U.S. patent application number 10/869024 was filed with the patent office on 2005-12-22 for apparatus and method for detecting swallowing activity.
This patent application is currently assigned to Bloorview MacMillan Children's Centre, a corp. registered under the Ontario Corporations Act. Invention is credited to Berall, Glenn, Casas, Michael J., Chau, Thomas T.K., Kenny, David J..
Application Number | 20050283096 10/869024 |
Document ID | / |
Family ID | 35481594 |
Filed Date | 2005-12-22 |
United States Patent
Application |
20050283096 |
Kind Code |
A1 |
Chau, Thomas T.K. ; et
al. |
December 22, 2005 |
Apparatus and method for detecting swallowing activity
Abstract
An apparatus and method for detecting swallowing activity is
provided. In an embodiment, a method includes receiving an
electronic signal from an accelerometer that represents swallowing
activity, extracting at least two features from the signal,
classifying the signal as a type of swallowing activity based on
the extracted features, and generating an output of the
classification. Exemplary activities include swallows, aspirations,
movement and vocal artifacts. By indicating whether an activity is
a swallow or an aspiration, the manner in which a patient afflicted
with an increased likelihood for aspirations is fed can be adjusted
to increase the likelihood of achieving a swallow instead of an
aspiration during feeding. In turn this could reduce
hospitalizations for aspiration pneumonia in patients with acute or
chronic injury.
Inventors: |
Chau, Thomas T.K.; (Toronto,
CA) ; Kenny, David J.; (Toronto, CA) ; Casas,
Michael J.; (Toronto, CA) ; Berall, Glenn;
(Toronto, CA) |
Correspondence
Address: |
TORYS LLP
79 WELLINGTON ST. WEST
SUITE 3000
TORONTO
ON
M5K 1N2
CA
|
Assignee: |
Bloorview MacMillan Children's
Centre, a corp. registered under the Ontario Corporations
Act
Toronto
CA
|
Family ID: |
35481594 |
Appl. No.: |
10/869024 |
Filed: |
June 17, 2004 |
Current U.S.
Class: |
600/593 ;
128/925; 600/595 |
Current CPC
Class: |
A61B 5/6822 20130101;
A61B 5/7264 20130101; A61B 5/11 20130101; A61B 2562/0219 20130101;
A61B 5/4205 20130101 |
Class at
Publication: |
600/593 ;
600/595; 128/925 |
International
Class: |
A61B 005/103; A61B
010/00; A61B 005/117 |
Claims
1. A method for detecting swallowing activity comprising the steps
of: receiving an electronic signal representing swallowing
activity; extracting at least two features from said signal;
classifying said signal as a type of swallowing activity based on
said features; and, generating an output representing said
classification.
2. The method of claim 1 wherein said electronic signal is
generated by an accelerometer.
3. The method of claim 2 wherein said features include at least one
of stationarity, normality and dispersion ratio.
4. The method of claim 3 wherein said classifying step is performed
using a radial basis neural network.
5. The method of claim 1 wherein said swallowing activity includes
at least one of a swallow and an aspiration.
6. The method of claim 2 wherein said extracting step includes
stationarity as one of said features, said extracting step of
stationarity including the following sub-steps: dividing said
signal into a plurality of non-overlapping bins; determining a
total number of total number of reverse arrangements,
(A.sub.Total,) in a mean square sequence is determined; extracting
said stationarity feature (z), determined according to the
following equation: 11 z = A Total - A A where: .mu..sub.A is the
mean number of reverse arrangements expected for a stationary
signal of the same length. .sigma..sub.A is the standard deviation
for an equal length stationary signal
7. The method of claim 6 wherein each of said bins is between about
one ms and about nine ms in length.
8. The method of claim 6 wherein each of said bins is between about
three ms and about seven ms in length.
9. The method of claim 6 wherein each of said bins is about five
milliseconds ("ms") in length.
10. The method of claim 2 wherein said extracting step includes
normality as one of said features, said extracting step of
normality including the following sub-steps: standardizing said
signal to have zero mean and unit variance ("s"). dividing said
standardized signal into a plurality of bins ("I") each of about
0.4 Volts, where 12 max ( s ) - min ( s ) 0.4 ,and wherein a
highest bin extends to infinity and a lowest bin extends to
negative infinity. determining observed frequencies ("n") for each
said bin by counting the number of samples in the standardized
signal ("s") that fell within each said bin. determining expected
frequencies {circumflex over (m)} for each said bin is determined
under the assumption of normality, using a Chi-square (X.sup.2)
statistic using the following: 13 X ^ 2 = i = 1 I ( n i - m ^ i ) 2
m ^ i determining said normality feature using the following:
log.sub.10({circumflex over (X)}.sup.2)
11. The method of claim 2 wherein said extracting step includes
dispersion ratio as one of said features, said extracting step of
dispersion ratio including the following sub-steps: determining a
mean absolute deviation of said signal according to the following:
14 S 1 = 1 n i = 1 n | x i - med ( x ) | determining an
interquartile range, S.sub.2, of said signal extracting said
dispersion ratio according to the following: 15 S 1 S 2
12. A device for detecting swallowing activity comprising: an input
device for receiving an electronic signal from a sensor, said
electronic signal representing swallowing activity; a microcomputer
connected to said input device and operable to extract at least two
features from said signal; said microprocessor further operable to
classify said signal as a type of swallowing activity based on said
features; and, an output device connected to said microcomputer for
generating an output representing said classification.
13. The device claim 12 wherein said sensor is an
accelerometer.
14. The device of claim 13 wherein said features include at least
one of stationarity, normality and dispersion ratio.
15. The device of claim 14 wherein said classifying is performed
using a radial basis neural network.
16. The device of claim 12 wherein said swallowing activity
includes at least one of a swallow and an aspiration.
17. The device of claim 13 wherein said extracting includes
stationarity as one of said features, said extracting of
stationarity including: dividing said signal into a plurality of
non-overlapping bins; determining a total number of total number of
reverse arrangements, (A.sub.Total,) in a mean square sequence is
determined; extracting said stationarity feature (z), determined
according to the following equation: 16 z = A Total - A A where:
.mu..sub.A is the mean number of reverse arrangements expected for
a stationary signal of the same length. .sigma..sub.A is the
standard deviation for an equal length stationary signal
18. The device of claim 17 wherein each of said bins is between
about one ms and about nine ms in length.
19. The device of claim 17 wherein each of said bins is between
about three ms and about seven ms in length.
20. The device of claim 17 wherein each of said bins is about about
five milliseconds ("ms") in length.
21. The device of claim 13 wherein said extracting includes
normality as one of said features, said extracting of normality
including: standardizing said signal to have zero mean and unit
variance ("s"). dividing said standardized signal into a plurality
of bins ("I") each of about 0.4 Volts, where 17 max ( s ) - min ( s
) 0.4 ,and wherein a highest bin extends to infinity and a lowest
bin extends to negative infinity; determining observed frequencies
("n") for each said bin by counting the number of samples in the
standardized signal ("s") that fell within each said bin;
determining expected frequencies {circumflex over (m)} for each
said bin is determined under the assumption of normality, using a
Chi-square (X.sup.2) statistic using the following: 18 X ^ 2 = i =
1 I ( n i - m ^ i ) 2 m ^ i determining said normality feature
using the following: log.sub.10({circumflex over (X)}.sup.2)
22. The device of claim 13 wherein said extracting includes
dispersion ratio as one of said features, said dispersion ratio
including: determining a mean absolute deviation of said signal
according to the following: 19 S 1 = 1 n i = 1 n | x i - med ( x )
| determining an interquartile range, S.sub.2, of said signal
extracting said dispersion ratio according to the following: 20 S 1
S 2
Description
FIELD OF THE INVENTION
[0001] The present invention relates generally to the diagnosis of
aspiration and more particularly relates to an apparatus and method
for detecting swallowing and related activity.
BACKGROUND OF THE INVENTION
[0002] Dysphagia refers to any deglutition (swallowing) disorder,
including abnormalities within the oral, pharyngeal and esophageal
phases of swallowing. Dysphagia is common in individuals with
neurological impairment, due to, for example, cerebral palsy,
cerebrovascular accident, brain injury, Parkinson's disease, stroke
and multiple sclerosis. Individuals with dysphagia are often at
risk of aspiration. Aspiration refers to the entry of foreign
material into the airway during inspiration. Aspiration may
manifest itself in a number of different ways. The individual may
begin to perspire and the face may become flushed. Alternatively,
the individual may cough subsequent to swallowing. In silent
aspiration, there are no overt clinical or easily recognizable
signs of bolus inhalation. The invention is particularly useful for
individuals with silent aspiration, but is also applicable for
other manifestations of aspiration. Aspiration bears serious health
consequences such as chronic lung disease, aspiration pneumonia,
dehydration, and malnutrition.
[0003] Dysphagia afflicts an estimated fifteen million people in
the United States. Certain sources indicate that fifty thousand
people die each year from aspiration pneumonia (Dray et al., 1998).
The occurrence of diffuse aspiration bronchiolitis in patients with
dysphagia is not uncommon, regardless of age (Matsuse et al.,
1998). Silent aspiration is especially prominent in children with
dysphagia, occurring in an estimated 94% of that population
(Arvedson et al., 1994). Half of stroke survivors have swallowing
difficulties (Zorowitz & Robinson, 1999), which translates to
500,000 people per year in the United States, (Broniatowski et al.,
2001), and aspiration is reported in 75% of these cases while 32%
report chest infections (Perry & Love, 2001). The incidence of
dysphagia is particularly significant in acute care settings
(25-45%), chronic care units (50%) (Finiels et al., 2001) and homes
for the aged (68%) (Steele et al., 1997). Dysphagia tremendously
diminishes quality of life for people of all ages, compromising not
only medical, but social, emotional and psychosocial
well-being.
[0004] The modified barium swallow using videofluoroscopy is the
current gold standard for confirmation of aspiration (Wright et
al., 1996). Its clinical utility in dysphagia management continues
to be asserted (e.g., Martin-Harris, 2000; Scott et al., 1998). The
patient ingests barium-coated material and a video sequence of
radiographic images is obtained via X-radiation. The modified
barium swallow procedure is invasive and costly both in terms of
time and labor (approximately 1,000 health care dollars per
procedure in Canada), and renders the patient susceptible to the
effects of ionizing radiation (Beck & Gayler, 1991).
[0005] Fibreoptic endoscopy, another invasive technique in which a
flexible endoscope is inserted transnasally into the hypopharynx,
has also been applied in the diagnosis of post-operative aspiration
(Brehmer & Laubert, 1999) and bedside identification of silent
aspiration (Leder et al., 1998). Fibreoptic endoscopy is generally
comparable to the modified barium swallow in terms of sensitivity
and specificity for aspiration identification (e.g., Madden et al.,
2000; Leder & Karas, 2000), with the advantage of bedside
assessment.
[0006] Pulse oximetry has been proposed as a non-invasive adjunct
to bedside assessment of aspiration (e.g., Sherman et al., 1999;
Lim et al., 2001). However, several controlled studies comparing
pulse oximetric data to videofluorscopic (Sellars et al., 1998) and
fiberoptic endoscopic evaluation (Leder, 2000; Colodny, 2000) have
raised doubts about the existence of a relationship between
arterial oxygen saturation and the occurrence of aspiration.
[0007] Cervical auscultation involves listening to the breath
sounds near the larynx by way of a laryngeal microphone,
stethoscope or accelerometer (Zenner et al., 1995) placed on the
neck. It is generally recognized as a limited but valuable tool for
aspiration detection and dysphagia assessment in long-term care
(Zenner et al., 1995; Cichero & Murdoch, 2002; Stroud et al.,
2002). However, when considered against the gold standard of
videofluoroscopy, bedside evaluation even with cervical
auscultation yields limited accuracy (40-60%) in detecting
aspirations (Sherman et al., 1999; Selina et al., 2001; Sellars et
al., 1998). Indeed, our recent research shows that aspirations
identified by clinicians using cervical auscultation, represent
only a quarter of all aspirations (Chau, Casas, Berall & Kenny,
submitted).
[0008] Swallowing accelerometry (Reddy et al., 2000) is closely
related to cervical auscultation, but has entailed digital signal
processing and artificial intelligence as discrimination tools,
rather than trained clinicians. In clinical studies, accelerometry
has demonstrated moderate agreement with videofluoroscopy in
identifying aspiration risk (Reddy et al., 1994) where as the
signal magnitude has been linked to the extent of laryngeal
elevation (Reddy et. al, 2000). Recently, fuzzy committee neural
networks have demonstrated extremely high accuracy at classifying
normal and "dysphagic" swallows (Das et al., 2001). However, prior
art swallowing accelerometry only provides limited information in
classifying normal from "dysphagic" swallows and does not provide
broader information about the clinical status of the patient.
[0009] Administration of videofluoroscopy or nasal endoscopy
requires expensive equipment and trained professionals such as a
radiologist, otolaryngologist or speech-language pathologist
(Sonies, 1994). Invasive procedures are not well-tolerated by
children and cannot be practically administered for extended
periods of feeding. There is a need for an economical, non-invasive
and portable method of aspiration detection, for use at the bedside
and outside of the institutional setting.
SUMMARY OF THE INVENTION
[0010] It is an object of the present invention to provide a novel
apparatus and method for detecting swallowing activity that
obviates or mitigates at least one of the above-identified
disadvantages of the prior art.
[0011] An aspect of the invention provides a method for detecting
swallowing activity comprising the steps of:
[0012] receiving an electronic signal representing swallowing
activity;
[0013] extracting at least two features from the signal;
[0014] classifying the signal as a type of swallowing activity
based on the features;
[0015] and,
[0016] generating an output representing the classification.
[0017] The electronic signal can be generated by an accelerometer.
The features can include at least one of stationarity, normality
and dispersion ratio. The classifying step can be performed using a
radial basis neural network.
[0018] The swallowing activity can include at least one of a
swallow and an aspiration.
[0019] The extracting step can include stationarity as one of the
features, the extracting step of stationarity including the
following sub-steps:
[0020] dividing the signal into a plurality of non-overlapping
bins;
[0021] determining a total number of total number of reverse
arrangements, (A.sub.Total,) in a mean square sequence is
determined;
[0022] extracting the stationarity feature (z), determined
according to the following equation: 1 z = A Total - A A
[0023] where:
[0024] .mu..sub.A is the mean number of reverse arrangements
expected for a stationary signal of the same length.
[0025] .sigma..sub.A is the standard deviation for an equal length
stationary signal
[0026] Each of the bins can be between about one ms and about nine
ms in length. Each of the bins can be between about three ms and
about seven ms in length. Each of the bins can be about five
milliseconds ("ms") in length.
[0027] The extracting step can include normality as one of the
features, the extracting step of normality including the following
sub-steps:
[0028] standardizing the signal to have zero mean and unit variance
("s").
[0029] dividing the standardized signal into a plurality of bins
("I") each of about 0.4 Volts, where 2 max ( s ) - min ( s ) 0.4
,
[0030] and wherein a highest bin extends to infinity and a lowest
bin extends to negative infinity.
[0031] determining observed frequencies ("n") for each the bin by
counting the number of samples in the standardized signal ("s")
that fell within each the bin.
[0032] determining expected frequencies {circumflex over (m)} for
each the bin is determined under the assumption of normality, using
a Chi-square (X.sup.2) statistic using the following: 3 X ^ 2 = i =
1 I ( n i - m ^ i ) 2 m ^ i
[0033] determining the normality feature using the following:
log.sub.10({circumflex over (X)}.sup.2)
[0034] The extracting step can include a dispersion ratio as one of
the features, the extracting step of dispersion ratio including the
following sub-steps:
[0035] determining a mean absolute deviation of the signal
according to the following: 4 S 1 = 1 n i = 1 n | x i - med ( x )
|
[0036] determining an interquartile range, S.sub.2, of the
signal
[0037] extracting the dispersion ratio according to the following:
5 S 1 S 2
[0038] Another aspect of the invention provides a device for
detecting swallowing activity comprising an input device for
receiving an electronic signal from a sensor. The electronic signal
can represent swallowing activity. The device also comprises a
microcomputer connected to the input device that is operable to
extract at least two features from the signal. The microprocessor
is further operable to classify the signal as a type of swallowing
activity based on the features. The device also includes an output
device connected to the microcomputer for generating an output
representing the classification.
BRIEF DESCRIPTION OF THE DRAWINGS
[0039] The invention will now be described by way of example only,
and with reference to the accompanying drawings, in which:
[0040] FIG. 1 is a schematic representation of an apparatus for
detecting swallowing activity in accordance with an embodiment of
the invention;
[0041] FIG. 2 is a flow chart depicting a method for detecting
swallowing activity in accordance with another embodiment of the
invention;
[0042] FIG. 3 is a set of graphs showing exemplary signals that can
be detected using the apparatus in FIG. 1; and
[0043] FIG. 4 is a graph showing exemplary output that can be
generated by the method in FIG. 2.
DETAILED DESCRIPTION OF THE INVENTION
[0044] As used herein the terms "swallow" and "penetration" are
distinguished from the term "aspiration". As used herein, a
"swallow" is the safe passage of foodstuffs from the oral cavity,
through the hypopharynx and into esophagus. Further, a swallow is
accompanied by a period of apnea with no entry of foodstuffs into
the protected airway. "Penetration" is the entry of foreign
material into the airway but not accompanied by inspiration.
However, an "aspiration" is the entry of foreign material into the
airway during inspiration. As used in relation to the embodiments
discussed below, the term "swallowing activity" means a swallow or
an aspiration or the absence of either, but in other embodiments
"swallowing activity" can refer to other types of activities
including penetration.
[0045] Referring now to FIG. 1, an apparatus for detecting
swallowing activity is indicated generally at 30. Apparatus 30
includes an accelerometer 34 that is positioned on the throat of a
patient 38. In a present embodiment, accelerometer 34 is placed
infer-anterior to the thyroid notch, so that the axis of the
accelerometer 34 is aligned to measure anterior-posterior
vibrations. Apparatus 30 also includes a computing device 42 that
is connected to accelerometer 34 via a link 46. Link 46 can be
wired or wireless as desired and corresponding to appropriate
interfaces on accelerometer 34 and device 42. Apparatus 30 is
operable to receive acceleration signals from accelerometer 34 that
reflect swallowing activity in patient 38.
[0046] In a present embodiment, accelerometer 38 is the EMT 25-C
single axis accelerometer from Siemens Canada, Mississauga, Ontario
Canada ("EMT 25-C"). Other accelerometers that can be used will
occur to those of skill in the art.
[0047] In a present embodiment, computing device 42 is based on the
computing environment and functionality of a personal digital
assistant that includes a chassis 50 that frames a display 54 for
presenting user output and a plurality of keys 58 for receiving
user input. Computing device 42 thus includes an interface to allow
device 42 to connect to accelerometer 34 via link 46. Computing
device 42 thus includes any suitable arrangement of microprocessor,
random access memory, non-volatile storage, operating system, etc.
As will be explained in greater detail below, computing device 42
is operable to receive signals from accelerometer 34 and to detect
swallowing activity from such signals, and report on those
activities by presenting output on display 54.
[0048] In order to help explain certain of these implementations
and various other aspects of apparatus 30, reference will now be
made to FIG. 2 which shows a method for detecting swallowing
activity and which is indicated generally at 200. However, it is to
be understood that apparatus 30 and/or method 200 can be varied,
and need not work exactly as discussed herein in conjunction with
each other, and that such variations are within the scope of the
present invention.
[0049] Beginning first at step 210, signals representing swallowing
activity are received. When method 200 is implemented using
apparatus 30, step 210 refers to the generation of electrical
signals by accelerometer 34 and the receipt of those signals at
computing device 42. The use of accelerometer 34 means that
acceleration signals representing the swallowing activity of
patient 38 are received, and due to the unique characteristics of
the EMT 25-C accelerometer used in the present embodiment, unique
features can be found in the appearance of those signals. FIG. 3
shows examples of signals that can be received using EMT 25-C,
indicated generally at 300, and specifically at 304, 308 and 312.
Speaking in very general terms, signal 304 is an example of typical
paediatric aspiration signals that portray weak or wide-sense
stationarity; signal 308 is an aspiration signal that portrays
nonstationarity due to evolving variance; and signal 312 is an
aspiration signal that portrays nonstationarity due to time-varying
frequency and variance structure.
[0050] However, it is to be understood that signals 300 are simply
raw data, and can represent aspirations or swallows or motion
artifact. It has been determined by the inventors that the
distribution of median acceleration magnitude is right-skewed for
both aspiration and swallows. Due to the skewness of the
distribution, gamma distribution is used to estimate the spread and
location parameters within signals 300. In particular, the spread a
and location b parameters of the gamma distributions for
aspirations and swallows that can be associated with signals such
as signals 300 are summarized in Table I.
1TABLE I Location and spread of signal accelerations Aspirations
Swallows Maximum Maximum likelihood 95% confidence likelihood 95%
confidence Parameter estimate interval estimate interval Spread (a)
1.3647 g [0.9343, 1.7952] 3.642 g [2.2713, 5.0128] Location 1.176 g
[0.732, 1.62] 0.063 g [0.041, 0.086] (b)
[0051] The stationarity and normality characteristics of signals
300 are summarized in Table II. Stationarity is measured by the
nonparametric reverse arrangements tests while normality is
measured by a chi-squared distribution-based test of histogram bin
counts. Further details about stationarity and normality can be
found in "Random Data Analysis and Measurement Procedures" 3.sup.rd
Edition, Julius S. Bendat and Allan G. Pierson, John Wiley &
Sons Inc., (c) 2000, New York ("Bendat"), the contents of which are
incorporated herein by reference. Chapter 10 of Bendat discusses
for tests for stationarity, while Chapter 4 of Bendat discusses
regarding normality.
[0052] Table II thus shows a very general, exemplary, summary of
how aspirations and swallows can correspond to the stationarity and
normality characteristics of received signals such as signals
300.
2TABLE II Stationarity and normality characteristics of signal
accelerations Aspirations Swallows Stationarity 41% not stationary
46% not stationary Normality 90% violating normality 100% violating
normality
[0053] Due to the skewness of the distributions of the bandwidths,
a gamma distribution is used to determine the location estimate.
The frequency bandwidths can be calculated using a discrete wavelet
decomposition at ten levels and determining the level at which the
cumulative energy (starting from the final level of decomposition)
exceeded 85% of the total energy. This determines the 85% bandwidth
for the signal in question.
[0054] The location estimate of the about 85% frequency bandwidth
can be between about 700 Hz to about 1100 Hz for aspiration
signals, and more preferably can be between about 900Hz and about
950 Hz, and even more preferably between about 910Hz and about 940
Hz, and still further preferably about 928 Hz for aspiration
signals.
[0055] The location estimate of the about 85% frequency bandwidth
can be between about 400 Hz to about 700 Hz for swallow signals,
and more preferably can be between about 500Hz and about 650 Hz,
and even more preferably between about 590 Hz and about 630 Hz, and
still further preferably about 613 Hz for swallows.
[0056] Having received signals at step 210, method 200 advances to
step 220. At step 220, a determination is made as to whether an
event is present inside the signals received at step 210. The
criteria for making such a determination is not particularly
limited. In a present embodiment, when computing device 42 receives
a signal magnitude from accelerometer 34 that exceeds an "on"
threshold (in a present embodiment of about 0.025 Volts ("V")) for
a pre-determined "onset" period (in a present embodiment about
thirty milliseconds ("ms")), event initiation is identified and
signal recording begins. The next about 12,000 samples are
recorded, corresponding to about 1.2 seconds ("s") of data.
Back-trimming is then performed to determine when the signal
activity substantially ceased. Such back-trimming involves counting
the number of data samples below about 0.05 V, starting from the
end of the recording. Once this count exceeds about thirty data
points, the end of the useful signal is deemed to have been
identified and the end of the signal is trimmed therefrom. In a
present embodiment, 12000 samples are recorded, but about 15,000
samples (i.e. about 1.5 s of above threshold signal activity) can
also be recorded for analysis as a single signal. In other
embodiments other numbers of samples can be recorded, as desired.
If the foregoing criteria are not met, then it is determined at
step 220 that an event has not occurred and method 200 returns to
step 210. However, if the criteria is met then method 200 advances
from step 220 to step 230, and the signals that are recorded at
step 220 is retained for use at step 230.
[0057] Next, at step 230, features are extracted from the recorded
signals. In a presently preferred embodiment, stationarity,
normality and dispersion ratio are three features that are
extracted.
[0058] In order to extract the stationarity feature, the procedure
in Chapter 10 of Bendat is employed. First, the received signal, is
divided into non-overlapping bins each of about five milliseconds
("ms") (i.e. for a total of fifty samples) in length. (The received
signal can, however, be divided into non-overlapping bins of
between about one ms and about nine ms, or more preferably between
about three ms and about seven ms.) Where the signal length,
defined herein as "L" is not an integral multiple of fifty, the
signal was trimmed at the beginning and end of the signal by
approximately (L mod 50)/2. Next, the mean square value within each
window was computed. Next, the total number of reverse
arrangements, referred to herein as A.sub.Total, in the mean square
sequence is determined. Finally, z-deviate serves as the
stationarity feature which is determined according to Equation
1.
[0059] where: 6 z = A Total - A A Equation 1
[0060] .mu..sub.A is the mean number of reverse arrangements
expected for a stationary signal of the same length.
[0061] .sigma..sub.A is the standard deviation for an equal length
stationary signal.
[0062] In order to extract the normality feature, an adaptation of
the procedure in Chapter 4 of Bendat is employed. First, the signal
is standardized to have zero mean and unit variance. The
standardized signal is referred to herein as "s". Next, the
amplitude of the standardized signal, s, is divided into I bins
each of about 0.4 Volts, where 7 I = max ( s ) - min ( s ) 0.4
.
[0063] The highest bin extended to infinity and the lowest bin
extended to negative infinity.
[0064] Next, the observed frequencies n for each bin are determined
by counting the number of samples in the standardized signal that
fell within each bin. The expected frequencies {circumflex over
(m)} for each bin is determined under the assumption of normality.
The Chi-square statistic was computed as shown in Equation 2. 8 X ^
2 = i = 1 I ( n i - m ^ i ) 2 m ^ i Equation 2
[0065] Finally, the normality feature is computed as shown in
Equation 3.
[0066] log.sub.10 ({circumflex over (X)}.sup.2) Equation 3
[0067] In order to determine the dispersion ratio feature, the mean
absolute deviation of each signal is determined according to
Equation 4. 9 S 1 = 1 n i = 1 n | x i - med ( x ) | Equation 4
[0068] Next, the interquartile range, S.sub.2, of each signal is
determined. The interquartile range is defined in Chapter 2 of
"Introduction to robust estimation and hypothesis testing", Rand R.
Wilcox, 1997, Academic Press, CA. Finally, the dispersion ratio
feature is determined according to Equation 5. 10 S 1 S 2 Equation
5
[0069] Having extracted these features from the signal, method 200
advances to step 240, at which point the signal is classified based
on the features extracted at step 230. In a presently preferred
embodiment, the classification is performed using a radial basis
function neural network implemented on the microcontroller of
device 42 to classify swallowing events in real-time, as either
swallows or aspirations. Further details about such a radial basis
function neural network can be found in Chapter 5 of "Neural
Networks for Pattern Recognition", Christopher Bishop, 1995,
Clarendon Press, Oxford ("Bishop"), the contents of which are
incorporated herein by reference. The network is operable to take
the three extracted features as inputs, and output a single number
as its classification of the detected type of swallowing activity.
In particular, an output level of about 0.1 is assigned to
represent swallows and an output level of about 0.9 to represent
aspirations. The network architecture consists of three inputs
corresponding to each extracted feature, eighty-nine radial basis
function units determined from an interactive training procedure as
outlined in "Bishop" and one output unit, representing swallowing
or aspiration. While eighty-nine radial basis units is presently
preferred, in other embodiments from about seventy-five to about
one-hundred radial basis units can be used, and in other
embodiments from about eighty to about ninety-five radial basis
units can be used, all corresponding to one output. The first layer
is nonlinear and the second layer is linear. Put in other words,
the first layer of the network consists of the nonlinear radial
basis functions while the second layer of the network is a weighted
linear summation of the radial basis function outputs.
[0070] Referring now to FIG. 4, a scatter plot is shown for the
results of performing steps 210-240 for a number of different
signals. The scatter plot in FIG. 4 is only two dimensional,
showing only a plot of the stationarity features vs. the normality
features. It can be seen that the squares on the scatter plot
indicate where aspirations actually occurred, whereas the circles
indicate swallows actually occurred. The scatter plot was generated
while performing method 200 in conjunction with videofluroscopy so
that the actual swallowing activity could be verified, not
withstanding the classification performed at step 230, so that the
classifications made at step 230 could be verified for accuracy.
The line indicated at 400 in FIG. 4 represents a rough dividing
line between classifications associated with swallows and
aspirations. While some measurements in the scatter plot show a
classification that does not reflect the actual type of swallowing
activity, the majority of swallowing events are in fact correctly
classified. Further improvement to the results shown in FIG. 4 are
obtained when the third feature, dispersion ratio, is used to
assist in the determination.
[0071] Method 200 then advances to step 250, at which point an
output is generated corresponding to the classification performed
at step 240. Thus, where a particular event was classified as a
swallow, then display 54 of device 42 would be instructed to
present the message "SWALLOW", whereas if the event was classified
as an aspiration then display 54 of device 42 would be instructed
to present the message "ASPIRATION". Such messages presented by
device 42 could also include colours (e.g. green associated with
swallows, red associated with aspirations) and/or auditory signals
(e.g. no sound for swallow, beeping for aspirations).
[0072] Using method 200, an individual feeding patient 38 can
adjust how the feeding is being performed in order to reduce
aspirations and increase swallows. Such adjustments to feedings can
be based on changing consistency or type of food, the size and/or
frequency of mouthfuls being offered to patient 38, and the
like.
[0073] It should now be understood that as method 200 is
implemented using device 42, the microcontroller of device 42 will
be provided with software programming instructions corresponding to
method 200.
[0074] While only specific combinations of the various features and
components of the present invention have been discussed herein, it
will be apparent to those of skill in the art that desired subsets
of the disclosed features and components and/or alternative
combinations of these features and components can be utilized, as
desired. For example, it is also to be understood that other types
of vibration sensors other than accelerometer 34 can be used with
appropriate modifications to computing device 42. While presently
less preferred, another sensor can include a sensor that measure
displacement (e.g microphone), while having computing device 42
record received displacement signals over time. Another type of
sensor can include a sensor that measures velocity, having
computing device 42 record received velocity signals over time.
Such signals can then be converted into acceleration signals and
processed according to the above, or other techniques of feature
extraction and classification thereof that work with the type of
received signal can be employed, as desired.
[0075] As an additional example, while at step 230 of method 200
stationarity, normality and dispersion ratio are three features
that are extracted, it is to be understood that in other
embodiments other features and/or combinations thereof can be
extracted that can be used to detect a swallowing event. For
example, while presently less preferred, it can be desired to
simply extract any two of stationarity, normality and dispersion
ratio in order to make a determination as to whether a particular
swallowing event is to be classified as a swallow or
aspiration.
[0076] Furthermore, while computing device 42 is a personal digital
assistant having a programmable microprocessor and display, in
other embodiments device 42 can simply be an electronic device that
includes circuitry dedicated to processing signals from an
accelerometer (or other sensor) and classifying those signals as
different types of swallowing activity. Similarly, the device can
simply include a set of indicator lights--e.g. a pair of indicator
lights, one light for indicating a swallow, the other for
indicating an aspiration. Whatever the format of device 42, device
42 can also include an interface for connection to a personal
computer or other computing device so that updated programming
instructions for detecting aspirations, swallows and/or other types
of swallowing activity can be uploaded thereto.
[0077] The above-described embodiments of the invention are
intended to be examples of the present invention and alterations
and modifications may be effected thereto, by those of skill in the
art, without departing from the scope of the invention which is
defined solely by the claims appended hereto.
* * * * *