similar study,
Pelletier and Lawless [21] reported that airway invasion
was significantly mitigated in patients with neurogenic
oropharyngeal dysphagia by using a high-intensity sour
stimulus (citric acid) as opposed to water. They specu-
lated that the observed improvement in swallowing
could be attributable to enhanced gustatory and trigem-
inal stimulation. Moreover, Chi-Fishman and Sonies
[23] used ultrasonography to study the effects of bolus
viscosity and volume on hyoid kinematics in healthy
individuals. They found significant main effects of visc-
osity on the duration of hyoid movement and of volume
on the extent of hyoid excursion and hyoid movement
velocity. Lastly, the principle that stimulus characteris-
tics might influence swallowing is supported by a mon-
key study by Martin et al. [24], which showed that the
majority of the swallow-related neurons in the tongue
primary motor cortex were associated with a mechanor-
eceptive field on the tongue dorsum.
In spite of the known effects of stimulus characteris-
tics on swallowing function, there is a lack of knowledge
to date regarding stimulus effects on dual-axis accelero-
metry signals. This is a particularly important issue for
abnormal swallow detection based on accelerometry,
because observed differences between healthy and
abnormal swallows might in fact reflect normal varia-
tions attributable to the influences of different stimuli.
Furthermore, this question needs to be addressed for
the development of an abnormal swallow detection
device, so that either the use of the device can be lim-
ited to one particular stimulus type or the detection
algorithm is designed to be stimulus-independent or sti-
mulus-adaptive. Although our previous study found no
significant stimulus effects on a small set of features
extracted from dual-axis accelerometry signals in a
healthy population [14], a more comprehensive investi-
gation is needed to provide a definitive answer. Depen-
dence on stimuli has been identified in other
swallowing-related signals such as electromyography
[25,26] and nasal airflow [27].
The main contribution of the present study is the
investigation of the effects of four different liquid stimuli
on a broad collection of time and time-frequency fea-
tures extracted from dual-axis swallowing accelerometry
signals, acquired from healthy individuals. The analysis
is strictly from a signal standpoint, since the relationship
between swallowing physiology and dual-axis accelero-
metry is still unclear as mentioned above. The results of
this study will provide new insight into the ways in
which modulations of swallowing function are mani-
fested in accelerometry signals. We anticipate that the
results of this study would be of interest to clinicians
interpreting accelerometry signals for dysphagia assess-
ment and to engineers developing diagnostic algorithms
based on accelerometry.
Methods
Signal Acquisition
Seventeen healthy adults (8 males) with no history of
dysphagia or any neurological impairment participated
in this study. Prior to acceptance into the study, all par-
ticipants underwent a standardized oral mechanism
examination and a water swallow screening test con-
ducted by a registered speech-language pathologist to
Lee et al. BioMedical Engineering OnLine 2010, 9:7
http://www.biomedical-engineering-online.com/content/9/1/7
Page 2 of 10

Page 3

confirm the absence of any clinical signs of dysphagia.
The mean age of the participants was 46.9 ± 23.8 years.
This study was approved by the research ethics commit-
tees at the Toronto Rehabilitation Institute, Bloorview
Kids Rehab, and University of Toronto. Each participant
gave informed consent prior to participation.
Each participant was asked to swallow 4 different liquid
stimuli: water, a 40% weight per volume thin liquid bar-
ium suspension (prepared using water and Liquid Poli-
bar™ barium, E-Z-EM), and commercially prepared pre-
thickened nectar-thick and honey-thick apple juices
(RESOURCE®, Novartis Nutrition). All stimuli were self-
administered via a cup. No standardized sip volume was
enforced, and the stimuli were served chilled. For each
stimulus, each participant completed sequences of 4 swal-
lows in two different tasks: discrete and sequential swal-
lowing. In the discrete task, participants were instructed
to take 4 sips of a comfortable size and to remove the
cup from the lips between consecutive sips. In the
sequential task, participants were asked to take 4 conse-
cutive sips without taking the cup off the lips. For both
tasks, it was made clear that there was no requirement to
finish the contents of the cup for each sequence. Partici-
pants were also asked to initiate each swallowing
sequence upon hearing an audible cue, which was gener-
ated 5 seconds after signal recording started. Participants
were allowed to complete each swallowing sequence at a
self-selected, comfortable pace. Each stimulus-task combi-
nation was repeated twice, with the exception of the
water-discrete combination, which was repeated 3 times,
yielding 17 sequences per participant in total.
Dual-axis accelerometry signals were acquired via an
accelerometer (ADXL322, Analog Devices) placed at a
midline cervical location just below the thyroid cartilage.
The attachment was accomplished with a double-sided
electrode collar (650455, VIASYS Healthcare). The two
axes of the accelerometer were oriented along the A-P and
S-I directions. See Figure 1 for an illustration of the sensor
placement and axial orientation. Each of the two channels
corresponding to the two axes was then passed through a
pre-amplifier with a bandpass filter (Model P55, Grass
Technologies, West Warwick, RI). The amplifier provided
10-times amplification, and the lower and upper cutoff fre-
quencies of the bandpass filter were set at 0.1 Hz and 3
kHz, respectively. Both signal channels were sampled by a
custom LabVIEW application at 10 kHz. These filtering,
amplification, and sampling specifications have been justi-
fied in a previous swallow accelerometry study [14].
Swallow Segmentation and Pre-processing
The 4 swallows in each sequence were segmented using
the sequential fuzzy c-means algorithm described in
[28], which was designed for dual-axis accelerometry
signals. Subsequently, the results of this automatic
segmentation were visually verified, and any incorrectly
segmented swallows were manually segmented based on
visual inspection. However, some swallows could not be
segmented even by the manual segmentation and hence
were excluded from the study. In the end, 1,114 swal-
lows were available for pre-processing and analysis.
Each segmented signal was then filtered through axis-
specific finite impulse response (FIR) filters that were
designed to annul the unwanted effects introduced by
the signal acquisition system [29]. The modified covar-
iance method of autoregressive (AR) modeling was uti-
lized to model the data acquisition system using
baseline recordings, and inverse filtering on the con-
structed AR model yielded the FIR filters. Next, the seg-
mented and filtered signals were denoised via a 10-level
discrete wavelet decomposition with the discrete Meyer
wavelet and soft-thresholding. The global denoising
threshold, Tden, was determined empirically based on
the 1st-level detail signal, d1, as follows [30]:
T
med dn
den
(| |) log
.
1
2
0 6745
(1)
where n is signal length and med is the median
operator.
Feature Extraction
For each swallow, a number of time and time-frequency
features were extracted from the pre-processed A-P sig-
nal, X = {x1, x2, ..., xn}, and S-I signal, Y = {y1, y2, ..., yn}.
A variety of features were deployed to obtain a compre-
hensive description of the dual-axis accelerometry sig-
nals corresponding to each swallow, from a signal
perspective. The following subsections describe the
computation of each feature in detail. For the features
that were identically extracted from both axes, the
description below is based only on the A-P signal, X.
The same formulation applies for the S-I signal.
Time Domain Features
• The mean of the amplitude values of the signal is a
measure of the location of the amplitude distribu-
tion, and was computed as follows:
Xi
i
n
n
x


1
1
(2)
• The variance of the amplitude distribution of the
signal is a measure of the spread of the distribution
as well as AC signal power. An unbiased estimate of
the variance was computed as follows:

XiX
i
n
n
x
22
1
1

1

()
(3)
Lee et al. BioMedical Engineering OnLine 2010, 9:7
http://www.biomedical-engineering-online.com/content/9/1/7
Page 3 of 10

Page 4

• The skewness of the amplitude distribution is a
measure of the asymmetry of the distribution, and
was computed as follows:



1
1
n
3
1
1
n
2
1
3 2
/
,
()
()
X
xi
X
in
xi
X
in









(4)
Skewness has been employed in characterization of
the dual-axis accelerometry signal [14].
• The kurtosis of the amplitude distribution is a
measure of the peakedness of the distribution, and
was computed as follows:



2
1
n
4
1
1
n
2
1
2
,
()
()
X
xi
X
in
xi
X
in









(5)
This higher moment has been utilized in a previous
dual-axis accelerometry study [14].
• The entropy rate measure introduced by Porta et
al. [31,32] quantifies the extent of regularity in a sig-
nal and is particularly useful for characterizing a sto-
chastic process in which some relationship among
consecutive data points is anticipated. These kinds
of relationships are expected in swallowing accelero-
metry given that swallowing is a well-defined physio-
logical process. Prior to the actual computation of
the entropy rate feature, X was normalized to zero
mean and unit variance, by subtracting μXand divid-
ing by sX. The normalized X was then quantized
into 10 equally spaced levels represented by integers
from 0 to 9, ranging from the minimum to maxi-
mum value. With the quantized signal denoted as
ˆ
ˆ ,ˆ ,...,ˆ
X x xxn

12
, sequences of consecutive points
in ˆXof length L, 10 ≤ L ≤ 30, L Î ℤ+, were coded
as a series of integers, ΩL= {w1, w2, ..., wn-L+1},
according to the following:

wx
ˆ
x
ˆ
x
ˆ
ii L
 
L
i L
 
L
i


1
1
2
20
1010 10

(6)
This implies that wiranged from 0 to 10L-1. Base 10
was used because there were 10 quantization levels.
The Shannon entropy of ΩLwas defined as follows:
SE Lpjpj
LL
L
e
j
( )( )log ( )





0
101
(7)
where p
j in ΩL, approximated by the corresponding sample
L

(j) represents the probability of the value
Figure 1 Accelerometry sensor placement. A sagittal view of the cervical region showing the orientation and polarity of the two axes of the
accelerometer and nearby anatomical structures.
Lee et al. BioMedical Engineering OnLine 2010, 9:7
http://www.biomedical-engineering-online.com/content/9/1/7
Page 4 of 10

Page 5

frequency. The normalized entropy rate was then
computed as follows:
NER L
SE L SE LSE perc L
SE
( )
( )()
( )
1
( )
1
( )

 
1

(8)
where perc(L) is the percentage of the coded integers
in ΩLthat occurred only once. Finally, an index of
regularity, r, was calculated as the entropy rate fea-
ture in this study:
 
1
min NER L
(( ))
(9)
which ranged from 0 (maximum randomness) to 1
(maximum regularity).
• The memory of a signal quantifies the temporal
extent of the correlation among nearby data points.
To compute a measure of signal memory, the auto-
correlation of the signal was computed from zero to
the maximum possible time lag and was normalized
so that the autocorrelation at zero lag was unity.
The memory was then estimated by the 1/e width,
which is the time duration from zero lag to the
point where autocorrelation becomes less than or
equal to 1/e ≈ 0.3679. This particular threshold has
been used before to estimate signal memory in dual-
axis accelerometry [14].
• The Lempel-Ziv (L-Z) complexity [33] is a measure
of the predictability of the signal. It has been applied
in a number of biomedical applications, ranging from
complexity characterization of DNA sequences [34]
to analyses of brain information transmission [35].
Further, Aboy et al. [36] discussed how to interpret
the L-Z complexity in the realm of biomedical signal
analysis and concluded that the L-Z complexity is a
useful scalar estimator of the bandwidth of a random
process and the harmonic variability in quasi-periodic
signals. Prior to the computation of the actual com-
plexity value, X was first converted to a sequence of
finite symbols by using 100 equally spaced quantiza-
tion levels ranging from the minimum to the maxi-
mum in X. Next, the quantized signal Sn
sn} was decomposed into k blocks so that Sn
F2, ..., Fk]. A block was a sequence of consecutive
symbols of length ℓ - j +1 expressed as follows:
1= {s1, s2, ...,
1= [F1,
 



S s s
j
sj n j
, ,
jj



,,...,,
1
1

(10)
The first block was simply initialized to be the first
symbol, i.e. F1= S1
determined to be
1= s1. Subsequent blocks were
m
h
h
Smm
m
m





1
1
1
1,,

(11)
where hmis the ending index for Fm, such that Fm
+1is a unique sequence of minimal length in the
sequence Shm
1
was computed as follows and utilized in this study:
1
1
. Finally, the normalized complexity
C
kn
n

log100
(12)
The logarithmic base of 100 came from the fact that
there were 100 quantization levels.
• The temporal swallow duration in seconds, 0.0001
(n-1), was utilized as a feature. Bolus consistency has
been shown to influence the duration of the pharyn-
geal swallow [2,20,23].
• Extending from the entropy rate measure above,
Porta et al. [37,38] also introduced a method of
quantifying the cross-entropy rate between two sto-
chastic processes. This measure describes the pre-
dictability of a data point in one signal given a
sequence of current and past data points in the
other signal. First, both X and Y were normalized,
quantized, and coded using the same methodology
as for the entropy rate feature, yielding L
Xand
L
Y, respectively, with 10 ≤ L ≤ 30, L Î ℤ+. In addi-

structed as follows:
tion,
L
X Y
/
X Y
/
1
X Y
/
2
n L
 
X Y
/
www
,,...,

2
was con-
wxyyy
i
X Y
/
i L
 
L
i L
 
L
i L
 
L
i


2
1
2
2
3
30
10 101010

(13)
where
are the quantized signals. Then, with SEX(L), SEY(L),
and SEX/Y(L) representing the Shannon entropies
(defined in (7)) of L
tively, the normalized cross-entropy of X given Y
was computed as follows:
ˆ
ˆ ,ˆ ,...,ˆ
x x
12
Xxn

and
ˆ
ˆ ,ˆ ,...,ˆ
y y
12
Yyn

X, L
Y, and L
X Y
/, respec-
NCERL
SEX YL
/( )
SEYLSEX
( )
1
percX YL
SEX
X Y
/( )
()( )
1
/( )

  
1

(14)
where percX/Y(L) is the percentage of the elements
in L
that occurred only once. Next, the uncou-
pling function was defined as follows:
X Y
/
UFLmin NCER
(
L NCER
( ),
L
X Y
,
X Y
/
Y X
/
( ) ( ))

(15)
Finally, the following index of synchronization was
computed and utilized as the cross-entropy rate fea-
ture in this study:
X YX Y
,
min UF
(
L
,
( ))

1
(16)
which ranged from 0 (X and Y are completely
uncoupled) to 1 (perfect synchronization).
Lee et al. BioMedical Engineering OnLine 2010, 9:7
http://www.biomedical-engineering-online.com/content/9/1/7
Page 5 of 10