U.S. patent application number 12/938995 was filed with the patent office on 2014-01-02 for physiological signal denoising.
The applicant listed for this patent is Marina Brockway. Invention is credited to Marina Brockway.
Application Number | 20140005988 12/938995 |
Document ID | / |
Family ID | 49778984 |
Filed Date | 2014-01-02 |
United States Patent
Application |
20140005988 |
Kind Code |
A1 |
Brockway; Marina |
January 2, 2014 |
PHYSIOLOGICAL SIGNAL DENOISING
Abstract
Physiological signals are denoised. In accordance with an
example embodiment, a denoised physiological signal is generated
from an input signal including a desired physiological signal and
noise. The input signal is decomposed from a first domain into
subcomponents in a second domain of higher dimension than the first
domain. Target subcomponents of the input signal that are
associated with the desired physiological signal are identified,
based upon the spatial distribution of the subcomponents. A
denoised physiological signal is constructed in the first domain
from at least one of the identified target subcomponents.
Inventors: |
Brockway; Marina; (St. Paul,
MN) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Brockway; Marina |
St. Paul |
MN |
US |
|
|
Family ID: |
49778984 |
Appl. No.: |
12/938995 |
Filed: |
November 3, 2010 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
61257718 |
Nov 3, 2009 |
|
|
|
61366052 |
Jul 20, 2010 |
|
|
|
Current U.S.
Class: |
703/2 ; 708/311;
708/322 |
Current CPC
Class: |
A61B 5/7253 20130101;
G06K 9/0051 20130101; G06F 17/14 20130101; A61B 5/0452 20130101;
H03H 17/0248 20130101; A61B 5/04017 20130101; G06K 9/0053
20130101 |
Class at
Publication: |
703/2 ; 708/322;
708/311 |
International
Class: |
G06F 17/10 20060101
G06F017/10; G06G 7/60 20060101 G06G007/60; G06F 17/14 20060101
G06F017/14 |
Claims
1. A method for computing a denoised physiological signal from an
input signal including a desired pseudoperiodic physiological
signal and noise, the method comprising: decomposing the input
signal from a first domain into subcomponents of the input signal
in a second domain that is different than the first domain;
identifying target subcomponents of the input signal that are
associated with the desired physiological signal based upon the
time-based distribution of the subcomponents; and reconstructing a
denoised physiological signal in the first domain from at least two
of the identified target subcomponents.
2. The method of claim 1, wherein a cycle of the input signal is
partitioned into at least two time windows based upon a
characteristic band of frequencies associated with the desired
signal within each of the at least two time windows, and
identifying target subcomponents includes: identifying the
locations of the at least two time windows based upon the location
of a feature point of the cycle, and identifying a target
subcomponent in a time window as a subcomponent that contains more
energy within the characteristic band of frequencies of the desired
signal, than energy outside the characteristic band of
frequencies.
3. The method of claim 1, wherein identifying target subcomponents
of the input signal that are associated with the desired
physiological signal based upon the time-based distribution of the
subcomponents includes identifying target subcomponents using
principal component analysis.
4. The method of claim 1, wherein identifying target subcomponents
of the input signal that are associated with the desired
physiological signal includes identifying noise subcomponents of
the input signal as subcomponents having more noise energy than
desired physiological signal energy, and identifying, as target
subcomponents, the subcomponents that are not noise
subcomponents.
5. The method of claim 1, wherein identifying target subcomponents
of the input signal that are associated with the desired
physiological signal includes identifying as noise subcomponents of
the input signal, the subcomponents having more noise energy than
desired physiological signal energy, removing at least one of the
noise subcomponents, and identifying, as target subcomponents, the
subcomponents remaining after removing the noise subcomponents.
6. The method of claim 1, wherein identifying target subcomponents
of the input signal that are associated with the desired
physiological signal includes identifying subcomponents having more
desired physiological signal energy than noise energy.
7-8. (canceled)
9. The method of claim 1, wherein the input signal has identifiable
cycles that can be partitioned into time windows, each time window
having an associated band of frequencies, and identifying target
subcomponents of the input signal that are associated with the
desired physiological signal includes identifying a cycle of the
input signal, partitioning the cycle into at least two time
windows, identifying noise subcomponents of the input signal in a
time window as subcomponents that are not associated with a
frequency band of a desired physiological signal within the time
window, and identifying the target subcomponents as subcomponents
in the time window that are not identified as noise
subcomponents.
10. (canceled)
11. The method of claim 1, wherein the desired physiological signal
is a mammalian ECG signal, and identifying target subcomponents of
the input signal that are associated with the desired physiological
signal includes identifying the location of a QRS complex in a
cardiac cycle of the ECG, identifying a first time window that
includes the QRS complex, identifying at least one time window in
the cycle that does not include the QRS complex, and identifying
the target subcomponents, within the identified at least one time
window, as subcomponents of the input signal that are associated
with a frequency band of the desired physiological signal within
the identified at least one window.
12. The method of claim 1, wherein the desired physiological signal
is a mammalian ECG signal, and identifying target subcomponents of
the input signal that are associated with the desired physiological
signal includes identifying the location of a QRS complex in a
cardiac cycle of the ECG, identifying a first time window that
includes the QRS complex, identifying at least one time window in
the cycle that does not include the QRS complex, identifying as
noise subcomponents, within the identified at least one time
window, subcomponents of the input signal that are not associated
with a frequency band of the desired physiological signal within
the identified at least one window, and identifying the target
subcomponents as subcomponents in the identified at least one time
window that were not identified as noise subcomponents.
13. The method of claim 1, wherein the desired physiological signal
is a mammalian ECG signal, and identifying target subcomponents of
the input signal that are associated with the desired physiological
signal includes identifying the location of a QRS complex in a
cardiac cycle of the ECG, identifying a first time window that
includes the QRS complex, identifying at least one time window in
the cycle that does not include the QRS complex, and comparing the
frequency content of the subcomponents, within the at least one
time window, with a characteristic band of frequencies identified
via spectral analysis of a database containing recordings of
representative ECG signals for the mammalian species, and
identifying target subcomponents as subcomponents that have more
power within the characteristic band of frequencies than power
outside of the characteristic band of frequencies.
14. (canceled)
15. The method of claim 1, wherein decomposing the input signal
from a first domain to a second domain, includes using one of: a
discrete cosine transform; a wavelet related transform; a
Karhunen-Loeve transform; a Fourier transform; a Gabor transform;
and a filter bank.
16. The method of claim 1, wherein identifying target subcomponents
includes using at least one of: spatially selective filtering;
independent component analysis; and periodic component
analysis.
17. The method of claim 1, wherein decomposing the input signal
from a first domain into subcomponents in a second domain includes
using a mathematical transform to transform the input signal from
the first domain into the subcomponents, and reconstructing a
denoised physiological signal in the first domain includes
combining at least two of the identified target subcomponents using
an inverse of the mathematical transform.
18. The method of claim 1, wherein the input signal includes at
least one of an electrocardiogram signal, a heart sound signal, a
blood pressure signal, a respiratory information signal, a blood
flow signal, a photoplethysmography signal, and a periodic
stimulation response signal.
19. The method of claim 1, further including, before decomposing
the input signal, segmenting the input signal according to the
timing of an applied stimulation to assemble a plurality of
signals.
20-21. (canceled)
22. A method for computing a denoised ECG signal from an input
signal including a desired ECG signal and noise, the method
comprising: decomposing the input signal into subcomponents;
identifying a location of the QRS complex of a cardiac cycle in the
ECG signal; identifying a first time window in the cardiac cycle
that includes the QRS complex; identifying at least one time window
in the cardiac cycle that does not include the QRS complex; for
each of the identified time windows, identifying target
subcomponents as subcomponents that contain more energy that is
within the band of frequencies characteristic of the desired ECG
signal in the time window than energy that is outside the band of
frequencies of the desired ECG signal; and reconstructing a
denoised physiological signal using at least two of the identified
target subcomponents.
23. The method of claim 22, wherein decomposing the input signal
into subcomponents includes transforming the input signal from a
first domain into subcomponents in a second domain using at least
one of a discrete cosine transform; a wavelet related transform; a
Karhunen-Loeve transform; a Fourier transform; a Gabor transform;
and a filter bank.
24. The method of claim 22, wherein identifying target
subcomponents includes identifying target subcomponents using at
least one of: principal component analysis; independent component
analysis; spatially selective filtering; and periodic component
analysis.
25. The method of claim 22, wherein the desired ECG signal is an
ECG of a species and identifying the band of frequencies
characteristic of the desired ECG signal in each of the identified
time windows includes: identifying a database containing recordings
of representative ECG signals for the species, identifying the
location of the QRS complex in cardiac cycles of the ECG,
partitioning the cardiac cycles into said identified time windows,
and using spectral analysis to identify the band of frequencies
characteristic of the desired ECG signal in the identified time
windows.
26-31. (canceled)
32. A system for denoising at least one of a P-wave and T-wave of
an input signal including a desired ECG signal and noise, the
system comprising: a logic circuit; and a memory circuit that
stores instructions that, when executed by the logic circuit, carry
out the steps of decomposing the input signal into subcomponents,
identifying a location of the QRS complex of a cardiac cycle in the
ECG signal, identifying at least one time window in the cardiac
cycle that includes the at least one of the P-wave and T-wave and
does not include the QRS complex, identifying target subcomponents
in the at least one time window as subcomponents that contain more
energy within the characteristic band of frequencies of the desired
ECG signal within the at least one time window than energy outside
the band of frequencies, and reconstructing at least one of
denoised P- and T-waves using the identified target
subcomponents.
33. The method of claim 1, wherein decomposing the input signal
from a first domain into subcomponents of the input signal in a
second domain includes decomposing the input signal from a first
domain having a dimension size defined by a number of observed
signal channels in the input signal, into a second domain having a
dimension size defined by said number of channels multiplied by a
number of subcomponents in each channel.
34. The method of claim 1, further including partitioning a cycle
of the input signal into at least two time windows based upon a
characteristic band of frequencies of the desired signal within at
least one of the time windows, at least one of said time windows
containing a signal wave with wider-band frequency content than at
least one other time window, and wherein identifying the target
subcomponents includes: combining at least two subcomponents as a
time-based function, using the time-based function to identify a
location of a signal wave with wider-band frequency content,
identifying the locations of the at least two time windows based
upon the location of said signal wave, and identifying a target
subcomponent in a time window as a subcomponent that contains more
energy within the characteristic band of frequencies of the desired
signal within the window than energy outside the characteristic
band of frequencies.
35. The method of claim 1, wherein identifying the target
subcomponents includes partitioning a cycle of the input signal
into at least two different time windows and selecting target
subcomponents in each time window based upon a time-frequency
representation of a cycle of the desired signal.
36. The method of claim 35, wherein the time-frequency
representation of a cycle of the desired signal includes a spectral
content distribution that varies synchronously with a feature point
of the cycle.
37. The method of claim 35 wherein the desired signal in each of
the at least two time windows is associated with a characteristic
band of frequencies, and the target subcomponent is a subcomponent
containing more energy inside the characteristic band of
frequencies than energy outside the characteristic band of
frequencies.
38. A method for removing noise from a non-QRS portion of a cardiac
cycle of an input signal including a desired ECG signal and noise,
the method comprising: decomposing the input signal into
subcomponents; identifying a location of the QRS complex of a
cardiac cycle in the ECG signal; identifying at least one time
window in the cardiac cycle that does not include the QRS complex;
identifying target subcomponents in the at least one time window
based upon spectral energy of each target subcomponent relative to
a characteristic spectral energy of the desired ECG signal within
the at least one time window; and reconstructing the non-QRS
portion of the cardiac cycle using the identified target
subcomponents in the at least one time window.
39. The method of claim 38 wherein the at least one time window
spans the duration of the cardiac cycle outside the QRS
complex.
40. The method of claim 38 wherein the input signal is from a
species and identifying the target subcomponents includes:
identifying a characteristic spectral energy of the desired ECG
signal in the at least one time window by characterizing the
spectral energy of at least one of P-waves and T-waves of the
species, and identifying target subcomponents as subcomponents with
spectral energy that overlaps said characteristic spectral
energy.
41. The method of claim 40 wherein identifying target subcomponents
as subcomponents with spectral energy that overlaps said
characteristic spectral energy includes identifying subcomponents
that, when combined, result in a quality of signal reconstruction
greater than about 90%.
42. The method of claim 40 wherein identifying target subcomponents
as subcomponents with spectral energy that overlaps said
characteristic spectral energy includes identifying subcomponents
that, when combined, result in a quality of signal reconstruction
greater than about 95%.
43. The method of claim 11, wherein identifying target
subcomponents as subcomponents of the input signal that are
associated with a frequency band of the desired physiological
signal includes identifying subcomponents of the input signal
having a spectral energy that substantially overlaps the frequency
band.
44. The method of claim 3, wherein identifying target subcomponents
using principal component analysis includes using principal
component analysis to decorrelate the subcomponents, and after
decorrelating the subcomponents, identifying ones of the
subcomponents as the target subcomponents.
Description
RELATED PATENT DOCUMENTS
[0001] This patent document claims the benefit under 35 U.S.C.
.sctn.119 of U.S. Provisional Patent Application Ser. No.
61/257,718, filed on Nov. 3, 2009; this patent document also claims
the benefit under 35 U.S.C. .sctn.119 of U.S. Provisional Patent
Application Ser. No. 61/366,052, filed on Jul. 20, 2010; each of
these provisional patent applications is fully incorporated herein
by reference.
FIELD OF INVENTION
[0002] Various aspects of the present invention relate to the
processing of physiological signals, and more particular aspects
relate to removing noise and extracting and compressing
information.
BACKGROUND
[0003] Implantable and external devices are used to monitor
physiologic signals of human and animal subjects. These devices may
incorporate various types of sensors and can measure and record
signals from those sensors for processing by a system or monitoring
center located remote from the subject, or in other cases, the
device may perform some or all of the desired signal processing and
forward the resulting information to a remote system for display,
recording, or further processing.
[0004] The signal processing is performed to extract information
from the signal in order to assess the physiological condition of
the monitored subject and often to evaluate the response of the
subject to a therapy or experimental protocol. For example, devices
that measure ECG and blood pressure are routinely used to assess
cardiovascular function. Clinicians can use this information to
make therapy decisions and researchers can use this information to
assess the safety and utility of experimental therapies. This
information is also used for closed loop control of therapy
delivery. In other examples, measurements of peripheral nerve
activity (PNA), respiration, blood oxygen, blood glucose, EEG, EMG,
heart sounds and blood flow signals are processed to extract
information for clinical or research purposes.
[0005] There is an increasing reliance on automatic processing to
extract information in order to reduce labor and costs and to more
consistently and accurately evaluate the condition of the subject.
In therapeutic devices automatic extraction of this information is
often essential for feedback control. However, accurate automatic
extraction of information is often challenging or is compromised by
the presence of noise.
[0006] In some physiologic signal processing applications,
automated analysis is complicated by the fact that measured signals
are the result of activity of multiple sources, referred to as
multi-source signals. An example of a multi-source signal is ECG
measured on the surface of the body where electrical activity is
sensed from both the atria and ventricles as well as skeletal
muscles. It is useful, for example, to observe atrial activity
independent of ventricular activity in order to improve the
detection of atrial arrhythmias Current techniques for providing
signal source extraction of multi-source signals, such as
independent component analysis (ICA), assume independence of
sources and performance is compromised when this assumption in
invalid, such as is the case when separating atrial and ventricular
activity in an ECG. In addition, ECGs are often recorded from
ambulatory subjects using a small number of sensing leads, further
complicating signal source extraction due to the mixing of sources
inherent in a small number of leads.
[0007] Other signals, such as peripheral nerve activity (PNA) and
brainstem auditory response, have proven difficult to analyze
because of very low signal-to-noise ratio (SNR). Visual analysis of
these signals is often inadequate to detect important features and
obtain a quantitative evaluation.
[0008] Many physiological signal processing techniques have been
difficult to successfully implement under certain conditions,
particularly when processing signals from ambulatory subjects where
the signals are often quite noisy. For example, measurements of ECG
parameters such as heart rate, QT interval, PR interval, as well as
systolic and diastolic blood pressure may contain errors as a
result of the presence of noise. Detection of ventricular and
atrial arrhythmias in ECG may have excessive incidence of false
positives due to the inability of a signal processing algorithm to
provide accurate detection, particularly in the presence of noise.
Likewise, the presence of noise may result in inaccurate and
inconsistent evaluation of cardiac pathologies that are reflected
in ECG morphology. Because of lack of confidence in the accuracy of
results, human review has often been used to confirm results or
correct errors made by automated analysis algorithms.
[0009] Inaccuracies in performance can also result in excess
telecom costs when monitoring ambulatory subjects. For example,
some types of ambulatory ECG monitoring devices employ on-board
signal processing to detect arrhythmias and forward the detected
arrhythmias to a monitoring center where they are further processed
and reviewed by a human being using a data review system. Because
of limitations in existing algorithms, there is a high rate of
false positive arrhythmia detections in the ambulatory device that
results in a high volume of data transmitted from the patient to a
monitoring center. This results in excessive telecommunications
expense, the need for additional memory in the ambulatory
monitoring device, and additional expense to manually review the
data received at the monitoring center.
[0010] Various methods of data compression are also limited in
their ability to provide high levels of compression with minimal
signal distortion in part due to the presence of noise. More
efficient data compression can reduce the volume of data that must
be stored in memory on an ambulatory monitoring device as well as
reduce the volume of data transmitted from the monitored subject.
In certain applications, this can result in a reduction in telecom
expense and a reduction in power consumption in the ambulatory
monitoring device, leading to a reduction in the device size and
extension of battery life.
[0011] The presence of noise in physiological signals can be a
limiting factor in providing accurate and consistent computerized
evaluations and extraction of information, but the removal of noise
has been complicated by the fact that the noise often has spectral
content that falls within the bandwidth of the signals of interest
(referred to as in-band noise). For example atrial signals can be
contaminated by electrical activity of the ventricle, and ECG
signals can be contaminated by EMG from the skeletal muscles. The
plethora of signal sources contained within a limited number of
channels measured in a surface ECG, with each channel containing
mixed interdependent signals, renders the independent observation
of the sources of these multisource signals a very difficult
problem. This problem is further complicated when signals are
acquired from closely spaced electrodes and are contaminated by
noise, as is usually the case when monitoring patients outside a
clinic or hospital. This characterization is not only common to
electrocardiogram (ECG) signals acquired with surface or
subcutaneous leads but is also common to electrograms (EGM)
measured with intracardiac leads, blood pressure signals, pulse
oximetry signals, peripheral nerve activity (PNA) recordings,
signals representing non-invasive measurements of intracranial
pressure, and other physiologic signals collected from ambulatory
subjects. Current filtering techniques such as bandpass filtering
are effective in removing noise without distorting the signal when
the spectral content of the noise and signal are separated in the
frequency domain. Many filtering techniques capable of removing
in-band noise such as independent component analysis require that
noise and signal content are uncorrelated and independent, an
inaccurate assumption for most physiological signals.
[0012] Removing at least some of the in-band noise, or denoising,
of physiological signals can be useful in the improvement of
accuracy of computerized evaluations and has been an objective of
many prior efforts. However, the success of many prior efforts has
been limited. Various techniques have also been limited in their
ability to report characteristics of information derived by an
algorithm relative to noise, artifact, or signal morphology
changes. These and other matters have presented challenges to the
design and implementation of devices, systems and methods for
processing physiological signals.
SUMMARY
[0013] Various aspects of the present invention are directed to
devices, methods and systems involving physiological signal
processing, in a manner that addresses challenges and limitations
including those discussed above.
[0014] In accordance with various example embodiments, a
physiological signal is denoised. In various implementations,
denoising is carried out using techniques including those referred
to herein as Multi-Domain Signal Processing (MDSP), for which a
multitude of exemplary embodiments are described. The captured
signal is received in a first domain and is decomposed into
subcomponents in a second domain of higher dimension than the first
domain. The resulting subcomponents in the second domain are
processed based upon the spatial distribution of the subcomponents
using, for example, spatially selective filtering or principal
component analysis to identify those subcomponents that are
primarily associated with noise in a time segment. Subcomponents
identified as primarily associated with noise are removed and the
remaining subcomponents are combined to reconstruct a denoised
signal in the first domain.
[0015] Another example embodiment is directed to a method for
computing a denoised physiological signal from an input signal
including a desired physiological signal and noise. The input
signal is decomposed from a first domain into subcomponents of the
input signal in a second domain of higher dimension than the first
domain. From the decomposed subcomponents, target subcomponents of
the input signal that are associated with the desired physiological
signal are identified based upon the spatial distribution of the
subcomponents. A denoised physiological signal is constructed in
the first domain from at least one of the identified target
subcomponents.
[0016] Another example embodiment is directed to a method for
reducing the power of an undesired signal in a recording of a
quasi-periodic mammalian physiological signal containing desired
components and undesired components. A transform is used to
decompose the physiological signal into subcomponents representing
the signal in a time-frequency domain. At least two time windows
are identified in a cycle of the signal, each time window having a
different band of frequencies associated with desired components of
the signal. In the at least two time windows, ones of the
subcomponents associated with the desired signal components are
identified as subcomponents having spectral content overlapping a
band of frequencies associated with desired signal components
within said at least two time windows. A denoised physiological
signal is constructed using at least one of the identified
subcomponents within each of the at least two time windows.
[0017] In another aspect of the present invention, a dynamic
signal-to-noise ratio is computed as the ratio of power contained
in the subcomponents primarily associated with signal and
subcomponents primarily associated with noise.
[0018] In another aspect of the present invention, the
subcomponents associated with a signal wave of a physiological
signal are denoised and are combined to form a denoised emphasis
signal of the signal wave. The emphasis signal is subsequently
evaluated to identify feature points of the signal wave.
[0019] In another aspect of the present invention, feature points
of signal waves are evaluated to detect clinically significant
events and signal morphology characteristics.
[0020] In another aspect of the present invention, a validity
metric, computed as a function of a dynamic signal-to-noise ratio
and signal morphology characteristics, is used to assess the
accuracy, consistency, and validity of feature points, morphology
characteristics, and detected events.
[0021] In yet another aspect of the present invention, the denoised
signal or its subcomponents and detected features are used to
facilitate signal compression to reduce stored and communicated
data volume.
[0022] In yet another aspect of the present invention, signal
compression and feature and event detection improvements are used
to implement a device of reduced size and enhanced performance.
[0023] The above summary is not intended to describe each
embodiment or every implementation of the present disclosure. The
figures and detailed description that follow more particularly
exemplify various embodiments.
BRIEF DESCRIPTION OF THE DRAWINGS
[0024] The invention may be more completely understood in
consideration of the following detailed description of various
embodiments of the invention in connection with the accompanying
drawings, in which:
[0025] FIG. 1 shows a data flow diagram of Multi-Domain Signal
Processing applied to denoising a captured signal, consistent with
an example embodiment of the present invention;
[0026] FIG. 2 shows a data flow diagram of Multi-Domain Signal
Processing applied to extraction of a signal from a captured
physiological signal, consistent with an example embodiment of the
present invention;
[0027] FIG. 3 shows scatter plots, in which FIG. 3a shows a scatter
plot of subcomponents of a noisy signal before application of PCA
and ICA, and FIG. 3b shows a scatter plot of subcomponents of the
same noisy signal of FIG. 3a after application of PCA and ICA,
according to another example embodiment;
[0028] FIG. 4 illustrates a data flow diagram for computing a
dynamic signal-to-noise ratio, according to an example embodiment
of the present invention;
[0029] FIG. 5 shows an example of denoising performance and QRS
detection in a rabbit ECG with non-sustained ventricular
tachycardia and illustrates a dynamic signal-to-noise ratio updated
for every cardiac cycle, according to an example embodiment of the
present invention;
[0030] FIG. 6 shows an example of denoising performance on a 3-lead
ECG signals using PCA, ICA, and MDSP, according to another example
embodiment of the present invention;
[0031] FIG. 7 shows results representing denoising performance of
MDSP, PCA, and bandpass filtering, in connection with another
example embodiment of the present invention;
[0032] FIG. 8 shows an example of denoising performance on a single
channel noisy ECG signal, in connection with an example embodiment
of the present invention;
[0033] FIG. 9 shows an example of atrial activity extraction from a
surface ECG signal containing atrial flutter, in connection with
another example embodiment of the present invention;
[0034] FIG. 10 shows an example of atrial activity extraction from
a surface ECG signal containing atrial fibrillation, in connection
with another example embodiment of the present invention;
[0035] FIG. 11 shows an example of denoising performance on single
channel peripheral nerve activity (PNA) recording corrupted with
noise, in connection with another example embodiment of the present
invention;
[0036] FIG. 12 shows a data flow diagram for computing an emphasis
signal and finding a feature point in a signal waveform, and
example signal waveforms and corresponding emphasis signals and
feature point markers, according to another example embodiment of
the present invention;
[0037] FIG. 13 shows an example of QRS detection performance for a
challenging ECG recording, according to another example embodiment
of the present invention;
[0038] FIG. 14 shows a data flow diagram for computing an emphasis
signal corresponding to a respiration pattern, according to an
example embodiment of the present invention;
[0039] FIG. 15 shows an example of respiration signal extraction in
an ECG recording, according to another example embodiment of the
present invention;
[0040] FIG. 16 illustrates an apparatus for efficient wireless
communication of an ECG along with data flow diagrams for ECG
signal compression, according to an example embodiment of the
present invention;
[0041] FIG. 17 shows an example of 3D ECG cycle plot as an
intermediate step in ECG compression, in connection with another
example embodiment of the present invention;
[0042] FIG. 18 shows a system for evaluating ECG strips captured by
an ambulatory monitoring device along with data flow diagrams for
determining if a strip contains an arrhythmia and assessing the
validity of the result, according to an example embodiment of the
present invention;
[0043] FIG. 19 illustrates an apparatus for improving the
signal-to-noise ratio of an ECG signal and related flow chart for
processing an ECG signal, according to an example embodiment of the
present invention;
[0044] FIG. 20 illustrates an approach for spatially selective
filtering, and partitioning of a cardiac cycle of an ECG signal,
according to another example embodiment of the present
invention;
[0045] FIG. 21 illustrates a signal flow diagram for an embodiment
of denoising applied to an ECG signal, according to another example
embodiment of the present invention; and
[0046] FIG. 22 illustrates a system for denoising an ECG signal,
according to another example embodiment of the present
invention.
[0047] While the invention is amenable to various modifications and
alternative forms, specifics thereof have been shown by way of
example in the drawings and will be described in detail. It should
be understood, however, that the intention is not to limit the
invention to the particular embodiments described. On the contrary,
the intention is to cover all modifications, equivalents, and
alternatives falling within the scope of the invention including
aspects defined in the claims.
DETAILED DESCRIPTION
[0048] Various example embodiments of the present invention relate
to the processing of physiological signals, and in many
implementations, to reduce noise, extract information, characterize
signals, and/or compress the volume of data. While the present
invention is not necessarily limited to such applications, various
aspects of the invention may be appreciated through a discussion of
examples using this context.
[0049] Many embodiments described here are broadly referred to as
including an approach referred to as "Multi-Domain Signal
Processing" (MDSP), for which many different embodiments are
described by way of example. In connection with various
embodiments, the term Multi-Domain Filtering (MDF) is used herein
to refer to embodiments that use one or more MDSP-based embodiments
to denoise physiologic signals. Various embodiments are also
directed to processing a broad range of physiological signals
including but not limited to signals corresponding to ECG, blood
pressure, respiratory, heart sounds, EEG, peripheral nerve
activity, activity, temperature, photoplethysmography, tissue
impedance, blood glucose, and EMG. Different embodiments are
further directed to one or more of: improving the accuracy and
consistency of information provided under a broad range of use
scenarios, extending battery life of monitoring devices, providing
a desirably-sized monitoring device (e.g., a reduced, or small
size), and improving and/or reducing the need for human review of
analysis results and the associated expense of doing so.
[0050] In the following discussion, reference is made to cited
references listed in a numbered order near the end of this
document, which are fully incorporated herein by reference. These
references may assist in providing general information regarding a
variety of fields that may relate to one or more embodiments of the
present invention, and further may provide specific information
regarding the application of one or more such embodiments.
[0051] According to an example embodiment, physiological signals of
a subject are captured by an implantable or external device. This
device can be implemented as a part of a system that measures,
processes, and evaluates physiological data from animal or human
subjects for research, therapy titration, diagnosis, or delivery of
medical care. The device may temporarily store the processed
signals for later transmission to a system remote from the subject
for further processing, display, and reporting to a medical care
provider or researcher. The processed signals may be transmitted in
real time or they may be used within a therapy delivery device as
part of a system to control or advise administration of a therapy.
In yet another embodiment, unprocessed or partially processed
signals may be transmitted to a device or system located outside
the body of the subject for processing, display, review, reporting,
or retransmission to another system.
[0052] In many applications, physiological signals processed in
accordance with the embodiments discussed herein are multisource
signals, meaning that the signal observed by electrodes or leads
includes components that are the result of many physiological
processes and sources that are often interdependent. For example,
an ECG signal measured at the surface of the body may include
signals emanating from sources such as the atria, ventricles,
electrical noise (e.g., from sources outside the body), noise from
muscular electrical activity, signals resulting from pathologies
such as conduction defects, scars in tissue from myocardial
infarction, ischemia resulting from reduced blood flow to a region
of the heart, and other representative examples. In these and other
contexts, various embodiments of the present invention are directed
to addressing challenges relative to the analysis of ECG signals,
such as those that benefit from independent analysis for denoising
and evaluation of cardiac function.
[0053] The terms "quasi-periodic", "signal wave", "feature point",
"parameter", and "event" are used in connection with the discussion
of various embodiments as follows. The term quasi-periodic refers
to a periodic signal with a period with a cycle length that varies
with time, and a signal wave is a particular portion or aspect of a
period of a physiological signal. For example, an ECG signal may
include signal waves referred to as P, Q, R, S, T and U waves.
Another example signal wave is the QRS complex portion of an ECG
(e.g., the portion of an ECG signal corresponding to the
depolarization of the left and right ventricles). In an arterial
pressure signal, the dicrotic notch is a signal wave.
[0054] Regarding the term "feature point," many physiological
signals can be characterized as having features and parameters. A
feature point is an identified point within a physiological signal,
which may be useful for characterizing the signal and related
characteristics of the signal's source. For instance, in
heart-related signals, such as ECG, arterial pressure, and blood
flow, most cardiac cycles have a feature point or points of
interest. Examples include a point corresponding to the onset of
the Q-wave (e.g., Q-wave onset) or offset of the T-wave (e.g.,
T-wave offset) in an ECG or systole in an arterial blood pressure
signal. Each of these feature points is described by time of
occurrence and amplitude, and consecutive feature points can be
combined to form a feature signal.
[0055] With respect to the term "feature point," it is sometimes
useful to combine feature points over a predetermined period of
time, referred to as a feature signal, to compute a parameter. For
example, systole feature points can be combined to compute a mean
systolic pressure. Computing a parameter can have the effect of
reducing or eliminating short-term physiological fluctuations
(e.g., changes with respiration) in the feature signal that are not
of interest to the user.
[0056] Regarding the term "event," physiological signals may
include information that can be used for identifying the onset and
offset of an event, or simply the fact that an event occurred. For
example, when monitoring the ECG of a subject it may be useful to
know that an arrhythmic event such as ventricular tachycardia or
atrial fibrillation has occurred. In additional embodiments,
information is combined from multiple signals to compute features
and parameters. For example, the QA interval (the time difference
between Q-wave onset and the upstroke of an arterial pressure wave)
can be used as a surrogate for cardiac contractility, and employs
both a pressure and an ECG signal from a subject. QA interval
feature points are computed for a cardiac cycle, can be used to
create a feature signal, and averaged over a predetermined period
of time to create a QA parameter.
[0057] Regarding the terms referring to a "desired physiological
signal," such a signal refers to a signal that is to be extracted.
This signal may correspond, for example, to a physiological signal
within an input signals that include the physiological signal and
noise. In this context, the noise may also include a physiological
signal that is not the desired physiological signal. For instance,
where a desired physiological signal is an ECG signal from a
subject's heart, other physiological signals included with an input
signal, such as a respiratory signal, are not desirably
extracted.
[0058] A "captured signal" is a signal that is sensed and recorded
(e.g., digitized), and may also be conditioned. Conditioning of the
signal may involve amplification and the application of a filter to
remove much of the noise that is outside the bandwidth of the
signal. Following digitization, additional filtering may be applied
to further remove noise from outside the signal bandwidth,
typically using linear filtering techniques such as a finite
impulse response filter. A "denoised" signal is a captured signal
that has been processed to remove noise, such as by removing
undesirable signal components or subcomponents having spectral
content within a bandwidth of a selected (e.g., target, or desired
signal), or in-band noise. Denoising can be useful for rendering
clarity to the desired signal as a result of suppression of
undesired signal components (e.g., noise), hence making the desired
signal more suitable for analysis.
[0059] In accordance with other example embodiments of the present
invention, in-band noise of a physiological signal is reduced using
a technique involving a multi-domain filtering-type of approach,
referred to in connection with various embodiments as MDF.
Resulting signals are thus denoised, in the context that at least
some noise components in the physiological signal have been
removed, relative to the resulting signal (e.g., as re-generated
from components of the physiological signal, in a different
domain). In one embodiment, signals captured in a first domain are
decomposed into subcomponents in a second domain of higher
dimension than the first domain. To remove in-band noise from the
captured physiological signal, the subcomponents in the second
domain are processed based on their spatial distribution using, for
example, spatially selective filtering or principal component
analysis to identify those subcomponents that are primarily
associated with noise at an instant in time. Those subcomponents
that are identified as primarily associated with noise are removed
and the residual subcomponents (those not having been removed) are
combined to reconstruct a denoised signal in the first domain. A
subcomponent occurring in a time window within a cycle of the
pseudo-periodic signal is said to be associated with a signal if it
contains frequencies within a band that has previously been
characterized as being present in the signal. A subcomponent within
a time segment is considered to be primarily associated with noise
if the energy associated with the signal is attributed primarily
(e.g., more than half) to noise. In some embodiments, a larger
tolerance for noise is set, to permit signals with larger noise
content to be processed without distorting the signal (e.g., where
a desired signal is expected to include a substantial portion of
noise). Under such conditions, processing of the detected signal
can be carried out with the understanding that noise forms a
majority of the power in the subcomponent. In some implementations,
a signal is treated as primarily associated with noise when at
least about 60% of the signal's energy at that time segment is
noise energy. Likewise, a subcomponent within a time segment is
considered to be primarily associated with signal if at least half
or more (e.g., 60%) of its energy is within the band of frequencies
characterized as being present in the signal.
[0060] In the context of various embodiments, references to the
removal of subcomponents may not involve any removal or
modification of the subcomponents, but rather involve a selective
combination of those subcomponents that have been determined to be
desirable. For example, for a physiological signal having various
subcomponents, certain components can be identified as likely
representing components of a signal corresponding to a particular
physiological characteristic that is to be analyzed. These
identified (desirable) components can be selectively combined,
leaving behind other subcomponents. In this context, undesirable
subcomponents are not necessarily removed, but rather, have not
been used when forming a recombined signal corresponding to a
received physiological signal.
[0061] The subcomponents that result from decomposition in
Multi-Domain Signal Processing (MDSP) embodiments are also used in
other aspects of physiologic signal processing, for various
embodiments. A characteristic of MDSP is that a signal wave in a
multisource physiological signal can be represented using a small
number of subcomponents that contain most of the energy of the
signal wave. For example, in an ECG decomposed using a discrete
wavelet transform, a group of 3 subcomponents may contain most of
the energy found in the P-wave of an ECG while most of the energy
of a QRS complex may be contained in 4 subcomponents. A
subcomponent or its time segment containing a significant amount of
energy of a signal at that time segment is said to be associated
with the signal at that time segment. As applicable here, a
significant amount of energy of a subcomponent is an amount of
energy, corresponding to the time segment, that is at least half of
the total energy of the signal during the time segment. In various
implementations, a subcomponent may be associated with more than
one signal wave.
[0062] Subcomponents or their time segments associated with a
signal wave can be identified, isolated, and used to construct a
signal wave independent of other signal waves in a multisource
signal. For example, subcomponents associated with a P-wave of an
ECG can be identified within the second domain independent of
ventricular electrical activity, and used to reconstruct a denoised
signal representative of atrial activity. Another group of
subcomponents is associated with the T-wave of an ECG, and are used
in the extraction of repolarization activity. Signal source
extraction of the signal subcomponents in a multisource signal can
lead to more accurate analysis and evaluation of certain
physiological functions. Another embodiment involving signal source
extraction is directed to the extraction of fetal ECGs from ECG
signals captured from a pregnant female. Yet another example
embodiment is directed to extracting an oxygen saturation signal
from a photoplethysmography signal.
[0063] Subcomponents can also be used to compute a denoised
emphasis signal that exaggerates the features of particular signal
waves to facilitate the identification and detection of feature
points. For example, subcomponents associated with the T-wave of an
ECG can be identified and used to compute a denoised emphasis
signal that exaggerates the T-wave relative to other ECG signal
waves to facilitate accurate and consistent detection of T-wave
offset.
[0064] In some embodiments, physiological signal subcomponents are
used to compute a dynamic signal-to-noise ratio (dSNR) that
represents the ratio of signal energy relative to noise energy on a
sample-by-sample basis, or longer period of time such as for one or
more cardiac cycles of an ECG. The dSNR can be used for computing a
validity signal for assessing the accuracy and reliability of
information derived from the captured physiological signal. In some
embodiments, the dSNR is used to directly assess the accuracy and
reliability of derived information. If dSNR for a cardiac cycle or
a portion of a cardiac cycle is low, for example, certain
information derived from the ECG for that cardiac cycle may not be
accurate and, if it is very low, a cardiac cycle may be
uninterpretable.
[0065] MDF denoising and other aspects of one or more MDSP-based
embodiments are directed to facilitating the efficient compression
of physiological signals, which can be used to reduce the volume of
data corresponding to sensed signals. Reducing the amount of data
(e.g., by eliminating or omitting noise) serves to help the
efficient storage of data, and also to reduce the amount of data
needed in communications which can be helpful, for example, to
simplify wireless transmission protocol. For example, various
MDSP-based embodiments involve compressing ECG signals at
compression rates of 15:1 to 20:1 (e.g., relative to an original
physiological signal prior to denoising) with minimal signal
distortion. An MDSP approach is used to achieve accurate
identification of cardiac cycles of a denoised ECG, and
significantly reduce in-band noise. These identification and
denoising approaches are used in connection with one or more of a
variety of compression algorithms, in accordance with various
embodiments, together with identified redundancies in adjacent
cardiac cycles to achieve efficient compression.
[0066] In some embodiments, an MDF-based approach is used to remove
noise from peripheral nerve activity (PNA) signals and facilitate
the accurate quantification of PNA. In these and/or other
embodiments, an MDSP-based approach is used across a broad range of
physiological signal processing applications with related
processing without necessarily relying upon assumptions regarding
the independence of signal and noise sources, and/or based upon an
assumption (for processing) that physiological signals being
processed are quasi-periodic signals.
[0067] In the following discussion, reference is made to various
documents or other references listed near the end of this patent
document, by numerals within square brackets. The information in
the documents/references to which these numerals refer may be
implemented in connection with one or more example embodiments, and
are fully incorporated herein by reference.
[0068] Various embodiments are directed to processing signals from
ambulatory subjects under conditions in which noise is common and
measurements are often obtained using a limited number of
closely-spaced electrodes, or a single cannulation of a vessel,
such as in the case of blood pressure measurements. Such signals
often include aspects of signals emanating from multiple sources
that are generally interdependent. For example, heart activity in
an atrial chamber usually initiates activity in a ventricular
chamber. Statistically, this situation is characterized by mutual
dependency of signal sources contaminated by noise. This violates a
central assumption of independence of sources that is fundamental
to successful application of many current techniques used to
separate signal source and noise in multisource signals, such as
independent component analysis (ICA) [1]. Various embodiments are
directed to processing related signals, without necessarily
assuming such independence, to obtain a denoised signal for a
desired signal (i.e., a signal corresponding to a characteristic to
be detected). The problem of extracting source signals from the
multisource physiological signals contaminated with noise can be
expressed mathematically as:
x(t)=As(t)+n(t) (1)
where s(t)=(s.sub.1(t), . . . , s.sub.P (t)) is a vector of source
signals that are mixed together by an unknown mixing matrix A of
size M.times.P, where M is a number of observed signals and P is a
number of source signals and additive noise n(t)=(n.sub.1(t), . . .
, n.sub.P (t)). The extraction of source signals is achieved by
estimating the inverse of a mixing matrix A and computing the
denoised and separated signals s(t)=(s.sub.1(t), . . . , s.sub.P
(t)) from the observed signals x(t)=(x.sub.1(t), . . . ,
x.sub.M(t)) according to the formula
s(t)=A.sup.-1x(t)-A.sup.-1n(t). When the number of leads M is less
than the number of sources P, which is most often the case in
physiologic monitoring, the mixing matrix A.sub.M.times.P is
underdetermined. In addition, the sources that make up multisource
physiological signals are often not independent. In this situation
the matrix A.sub.M.times.P is not invertible, rendering it
difficult to estimate an inverse of the M.times.P mixing matrix and
recover unmixed sources from mixed observations, as is the case
when using ICA techniques.
[0069] Turning now to the figures, FIGS. 1 and 2 show signal
denoising and signal source extraction, as carried out in a
three-step process in accordance with another example embodiment of
the present invention. The first step, 102 and 202 in FIGS. 1 and 2
respectively, involves decomposing an input signal 101 and 201 into
subcomponents in a second domain of larger dimension than the first
domain. Decomposition steps 102 and 202 are performed using one of
a variety of transforms that is used for accurate signal
reconstruction. Such transforms may involve using, for example, a
discrete cosine transform, a wavelet related transform, a
Karhunen-Loeve transform, a Fourier transform, a Gabor transform,
or a filter bank. In FIGS. 1 and 2, the set of subcomponents
resulting from the decomposition steps 102 and 202, respectively is
referred to as D2sub. The dimension of the first domain is defined
by the number of observed, or captured, signal channels. The
dimension of the second domain is defined by the number of channels
multiplied by the number of subcomponents in each channel. The
second step involves the identification of signal and noise
subcomponents in the second domain followed by removal of unwanted
subcomponents, including noise.
[0070] Referring to FIG. 1, to denoise a desired signal,
subcomponents of the set D2sub that are primarily associated with
noise are identified and removed to create a subset D2s of residual
denoised subcomponents. Referring to FIG. 2, if signal source
extraction is performed, such as when an atrial activity signal is
extracted from a surface ECG, the subcomponents associated with the
signal sources to be extracted are identified and the other
subcomponents are removed to create a subset D2soi containing
energy associated with the desired signal. In certain embodiments
involving both denoising and signal source extraction,
subcomponents to be removed are identified based upon their spatial
distribution using, for example, Spatially Selective Filtering
(SSF) or Periodic Component Analysis (.pi.CA). Ad decision steps
103 and 203 in FIGS. 1 and 2, the signal is processed at either
104, 204 if the dimension of the first domain is greater than 1, or
at block 105/205 if the dimension of the first domain is not
greater than one.
[0071] If the dimension of the first domain is greater than 1, then
SSF, .pi.CA, Principal Component Analysis (PCA), and Independent
Component Analysis (ICA) can be used in combination (e.g., as in
blocks 104 and 204). If the dimension of the first domain is equal
to 1, then SSF or PCA can be used to evaluate the spatial
distribution of the subcomponents, as in 105 and 205. Once the
subcomponents associated with noise are identified they are
effectively removed by setting them to zero. Following removal of
the subcomponents associated with noise, the subcomponents are said
to be denoised. In the case of signal source extraction, those
subcomponents associated with noise and those not associated with a
signal source to be extracted are identified and effectively
removed by setting them to zero.
[0072] In steps 106 and 206, the denoised signal subcomponents,
referred to as subset D2s, are reconstructed in the first domain to
provide denoised output signal 107 by performing the inverse of the
transform used to decompose the signal. If signal source extraction
is performed only those subcomponents not removed and identified as
associated with the desired signal sources, referred to as
subcomponent subset D2soi, are reconstructed into the extracted
desired signal 207 in FIG. 2.
[0073] Various embodiments directed to MDSP approaches use a
mathematical representation of physiological signals as a linear
combination of basis functions. Each of these basis functions is
chosen to fit a prominent feature, or signal wave, of a signal
source. For example, a basis function might represent a type of QRS
complex contained in an ECG signal. In one embodiment,
decomposition is achieved by representing the observed signals as a
linear combination of basis functions,
s i ( t ) = k = 1 K .PHI. k ( t ) d ik , i = 1 : P , x i ( t ) = k
= 1 K .PHI. k ( t ) c ik , i = 1 : M ( 2 ) ##EQU00001##
[0074] where d.sub.ik and c.sub.ik are decomposition coefficients
or subcomponents. In some embodiments, the elements .phi..sub.k may
be mutually dependent and form an overcomplete set referred to as a
"dictionary" D. The elements are selected to provide a sparse
representation of the signal sources, meaning that most of the
decomposition coefficients d.sub.ik and c.sub.ik in equation (2)
are close to or equal to zero. A combination of basis functions or
elements of the dictionary forms the second domain.
[0075] Higher sparsity provides for greater statistical
independence of signal source and noise subcomponents. This
property of sparse decomposition is used to facilitate effective
identification and extraction of signal source subcomponents and
removal of noise. Sparse decomposition, for example, can result in
a concentration of signal energy in a relatively small number of
large decomposition coefficients while noise is spread out across
many basis functions and represented by small decomposition
coefficients as described in equation (2).
[0076] The sparsity criteria can be used to choose a dictionary
that provides for sparse decomposition of the signal sources and
hence improved signal source extraction and denoising. In certain
embodiments it may be advantageous to choose an over-complete
dictionary D in order to increase the degree of sparsity of the
decomposition. The examples of dictionaries that are relevant to
this embodiment include eigenfunctions [2], wavelet related
transforms, including wavelet packets and shift-invariant or
stationary wavelets [3], cosine transform [4], Fourier basis [5] or
Gabor basis and their combinations [6]. A custom dictionary that
achieves a higher degree of sparsity can also be designed by a
number of methods for finding basis functions that represent
building blocks of the signal [7]. The examples of these methods
are matching pursuit [5], orthogonal matching pursuit [8], basis
pursuit [9], K-SVD [10], bounded error subset selection [11] or
orthogonal subspace pursuit. In general, the dictionary design
methods use a representative set of training signals, to find a
dictionary that leads to the best representation for each signal
wave in this set, under strict sparsity constraints:
min D , c { x - Dc F 2 } subject to c 0 .ltoreq. T ##EQU00002##
where .parallel..parallel..sub.0 is a l.sup.0 norm, counting the
nonzero entries of a vector, T is a constraint on the maximum
number of non-zero elements allowed, and
.parallel..parallel..sub.F.sup.2 is a Frobenius norm. For example,
a representative set of training signals can be a set of ECG
recordings with a broad range of morphologies that include normal
sinus rhythm, conduction abnormalities, pathologies, arrhythmias
for a particular species. The set should include a sample of each
morphology that the algorithm may experience during analysis of an
ECG recording of the species. In one embodiment the custom
dictionary is build using the following iterative procedure. A
vector is selected from the training set and is represented as a
sparse linear combination of initial dictionary elements. This can
be accomplished, for example, by finding elements that are
maximally correlated with the vector. This subset of elements form
a subspace that is decorrelated from the rest of the training set
(e.g., by using singular value decomposition (SVD)). The
decorrelated subset is added to the dictionary and the elements of
the training set that lie in this subspace are removed from the
training set. The process is repeated until an error threshold is
reached or an allowed number of iterations is exhausted.
[0077] In another embodiment decomposition into the second domain
is achieved by applying an adaptive filter that automatically
selects the frequency bands, or subbands of wavelet filters or a
filter bank, in which subcomponents of the signal sources are most
independent. Then each subcomponent x.sub.i(t),i=1: M, can be
represented as a sum of subband signals
i(t)=x.sub.i,1(t)+ . . . +x.sub.i,K(t) (3)
where K is the number of chosen subbands. The filter coefficients
can be adaptively tuned to reduce and/or minimize the mutual
information and increase independence between the filter outputs as
described by Zhang, et al. [12]
[0078] In an embodiment involving a wavelet-related transform
decomposition, the first decomposition step for an observed signal
with M channels into a wavelet-related dictionary D can be formally
described as iterative multiplication of the observed signal by an
operator matrix W.sup.j with each iteration step j making up a
decomposition level. In this embodiment, the operator matrix
W.sup.j is a Toeplitz matrix with the entries formed by a wavelet
impulse response. The size of the matrix operator W.sup.j is
N.times.N where N is the number of samples of the observed signal.
N is greater than both the number of sources P and the number of
observed signals M. Wavelets may be used to generate subcomponents
having characteristics of both sparse and subband decomposition. In
some implementations, a wavelet decomposition is computed via
iterative convolution with wavelet filters or filter banks as
described by Vaidyanathan [13].
[0079] Referring again to FIGS. 1 and 2, another example embodiment
is implemented at step 2, for an MDSP-based approach. The
statistical independence of at least some subcomponents related to
signal sources and noise is leveraged to identify those
subcomponents associated with signal sources and those associated
with noise. For example, in case of wavelet decomposition, the
equation (1) becomes
W.sub.x.sup.j=A.sub.jW.sub.s.sup.j+W.sub.n.sup.j, j=1:NL+1 (4)
where NL is the number of decomposition levels or subbands. The
mixing matrix A is indexed by j, denoting that the sources are
mixed differently in each subband according to their spectral
distribution.
[0080] The subcomponents associated with signal sources can be
identified based on spatial distribution of subcomponents in time
using one of several techniques, many of which involve finding the
mixing matrices A.sub.j and then inverting them to calculate the
signal source subcomponents. Examples of techniques that are
applicable to identification of subcomponents associated with noise
and subcomponents associated with signal sources include principal
component analysis (PCA), independent component analysis (ICA),
periodic component analysis (.pi.CA) techniques and spatially
selective filtering as shown in blocks 104 and 105 of FIG. 1 and
blocks 204 and 205 of FIG. 2. These techniques can be used
separately or in combination to achieve acceptable performance.
[0081] When using the PCA technique to identify subcomponents
associated with signal sources, an inverse of the mixing matrix
A.sub.j is estimated [14, 15] using singular value decomposition or
eigenvalue decomposition of the covariance matrix of signal
subcomponents. This PCA approach involves rotating and scaling the
data in order to orthogonalize independent subcomponents. The
orthogonalized subcomponents with low signal power are often
associated with noise and can be removed to achieve denoising. PCA
can be used as a preliminary step prior to applying ICA
technique.
[0082] In various embodiments involving the use of an ICA technique
to identify subcomponents associated with signal sources, an
inverse of the mixing matrix A.sub.j is estimated [4,16] as a
solution of an optimization problem that maximizes independence
between the signal sources. For example, ICA techniques can use
either higher-order statistics of signal components [17, 18] or
information-theoretic criteria to maximize independence.
Information theoretic criteria that can be applied include
maximization of negentropy or its approximation [19], minimization
of mutual information [19], maximum likelihood estimation [20, 21],
maximum a posteriori probability [22], or expectation-maximization
of Gaussian mixture models of sources [23]. These solutions can be
approximated via efficient numerical methods, such as FastICA [24]
and JADE [19] algorithms. The FIG. 3a provides a scatter plot image
of subcomponents in the second domain prior to the application of
PCA and ICA in the denoising process. FIG. 3b demonstrates an
example result of denoising and subcomponent separation achieved by
applying a combination of PCA and ICA techniques. Note that
application of PCA and ICA has achieved a decorrelation of the
subcomponents and alignment along the coordinate axis 302 of the
scatter plot, indicating a large degree of independence has been
achieved between the subcomponents (e.g., relative to alignment in
with coordinate axis 301 FIG. 3a). This high degree of independence
between subcomponents that occurs as a result of the denoising
process facilitates efficient removal of noise and signal source
extraction.
[0083] In embodiments involving the use of .pi.CA technique to
identify signal sources, an inverse of the mixing matrix A.sub.j is
estimated as a solution of an optimization problem that separates
subcomponents based on their periodicity or quasi-periodicity [25,
26]. Instead of diagonalizing and inverting the covariance matrix
as is done with PCA, the .pi.CA technique jointly diagonalizes the
covariance matrix and an autocorrelation matrix. The
autocorrelation matrix is calculated as an average of
autocorrelation matrices computed over time lags corresponding to
period lengths.
[0084] In one embodiment, a quasi-periodic signal can be
phase-wrapped by mapping the period length to a linear phase
.phi.(t) ranging from -.pi. to +.pi. assigned to each sample. The
autocorrelation matrix can then be calculated in the polar
coordinates in which cardiac cycles are phase aligned. The
technique involves extracting most pseudo-periodic subcomponents
corresponding to a desired physiologic signal. The technique is
efficient at identifying signal source and noise subcomponents but
relies on accurate detection of cycles such as cardiac or
respiratory cycles. It can be used in combination with spatially
selective filtering (SSF), a technique that facilitates better
cycle detection.
[0085] In another embodiment, referring to FIG. 2, signal sources
are identified and separated in the second step, blocks 204 and 205
as discussed above using spatially selective filtering. Spatially
selective filtering [27, 28, 29, 30] techniques detect
signal-related features and pass them across the subcomponents
while blocking features inherent to noise. The technique relies on
the differences of noise and signal distributions between
subcomponents. In one embodiment, spatially selective filtering is
facilitated by a decomposition whereby signal energy is
concentrated in a small number of large subcomponent coefficients
while noise is spread out across many decomposition levels and is
represented by small coefficients. Techniques similar to wavelet
thresholding [31] can be used to remove this noise. This may result
in a slight degradation of signal morphology.
[0086] In another embodiment, spatially selective filtering is used
to exploit the fact that most noise subcomponents are confined to
decomposition levels that represent high frequencies. In this
embodiment the locations of signal features are identified by
examining subcomponents corresponding to lower frequency or
correlation between subcomponents. For example, QRS wave location
can be identified as high amplitude peaks and valleys that occur
simultaneously across multiple subcomponents associated with lower
frequencies. The rest of the ECG waves (P, T, U) have lower
frequency content.
[0087] Referring to FIG. 20, a cardiac cycle of an ECG is
partitioned into time windows 2002 and 2003: 2002 containing the
QRS complex and 2003 spanning the remainder of the cardiac cycle.
The subcomponents associated with both low and high frequencies,
represented by region 2004, are preserved within the time window
2002 containing the identified QRS complex. The low-frequency
subcomponents associated with the desired signal, present in window
2003 and represented by region 2005, are preserved and
subcomponents associated with high frequencies are removed. In
another embodiment, subcomponents associated with low-frequencies,
represented by region 2007, are preserved throughout the cardiac
cycle and subcomponents associated with high frequencies are
preserved in a time window represented by region 2006. In another
embodiment, in signals with low amplitude subcomponents that are
associated with high frequency such as PNA, high amplitude EMG
noise can be detected as large peaks present in subcomponents
corresponding to high frequency. The EMG noise can then be removed
by zeroing subcomponents in the time window where the large peaks
were detected. In another embodiment, the high amplitude noise or
artifact is detected by identifying undesired peaks in
subcomponents that correspond to frequencies that do not overlap
the band of frequencies associated with the desired signal. For
example, a blood pressure signal can be denoised using this
approach.
[0088] In some embodiments the spatially selective filtering is
combined with one or more of PCA, ICA, or .pi.CA. For example, the
subcomponents that were identified as noise at the PCA stage can be
further filtered to remove only some segments that have spatial and
temporal characteristics of noise. In another example, the
subcomponents that were identified as signals at the PCA stage can
be further filtered to remove additional time segments from
subcomponents that have spatial and temporal characteristics of
noise.
[0089] The choice of a technique for identification of
subcomponents for denoising or signal source extraction depends
upon signal and noise characteristics of the physiologic signal,
spatial distribution of its subcomponents and the signal
acquisition and processing environment. For example, it may be
useful to process multi-channel signals with PCA and ICA techniques
and their combinations with spatially selective filtering. When
processing ECG signals obtained from an implanted single lead
device spatially selective filtering or .pi.CA may be applied, as
in blocks 105 and 205 in FIGS. 1 and 2, respectively.
[0090] In some embodiments, step three as shown in blocks 106 and
206 of FIGS. 1 and 2, respectively, involves reconstructing the
subcomponents remaining following denoising and signal source
extraction, s(t)=(s.sub.1(t), . . . , s.sub.P(t)) in the first
domain. This is accomplished by an inverse transform of extracted
signal sources or denoised signal sources. If wavelet-based
decomposition is employed, the reconstruction step can be formally
described as iterative multiplication by an inverse of the operator
matrix W.sup.j with each iteration step j corresponding to a
reconstruction level.
[0091] Referring to FIG. 4, other example embodiments involve
computing a dynamic signal-to-noise ratio (dSNR) in a manner that
can be updated on a sample-by-sample basis. dSNR can be used to
assess the accuracy and reliability of information derived the
physiological signal and, in some embodiments, can be combined with
other signal information such as ECG arrhythmic event information
to compute a validity metric to assess the validity of information
extracted from a signal. If dSNR for a cardiac cycle is low, for
example, information derived from the ECG for that cardiac cycle
may not be accurate.
[0092] A dynamic signal-to-noise ratio (dSNR) can be computed as
the ratio of the energies in signal and noise subcomponents. In one
embodiment, referring to FIG. 4, an input signal 401 in a first
domain is decomposed in process 402 into subcomponents, referred to
as set D2sub, in a second domain of higher dimension. At decision
point 403, the signal processing flow proceeds according to the
dimension of the first domain. If the dimension of the first domain
equals 1, then subcomponents primarily associated with noise,
referred to as set D2n, are identified in process 405 using SSF or
.pi.CA. Subcomponents of set D2n are combined to compute an
estimate of noise energy and residual subcomponents, referred to as
set D2s, are combined to compute an estimate of signal energy in
process 407. The SNR is then computed in process 409 to create
output dSNR value 410 according to formula:
SNR dB = 10 log 10 ( P signal P noise ) = 20 log 10 ( A signal A
noise ) ##EQU00003##
Where P.sub.signal and P.sub.noise are respective signal and noise
energy and A.sub.signal and A.sub.noise are respective signal and
noise amplitude.
[0093] If the dimension of the first domain is larger than one, a
PCA or ICA technique can be used in process 404 for initial
subcomponent denoising. The noise subcomponents extracted at this
initial denoising step can be discarded and the residual noise and
signal subcomponents can be identified using SSF or .pi.CA in
process 406. In process 408, the noise subcomponents identified at
this stage, D2n, are used to calculate an estimate of noise energy
and the residual subcomponents, D2s, are used to compute an
estimate of signal energy. The SNR is then computed in process 409
(e.g., as described above).
[0094] In other embodiments, a signal-to-noise ratio is estimated
using conventional approaches following denoising of the signal
using an MDSP embodiment. In one embodiment, the noise is measured
between signal waves by computing the peak amplitude and density of
zero crossings or variability of signal in a segment between signal
waves. In other embodiments, a signal-to-noise ratio is estimated
by computing a spectral distribution of the denoised signal. For
example, in an ECG signal, peaks in the spectral distribution are
evaluated to determine the relative power in the spectrum that
occurs within and outside of the normal range of the QRS complex,
T-wave, and P-wave.
[0095] In one embodiment dSNR is updated for each cardiac cycle.
Alternative implementations may update dSNR more or less often. For
example, it may be useful in some embodiments to compute a value of
dSNR for a window of two to ten cardiac cycles and use this value
in calculation of the validity metric for all cardiac cycles within
the window.
[0096] FIG. 5 provides an example of a dynamic confidence signal
(dCS) derived from dSNR computed for a rabbit ECG signal 501, with
dCS shown in the dashed line 505 in the top plot, and the bottom
plot showing automatic marking of QRS complexes (squares as 504)
and detected VT (circles as 503), in accordance with another
example embodiment. The dynamic confidence signal 505 (dashed) is
updated for each cardiac cycle. The dCS is updated every cardiac
cycle and the value of dCS reflects changes in signal amplitude
relative to the level of noise in the recording. Despite the low
signal-to-noise ratio, all QRS complexes (as indicated by the
squares 504 marking bottom ECG trace) and non-sustained episodes of
ventricular tachycardia (marked by circles 503 marking bottom ECG
trace) are detected.
[0097] In various embodiments, MDSP-based approaches as discussed
herein are used to denoise and extract information from a
physiologic signal acquired in a low SNR environment. For example,
a collar or spring-loaded clip placed on the neck of an animal with
embedded ECG electrodes can be used to collect ECG signals in
animals non-invasively. While such an approach may result in a low
SNR, denoising approaches such as discussed herein can be used to
glean meaningful data from the ECG signals. Another embodiment is
directed to placing a collar or spring-loaded clip on the neck of
an animal, or a collar placed around the tail or limb, which can
house light emitting diode transmitters and light receiving sensors
to collect photoplesythmography signals. In one embodiment, the
collar or clip includes two pairs of light transmitters and
receivers placed at different locations on the neck of the animal
in order to achieve redundancy and improved noise suppression. In
another embodiment a cage floor having ECG sensors is used to
collect ECG signals from the animal feet. These applications are
exemplary of many applications characterized by low SNR, for which
denoising approaches as discussed can be used to render the
applications viable for analysis.
[0098] In the following examples, performance aspects are discussed
as may be relevant to various embodiments is illustrated and
analyzed, many of which involve the analysis of ECG and PNA signals
that can be challenging to carry out using other approaches. In
many of the examples discussed here, the signal sources are
contaminated with noise and the number of sources is larger than
the number of observed signals. Although embodiments of the present
invention can be used with a broad range of physiological signals,
exemplary performance is illustrated with ECG and PNA signals that
are problematic for traditional signal processing approaches
because of contamination with in-band noise that could not be
substantially reduced without distorting signal morphology. Such
embodiments are applicable to the use of MDSP-based approaches as
discussed herein, to achieve certain performance-related conditions
or characteristics, which can be measured or relatively
characterized in manners as discussed herein. Accordingly, various
embodiments are directed to approaches to achieving such things as
levels or degrees of noise reduction, signal quality, cardiac (or
other signal) event detection accuracy and more, as facilitated
using aspects of the invention as described herein.
[0099] Various MDSP-based embodiments are directed to denoising a
physiological signal while preserving the morphology of a desired
signal within the physiological signal (e.g., as recorded). Quality
of signal reconstruction (QSR) is a metric commonly employed to
assess the ability of a noise removal technique to preserve
morphology of the desired signal in signals with moderate to high
signal to noise ratio. QSR is defined as the mean squared error
between the original signal x.sub.cl and denoised signal x.sub.den
calculated sample-by-sample as a percentage of the original signal
variance.
QSR = 100 % * ( 1 - i ( x cl i - x den i ) 2 i ( x cl i ) 2 )
##EQU00004##
[0100] QSR of a denoised signal would be close to 100% if the
original recording is contaminated with minimal noise and if
distortion introduced into the desired signal by denoising was very
small.
[0101] FIG. 6 shows plots characterizing an example MDF-based
embodiment directed to removing noise from input 3-lead ECG signals
601, 602, and 603 from a PhysioNet [32] database. Results of ICA
and PCA applied directly to the signal are also shown for
comparison. The ECG recordings of FIG. 6 represent signals with
portions that are relatively noise free (601) and portions that are
noisy (602 and 603), demonstrating the removal of noise with an
MDF-type approach as discussed herein while also preserving
morphology. In such a short recording, the character of the
denoising technique, whether MDF, ICA, or PCA, is relatively
consistent throughout the duration. Evaluating QSR for the portion
which is relatively noise free, therefore, provides an indication
of a technique's ability to preserve morphology while a visual
inspection of the noisy portion provides for a qualitative
assessment of the ability of a particular technique to suppress
noise. For the relatively noise free portion, the MDF-based
approach shows QSR values of 98%, 99%, and 92% for the top 610,
middle 611, and bottom 612 traces, respectively, indicating that
distortion was negligible. The PCA approach shows QSR values of
83%, 68%, and 71%, for the top 604, middle 605, and bottom 606
traces, respectively, indicating significant distortion of signal
content. The ICA approach shows QSR values of 2%, 1%, and 4%, for
the top 607, middle 608, and bottom 609 traces, respectively,
indicating very significant distortion of signal content.
Accordingly, various MDSP-based embodiments are directed to
processing signals to address challenges relating to such QSR
values that may be insufficient, if processed otherwise.
[0102] FIG. 7 shows characteristic results of denoising using an
MDF-type approach (in accordance with an example/experimental-type
embodiment), Butterworth bandpass filtering (BPF) with a pass-band
of 1 to 60 Hz, and PCA. A relatively noise-free ECG from the
PhysioNet Long-Term ST database (record s30661) is corrupted with
increasing levels of band-limited (0.05 to 70 Hz) white noise, and
processed in accordance with an example embodiment. The denoising
results are quantified measuring QSR and input and output signal
SNR according to the formula:
SNR dB = 20 log 10 ( .sigma. signal .sigma. noise )
##EQU00005##
where .sigma..sub.signal and .sigma..sub.noise are respective
clean-signal and noise standard deviations.
[0103] Referring to the plots 710 and 711 of FIG. 7, SNR and QSR
versus input signal SNR achieved by MDF, PCA, and BPF are
illustrated. Plot 710 illustrates the SNR improvement by comparing
input SNR (on x-axis) to denoised signal SNR (on y-axis). Plot 711
illustrates the corresponding QSR for the same range of input SNR.
For example, as illustrated in plot 710, an input signal with 4 dB
SNR is denoised with an MDF-based approach, a 9 dB SNR improvement
is achieved. Referring to the plot 711, 95% of the original clean
signal content is preserved following MDF denoising of an input
signal with 4 dB SNR. With increasing input SNR, QSR performance
for the MDF-based embodiment quickly approaches 100% with an
approximately linear denoising characteristic as measured by SNR on
a logarithmic scale. Similar processing with a two-lead ECG
generates similar results observed for an MDF-based embodiment. As
shown in the FIG. 7 the PCA and BPF results can be improved upon
using various embodiments as discussed herein.
[0104] Referring to the ECG tracings 707, 708, and 709 of FIG. 7,
these three tracings illustrate an input noise-free signal 707;
signal 708 corrupted with band-limited white noise to reduce SNR to
4 dB, and signal 709 denoised with an MDF-based approach. Adding
band-limited white noise to achieve a 4 dB SNR renders the T and P
waves indiscernible, as indicated in the middle tracing, while
application of MDF restores the P and T waves.
[0105] FIG. 8 shows another example embodiment, involving the use
of MDF to suppress noise while preserving morphology for a
single-channel ECG. Plot 801 is an input ECG recording corrupted
with noise, and plot 802 shows the result of applying MDF
filtering. The noise present in the input signal has similar
characteristics (frequency content and amplitude) as QRS complexes
which would result in QRS detection errors without denoising. The
application of MDF is used to avoid potential QRS detection errors,
and can remove most of the noise while preserving morphology of
QRS, P, and T waves, allowing for high accuracy detection of all
ECG features, including QRS complex, P, and T-waves.
[0106] The embodiments shown in FIGS. 6, 7, and 8 characterize an
example use of MDF to suppress in-band noise while substantially
preserving signal morphology, in accordance with various
embodiments. This combination of attributes is useful for a
multitude of applications, including clinical diagnosis and
research involving the measurement and analysis of physiological
signals. Waveform morphology is preserved for a variety of
applications, such as for detecting abnormalities in cardiac
function from an ECG, evaluating respiratory function from a
respiratory signal, evaluating a signal from a photoplethysmography
sensor for measuring oxygen saturation, evaluating sleep stages
from an EEG, and other applications. For example, preserved QRS
morphology is used in diagnosing bundle brunch block or ventricular
hypertrophy, T-wave morphology is used in measuring repolarization
abnormalities in clinical care and drug toxicity studies, and ST
segment changes are used when diagnosing ischemia, electrolyte
imbalance, and Brugada syndrome. In another example, the ability to
preserve P wave and QRS complex morphology facilitates the analysis
of time correspondence of P wave and QRS complex to diagnose AV
block.
[0107] In another embodiment, an MDSP-based approach is used for
denoising and signal source extraction for monitoring a fetal ECG.
In this embodiment, an ECG is recorded by placing sensing leads on
the surface of the skin of the mother, typically in the lower
abdomen.
[0108] In one embodiment, individual sources that make up a fetal
ECG are separated from the remaining subcomponents in the second
domain by applying ICA signal source extraction techniques in step
2 of the MDSP embodiment. In another embodiment the fetal ECG is
extracted from the other subcomponents by spatially selective
filtering, periodic component analysis, or their combination (e.g.,
in step 2 as shows in FIGS. 1 and 2 above). In this embodiment, the
undesired sources, such as maternal ECG, are treated as noise and
removed, leaving the denoised fetal ECG.
[0109] In another embodiment, an MDSP-based technique is used for
measuring a degree of synchronization of uterine contractions of a
pregnant female for predicting and detecting labor. In this
embodiment, an electrohysterogram (EHG) is collected from the
female abdomen. In one embodiment, an MDF-based approach as
discussed herein is used to remove ECG artifacts from the EHG
signal. In one embodiment, the degree of synchronization of
contractions in denoised EHG signal is measured by amplitude
correlation via linear or nonlinear regression and the frequency
relationship measured as coherence, or the amplitude and phase
synchronization in the time-frequency domain. In another
embodiment, a cross wavelet coherence function is used to measure
amplitude correlation between two contraction bursts [33, 34]. In
another embodiment, an envelope of a multi-channel EHG signal is
calculated using a Hilbert transform or low-pass filtering of a
rectified EHG signal to measure amplitude of a contraction wave and
its spatial synchronization.
[0110] In other embodiments, a MDSP approach as discussed herein is
used for detecting atrial arrhythmias in an ECG. In these
embodiments, atrial activity can be extracted by applying ICA, SSF,
PCA, .pi.CA, or their combination, as part of step 2 (process 204
and 205) as shown in FIG. 2. In this embodiment, step 2 can involve
removing noise as well as the subcomponents associated with
ventricular activity or segments of the subcomponents as identified
by SSF.
[0111] In FIG. 9, the results of atrial activity extraction are
illustrated on an atrial flutter recording from the PhysioNet
database [32]. Plots 901, 902, and 903 show ECG recorded from
surface leads. Plot 904 is a recording from an intracardiac
catheter located near an atrial free wall. The intracardiac
recording is shown as an exemplary benchmark of atrial activity.
Plot 905 shows the atrial activity separated from the surface ECG
recordings using an MDSP-based embodiment. Note that the locations
of P-waves on the plot 905 of extracted atrial activity matches
locations of atrial depolarizations recorded from intracardiac lead
shown in 904. The gaps in atrial activity coincide with ventricular
depolarizations (i.e., QRS complexes) and repolarizations
(T-waves). Analysis of the separated atrial activity can
significantly improve the accuracy of atrial flutter detection.
[0112] In FIG. 10, the results of atrial activity extraction are
illustrated on an atrial fibrillation recording from the PhysioNet
database [32]. During atrial fibrillation the P-waves often
degenerate into more fractionated and variable f-waves which are
barely visible in surface recording traces 1001, 1002, and 1003
shown in FIG. 10. Despite that, the atrial activity signal 1005
extracted using an MDSP embodiment clearly shows the f-waves that
coincide with the atrial depolarization trace 1004 recorded using
an intracardiac lead. The only gaps are when atrial activity
coincides with ventricular depolarizations (i.e., QRS complexes)
and repolarizations (T-waves).
[0113] FIGS. 9 and 10 demonstrate that various MDSP-based
embodiments are capable of extracting atrial activity from ECG
recordings. There are a number of ways that extracted atrial
activity can be used to improve the accuracy of atrial flutter and
fibrillation detection and discrimination. For example, spectral
analysis of the extracted atrial activity signal can be used to
quantify P-wave regularity and frequency and discriminate between
atrial fibrillation and flutter. In one embodiment, atrial rate can
be estimated by spectral analysis. In another embodiment, extracted
subcomponents corresponding to atrial activity can be analyzed in
the second domain to estimate the atrial rate. In another
embodiment, a P-wave similarity measure is used to detect and
discriminate between atrial fibrillation and flutter. In another
embodiment, zero crossings can be used to estimate atrial rate in
the segments where atrial activity is present. The analysis of
atrial activity can be combined with analysis of cardiac cycle
variability and regularity as well as PR interval variability to
further improve the accuracy of atrial fibrillation and flutter
detection.
[0114] In other embodiments an MDSP-based approach is used for
separating ventricular depolarization and repolarization activity.
This may be useful for assessing repolarization activity for the
risk stratification of sudden cardiac death. Once repolarization
activity is extracted, it can be used to assess T-wave alternans
and morphology, ST elevation in ischemia, and other cardiac
abnormalities that can be useful for assessing cardiovascular
risk.
[0115] Another embodiment is directed to using MDF for removing
high amplitude noise from PNA signals. In particular, this is
useful when removing high amplitude noise from PNA recordings such
as those from the vagal nerve, sympathetic nerves, or motor nerves.
In this embodiment, the high amplitude noise or artifact is
detected by identifying undesired peaks in subcomponents that do
not overlap the band of frequencies associated with desired signal.
In FIG. 11 exemplary performance of an MDF embodiment is
illustrated on a single channel PNA input signal 1101 with high
amplitude EMG noise. Plot 1101 is a PNA recording from a rat renal
nerve while plot 1102 is the denoised signal following application
of MDF. The MDF embodiment extracts the signal source from a very
low SNR acquired signal while preserving PNA information. It
removes nearly all noise and artifacts from the recording including
myoelectric artifacts (EMG), electrical, and other noises while
preserving information regarding the underlying PNA. An MDF
embodiment could be used to extend the life of preparations where
PNA is recorded by improving the ability to extract accurate
information from PNA signals with low SNR. In PNA recording
preparations for research in animal models, it is also common to
dose the subject with a drug that shuts down PNA so that baseline
noise can be measured and subsequently subtracted from the neural
signal. This procedure can be dangerous for the subject and is
labor intensive and cumbersome for the researcher. The MDF
embodiment removes background noise automatically and eliminates
the need for such an intervention.
[0116] In connection with another example, embodiment, PNA is
quantified from recordings such as discussed above, via the
computation of integrated sympathetic nerve activity. In one
embodiment for computing a PNA envelope, an orthogonal component of
the denoised PNA signal is computed using a transform such as a
Hilbert transform, or similar transform. An envelope is computed as
the square root of the sum of the squared denoised PNA signal and
its orthogonal denoised component. This provides an accurate
representation of neural activity without the phase delays inherent
in conventional approaches. The neural activity represented by the
signal envelope has a much lower frequency content compared to the
raw PNA signal and thus substantially reduces the bandwidth and
sampling rate requirement of a system for measuring PNA
activity.
[0117] Monitoring devices that transmit raw PNA signals as may be
used in accordance with this or other embodiments, such as that
available from Telemetry Research, Auckland, NZ, employs a sampling
rate of about 8,000 Hz. By employing this embodiment to calculate
an accurate PNA envelope, the sampling rate can be reduced to 100
Hz or less, resulting in a reduction of transmitted bandwidth of a
factor of 80 and a reduction in current drain of a wireless
transmitter used to transmit data from an ambulatory subject.
[0118] In another embodiment, an MDSP-based approach is used for
removing noise and extracting signal sources from signals acquired
as a result of programmed periodic stimulation, such as auditory
brainstem response and in peripheral nerve stimulation therapies
where evoked response is analyzed to titrate therapy or peripheral
nerve recruitment. These signals are often characterized by low SNR
and a limited number of observed channels. These signals can be
segmented in the time domain based upon knowledge of timing of
stimulation. Segmentation in the time domain allows for the
creation of the equivalent of multiple channels from the observed
signal, hence increasing the number of dimensions in the first
domain. Following decomposition into the second domain of larger
dimension than the first domain, one or more MDSP-based embodiments
previously described can be applied for denoising and signal source
extraction. For auditory brainstem response, an MDSP-based
embodiment can be used to improve the accuracy of intracranial
pressure estimation, such as with a system similar to that
described in US Patent Application Publication 20080200832, and
U.S. Pat. No. 6,589,189, which are fully incorporated herein by
reference.
[0119] In the case of a peripheral nerve stimulation therapy,
evoked responses can be analyzed to titrate therapy targeted at
specific nerve fibers or monitor neuropathy progression. For
example, in vagal nerve stimulation applied for cardiovascular
therapy, the stimulation protocols are optimized to selectively
recruit efferent smaller fibers that control heart function and
block stimulation of efferent larger fibers and afferent fibers
that could invoke pain or coughing reflex. Such approaches may be
implemented in accordance with that described in U.S. Patent
Application Publication 20080065158, which is incorporated herein
by reference. A device employed for neural stimulation may
incorporate a feedback control of stimulation by observing
parameters of an evoked response. In this type of neural
stimulation device, an MDSP-based embodiment could be applied to
assess evoked response resulting from nerve stimulation targeted at
a specific nerve fiber type. In some implementations, an MDF-based
embodiment is used to achieve an accurate assessment of the
response of particular nerve fibers to changes in programmed
stimulation parameters such as timing, frequency, pulse width,
pulse repetition, duty cycle and amplitude of stimulation in order
to appropriately affect therapy.
[0120] In another embodiment, an MDSP-based embodiment is used to
monitor neuropathy progression by measuring changes in nerve
conduction velocity of small diameter axons. The changes in these
axons can serve as an earlier marker of neuropathy development. An
MDSP-based embodiment that involves segmenting the evoked response
signal and denoising as described above is used to remove
background noise and for measurement of evoked response amplitude
and time. In many embodiments, this approach is used to facilitate
the detection of changes in axons that can otherwise be challenging
to detect due to low amplitude and phase cancellation of evoked
response potentials.
[0121] In another embodiment, MDSP is used for detecting feature
points of a physiological signal such as a QRS onset, P-wave onset,
or T-wave offset in an ECG signal or systole and diastole in an
arterial blood pressure signal. Referring to FIG. 12, this feature
point detection embodiment uses one or more selected denoised
signal subcomponents to compute an emphasis signal that emphasizes
a signal wave or feature point of interest, computed as a linear or
non-linear combination of selected subcomponents that are
associated with the signal wave of interest. In another embodiment,
the emphasis signal is computed by performing the inverse of the
transform used for decomposing the physiological signal on the
selected subcomponents. In some embodiments it may be useful to
differentiate the signal following the inverse transform.
[0122] Referring to FIG. 12, subcomponents used to compute the
emphasis signal are generated as a result of the decomposition
process 1205 of input signal 1200 using an MDSP embodiment, as
described herein. The subcomponents that contain the majority of
the energy of the signal waves are selected for the emphasis signal
and are denoised in process 1206 using an MDSP embodiment. The
emphasis signal computed in process 1207 may include frequency
subbands matching the spectral energy of particular signal waves of
interest or a subset of basis functions tuned to the signal wave or
feature point of interest and its variations across a range of
normal, perturbed, and pathological conditions. In another
embodiment wavelet related subcomponents can be used to compute the
emphasis signal in process 1207. The specific subcomponents used
depend upon the decomposition technique used, sampling rate, and
the species from which the physiologic signal was recorded. Example
emphasis signals of an ECG are shown in 1202. 1204, 1211, and 1213
of FIG. 12, along with corresponding ECG waveforms 1201, 1203,
1210, and 1212. The vertical dashed lines in FIG. 12 show the point
of feature point detection in each ECG waveform and corresponding
emphasis signal.
[0123] The transition points of the emphasis signal are evaluated
in process 1208 to detect feature points, shown by example in 1214
and 1215, of the physiologic signal to create output 1209
representing the time of the feature point. In one embodiment, the
pattern of significant peaks, valleys, and zero crossings within
the emphasis signal are used to detect feature points. In another
embodiment, the feature points are detected by applying a threshold
to the emphasis signal. In yet another embodiment, the feature
points are detected by applying pattern or template matching.
[0124] The following illustrative example demonstrates exemplary
performance of an embodiment involving the detection QRS complexes
in an ECG. For QRS detection, an MDSP-based approach is used to
compute an emphasis signal from a combination of denoised
subcomponents that are algorithmically selected to include, for
example, complexes that are wide, narrow, premature, fractionated,
biphasic, monophasic, fibrillatory, tachycardic, or complexes that
have been distorted as a result of a pharmacological agent.
[0125] FIG. 13 provides an illustration of an ECG signal 1301 that
is processed in accordance with such an approach, to address
problems that may relate to the presence of bigemeny, tall T-waves,
and low QRS amplitude, in addition to rapid changes in QRS
amplitude. An emphasis signal computed as described in this MDSP
embodiment facilitates the detection of all QRS complexes, as shown
by the circles as in 1302 of FIG. 13, with no false detection on
T-waves. When tested on the MIT BIH Arrhythmia database [35], a QRS
detection accuracy of 99.8% sensitivity and 99.8% positive
predictive value can be achieved on single lead ECGs. Feature
detection can be enhanced by using and analyzing multiple lead
recordings, using an MDSP-based approach to leverage redundancies
between the leads in a manner that is more efficient at removing
noise.
[0126] In another example embodiment, an MDSP approach is used for
extracting a respiration signal from an ECG, blood flow,
photoplethysmosgraphy, thoracic impedance, or an arterial blood
pressure signal. FIG. 14 shows one such implementation, in which a
respiration signal is extracted using one or more selected denoised
signal subcomponents to compute an emphasis signal that is
associated with a respiratory pattern. In some embodiments, the
subcomponents are filtered with a low-pass filter to extract the
low-frequency respiration signal.
[0127] In other embodiments, the emphasis signal is combined with a
heart rate signal to improve the accuracy of computed respiratory
parameters. For example, canines have pronounced respiratory sinus
arrhythmia which is characterized by heart rate changes that
correlate to respiration; these characteristics can be used in
connection with these embodiments, for extracting, processing or
otherwise using canine respiration signals.
[0128] Referring again to FIG. 14, subcomponents resulting from
decomposition in process 1402 of input signal 1401 are used to
compute an emphasis signal, using an MDSP approach such as
described in connection with one or more example embodiments
herein. Subcomponents are denoised in process 1403 using an MDSP
embodiment as described herein. Subcomponents that contain a
majority of the energy of the respiration pattern are selected and
combined to create an respiratory emphasis signal in process 1404,
also using an MDSP embodiment as discussed herein. The emphasis
signal may include frequency subbands matching the spectral energy
of particular signal waves of interest or a subset of basis
functions tuned to the signal wave or feature point of interest and
its variations across a range of normal, perturbed, and
pathological conditions. The emphasis signal is processed in 1405
using zero crossings or spectral analysis to compute respiratory
parameters.
[0129] In another embodiment (e.g., relative to FIG. 14), wavelet
related subcomponents are used to compute an emphasis signal. The
specific subcomponents that are used are selected relative to one
or more of the decomposition technique used, sampling rate, and the
species from which the physiologic signal was recorded.
[0130] In some embodiments, the noise level of an ECG is measured
using an embodiment described above, and zero crossings that occur
too frequently during noisy segments are discarded (or simply not
used). In another embodiment, the emphasis signal can be low-pass
filtered prior to measurement of respiration rate and tidal volume.
The tidal volume can be extracted by measuring peak and valley
amplitude between valid zero crossings. In another embodiment, the
tidal volume is computed as a function of area under the curve
between valid zero crossings.
[0131] FIG. 15 illustrates another example embodiment as directed
to the processing of a primate ECG. Plot 1501 is a noisy input ECG,
plot 1502 is an ECG denoised with a MDF-based approach as described
herein, and plot 1503 is an extracted respiration emphasis signal.
The residual noise level is measured and is shown on the bottom
plot as a dotted line. This approach is used to assess the validity
of the respiration signal. Zero crossings can be discarded or
otherwise not used if the noise exceeds a predetermined threshold
(e.g., as with the sample 23,500 in FIG. 15).
[0132] An additional MDSP embodiment is directed to detecting
events in a physiological signal by combining aspects of feature
point detection and signal source extraction as discussed herein,
to detect cardiac abnormalities and arrhythmias of ventricular and
atrial origin, such as ventricular fibrillation, tachycardia,
bradycardia, atrial fibrillation and flutter, AV block and others.
In one embodiment, the intervals between QRS complexes are computed
to detect rate abnormalities indicative of tachycardia or
bradycardia. In another embodiment, a transition to tachycardia and
QRS complex morphology are evaluated to discriminate between
sinus-, supra-ventricular or life-threatening ventricular
tachycardia. In another embodiment, the ventricular rhythm
statistics and separated atrial activity rate are evaluated to
detect atrial fibrillation and flutter.
[0133] Another embodiment of involving an MDSP-based approach is
used to achieve efficient compression of quasi-periodic signals
such as ECG signals, such as by suppressing noise while preserving
signal morphology and providing accurate feature point detection.
In one embodiment, arrhythmic events are detected and ECG traces
corresponding to these events are compressed to reduce the data
storage and transmission bandwidth required to communicate the
signal to a location remote from the monitored subject. Referring
to FIG. 16, an ECG signal sensed by electrodes 1601 is conditioned
by signal conditioning circuits 1602. The digitized signal 1609 is
decomposed in process 1610 and denoised in process 1611 using MDSP
techniques described herein. An emphasis signal is computed and QRS
complexes detected in process 1612 using an MDSP embodiment
described herein. Cardiac events are detected in process 1613 using
predefined thresholds for heart rate and morphology. The denoised
ECG traces of detected arrhythmic events are reconstructed in
process 1614. The ECG traces are segmented into cardiac cycles and
are aligned using a feature point of the QRS complex in time to
form an image in process 1615.
[0134] The two-dimensional (2D) image plot thus formed includes
consecutive cardiac cycles in one dimension and the ECG signal of
each cardiac cycle in the other dimension. An illustration of the
2D image is shown as a 3D plot by way of example in FIG. 17. Plot
1701 shows a 3D plot of sequential cardiac cycles with cardiac
cycle length equalization (e.g., sequential RR intervals are padded
by a constant value at the end of the cycle to ensure that all
cardiac cycle lengths are equal). The redundancy between adjacent
beats results in more efficient compression. However, due to
physiologic factors such as respiration or rhythm disturbances the
adjacent beat redundancy might be low. In the 3D ECG plot 1702,
cardiac cycles are sorted by length (e.g., RR interval), resulting
in smoother beat transitions that lead to more efficient
compression. Accordingly, the image can be efficiently compressed
by leveraging redundancies between adjacent cardiac cycles. In
addition, redundancies across adjacent subbands or wavelet scales
can be utilized by wavelet or cosine transforms of the image.
Examples of techniques used in process 1616 that could be utilized
to achieve efficient compression of the 2D image include transform,
subband or wavelet based encoding techniques such as embedded
zerotree wavelet (EZW) [36], set partitioning in hierarchical trees
(SPIHT) [37], modified SPIHT [38] and embedded block coding with
optimal truncation (EBCOT) encoding algorithms [39]. Compression
ratios on the order of 15:1 to 20:1 with less than 5% distortion
can be achieved using this technique on denoised ECG signals.
[0135] In another embodiment, a quasi-periodic signal is compressed
by phase wrapping cardiac cycles and converting the source signals
into a polar or cylindrical system of coordinates (38). Then the
signal can be efficiently represented by a 3 dimensional plot of
phase-aligned cardiac cycles and compressed.
[0136] In another embodiment a denoised ECG is compressed by
computing a template QRS complex and subtracting it from detected
normal QRS complexes. The residual signal has lower frequency
content and can be compressed by a lossy compression technique that
can include subsampling and quantization. The decompression
technique makes use of lossless coding of the QRS template and QRS
complex locations as well as lossy coding of the residual signal to
reconstruct the ECG signal with small amount of distortion.
[0137] Other MDSP-based embodiments are used to compress blood
pressure, pulse oximetry signals, respiration, heart sounds, and
other pseudoperiodic signals. In order to achieve high levels of
compression with any of these embodiments without introducing
significant signal distortion, accurate QRS or cardiac cycle
detection and effective noise suppression (denoising) are used. In
some implementations, creating the intermediate representation of
the signal that leverages redundancy between the cycles by sorting
cycles by length, as in plot 1702 of FIG. 17, is used to achieve a
high compression rate. Accurate cycle detection is used to mitigate
the introduction of noise and artifacts into the reconstructed
signal following compression and decompression.
[0138] Another embodiment is directed to an MDSP-based approach
used to compress PNA signals and other non-quasi periodic signals,
where the denoising provided by MDSP leads to a sparse signal as
demonstrated in plot 1102 of FIG. 11. Note that the denoised signal
is almost always near zero in the absence of a neural spike. The
sparse signal is compressed using one or more compression schemes,
such as direct time-domain coding or transform based coding [40].
In general, a compression scheme for a particular physiological
signal is selected based upon the ability of a given scheme to
leverage redundancies in the signal.
[0139] Referring to FIG. 18, another example embodiment is directed
to using an MDSP-based approach to evaluate ECG strips captured by
an ambulatory monitoring device 1801 with arrhythmic event
detection capability. The captured ECG strips are forwarded to a
data review system 1804 and evaluated using an MDSP-based algorithm
1800 implemented on a computing device in data review system 1804.
Algorithm 1800 is used to evaluate the captured ECG strips 1805 and
assign a classification to each strip. Each ECG strip is assigned
one of four classifications using an MDSP-based approach for
arrhythmia detection and validity, the classifications including:
A) arrhythmia is present, B) no arrhythmia present, C) strip is
uninterpretable, or D) uncertain (e.g., the ECG strip cannot be
placed with confidence in classification A, B, or C). If an ECG is
classified as D, an automatic indicator can be triggered to suggest
human review to determine if an arrhythmia is present. For ECG
strips falling in classifications B and C, no further review is
needed. Depending upon the nature of the arrhythmia and the
clinical care process, ECG strips falling in classification A may
be reviewed by a person to assign a suggested diagnosis prior to
forwarding to a decision maker. Using this approach and considering
a relatively low percentage of the ECG strips evaluated requiring
review, labor and costs associated with providing review services
can be substantially reduced. In addition, the quality of review
services can be improved, since an accurate computer-based
algorithm can provide better consistency due to elimination of
subjectivity.
[0140] The system shown in FIG. 18 includes ambulatory monitoring
devices 1801, or subject devices, that are worn by patients being
evaluated for a heart rhythm disorder or for research. Ambulatory
monitoring device 1801 includes a computing circuit configured with
an algorithm to evaluate the ECG signal from the patient and, if an
arrhythmia is detected, capture an ECG strip containing the
arrhythmia in memory. Such captured ECG strips may, for example, be
one to five minutes in duration and stored in memory for later
wireless communication to a base station 1802 (e.g., located in the
patient's home), and from the base station 1802 to a data review
system 1804 via telecom or data network 1803. The data review
system 1804 may be located at a center where the received ECG
strips can be reviewed, if necessary, to verify the presence of an
arrhythmia or to suggest a diagnosis. The information derived from
the ECGs is then packaged into a report which is forwarded to a
researcher or clinician for use in decision making.
[0141] The results contained in a report may be provided to
physicians, clinics, hospitals, or to organizations engaged in drug
safety research, and can be delivered via a service provider
system/review center that processes the resulting signals using
data review system 1804, possibly in combination with trained
healthcare personnel. Such a review center may provide services to
a large number of clinics and physicians, or it may be housed
within a clinic or a research facility and provide service to one
or a small number of clinics or research groups or businesses.
[0142] In some embodiments, subject devices forward full-disclosure
ECG recordings, and in yet other embodiments ECG strips of 10
second to 5 minutes are captured at regular intervals for analysis
at the review center without regard to the content or nature of the
ECG signal. In some embodiments, the subject device from which
results are communicated as above is implanted in a patient. One
type of implantable device is the Reveal XT from Medtronic of
Minneapolis, Minn.
[0143] Flow chart 1800 in FIG. 18 shows an example embodiment
directed to processing and evaluation of ECG strips received by
data review system 1804. A captured ECG strip 1805 received at data
review system 1804 is evaluated for the presence or absence of
arrhythmias using an MDSP-based embodiment as described herein.
Criteria 1806 input from a care provider or other decision maker is
used by process 1807 to determine if an arrhythmia is present.
Examples criteria 1806 include a heart rate above which a rhythm is
considered to be a tachycardia, a heart rate below which a rhythm
is considered to be a bradycardia, and a minimum duration atrial
fibrillation (AF) episode required for an occurrence of AF to be
reported as an arrhythmia. A dynamic signal-to-noise ratio (dSNR)
is computed for the ECG strip in process 1808 as described herein.
A validity metric (VM) may be additionally computed using the dSNR
and signal morphology in process 1808. VM is compared to a Validity
Threshold VT1 in decision point 1809. If VM exceeds VT1, the result
is considered valid and captured ECG strip 1805 is classified "A"
if process 1807 detected an arrhythmia and "B" if process 1807 did
not detect an arrhythmia. If VM does not exceed VT1 at decision
point 1809, then VM is compared to a threshold VT2 at decision
point 1810 (VT2<VT1). If VM exceeds VT 2, then the result is
determined to be uncertain and the ECG strip is classified as "D"
to indicate that review by a person should be carried out to
determine whether an arrhythmia is present in the ECG strip. If VM
does not exceed VT2, then the ECG strip is classified as
uninterpretable ("C") and is considered uninterpretable because the
noise level is too high to be evaluated by either human or by
automated algorithm.
[0144] In some implementations, classifications A, B, and C are
assigned with a degree of certainty, designated by the validity
metrics VM1 and VM2, sufficiently high that any error in
classification can be tolerated by the user (e.g., 90% likelihood).
In come embodiments, the threshold of certainty of the
classification, VM1 and VM2, is determined by the user. Captured
ECGs assigned classification D (uncertain) generally include
segments for which a computing circuit configured with executable
software to carry out an MDSP-based processing approach is unable
to make a determination of the rhythm as being classified as A, B,
or C with a sufficiently high degree of certainty. Segments with
classification D may, for example, contain a level of noise such
that the denoised signal is not interpretable by the algorithm, but
may be interpretable by a technician, or it may contain an unusual
morphology that the software was unable to recognize.
[0145] The embodiments described here for analyzing physiologic
signals may be implemented in a variety of platforms that include a
logic circuit or computer, with reference made herein to a logic
circuit or computing circuit being applicable to a variety of such
circuits operating in accordance with one or more embodiments
discussed herein. In one embodiment, a microprocessor (such as
Pentium or Core microprocessors available from Intel of Santa
Clara, Calif.) in a personal computer running an operating system
such as Windows (available from Microsoft of Seattle, Wash.) or the
Unix standard (set by The Open Group of San Francisco, Calif.) is
used to execute programming to carry out MDSP-based functions as
discussed herein. In another embodiment, a microcontroller suitable
for low-power applications, such as the MSP 430 available from
Texas Instruments of Dallas, Tex., is implemented to carry out
MDSP-based functions as discussed herein. When a review center as
described above is involved, a microprocessor may be used for
simplicity of implementation and relatively low degree of concern
about power consumption. Many implementations involving an
MDSP-based approach in an ambulatory monitoring device employ a
microcontroller (such as the MSP 430 above) to reduce/minimize
power consumption. In other embodiments where power consumption and
size are of high priority, implementation in a silicon-based state
machine using a hardware description language such as VHDL may be
useful.
[0146] In some embodiments, referring to FIGS. 16 and 19, aspects
of the present invention are implemented using a battery powered or
passively-powered device (e.g., via radio frequency power) that is
worn by or implanted within a human or animal subject. Depending
upon the application and specific design requirements, referring to
FIG. 18, various aspects of MDSP-based embodiments discussed herein
may be partitioned between implementation within subject device
1801 and implementation within the data review system 1804.
[0147] Referring to FIG. 19, an apparatus for improving the
signal-to-noise ratio of a physiological signal is shown, in
accordance with another example embodiment. While referencing an
ECG signal, the apparatus shown in FIG. 19 may be implemented with
other signals such as blood pressure, respiration,
photoplethysmography, blood glucose, blood flow, heart sounds, PNA,
EMG, and EEG. In this example embodiment, ECG is sensed using
either surface or implanted electrodes 1901. Signal conditioning
circuits 1902 receives the signal from sensing electrodes 1901 and
is conditioned to amplify and filter the signal to remove much of
the noise outside the bandwidth of the ECG signal.
Analog-to-digital conversion (ADC) is accomplished by an ADC on
board a Texas Instruments MSP430 microcontroller 1903 (shown by way
of example, and implementable with other processors).
[0148] Referring to the right side of FIG. 19, digitized signal
1906 is processed by computer instructions executed by a 16-bit
RISC MCU of the MSP430 1903. The digitized signal 1906 is
decomposed into a combination of basis functions (subcomponents) in
a second domain of higher dimension than the first domain in
process 1907. The dimension of the first domain is equal to the
number of sensed ECG signals. Decomposition in process 1907 is
performed using one of a discrete cosine transform, a wavelet
related transform, a Karhunen-Loeve transform, a Fourier transform,
a Gabor transform, and a filter bank. A subset of subcomponents
containing primarily signal energy is identified using MDSP
techniques (e.g., one or more of SSF, PCA, ICA, and .pi.CA) in
process 1908. In process 1909, a denoised signal is reconstructed
from the subcomponents identified as primarily containing signal
energy by performing the inverse of the transform used to decompose
the signal. It should be noted that the concepts described here can
also be applied to other MDSP-based embodiments beyond denoising,
including feature point detection, event detection, and computing a
dynamic signal-to-noise ratio to evaluate the accuracy of
information extracted from a physiological signal.
[0149] Referring back to FIG. 16, another example embodiment
involves using an MDSP-based approach for implementing a
battery-powered apparatus capable of wirelessly communicating
physiological signals. Referring to FIG. 18, this apparatus is one
embodiment of subject device 1801. The example refers specifically
to ECG signals, but a similar embodiment can be used for other
pseudoperiodic physiological signals such as arterial blood
pressure, respiration, blood oxygen saturation derived from PPG,
heart sounds, and blood flow for example. In this example
embodiment, ECG is sensed using either surface or implanted
electrodes 1601. Signal conditioning circuits 1602 receive the
signal from electrodes 1601 and amplify and filter the signal to
remove much of the noise outside the bandwidth of the ECG signal.
Analog-to-digital conversion (ADC) of the conditioned signal is
accomplished by an ADC on board a Texas Instruments MSP430
microcontroller 1603.
[0150] Referring to the right side of FIG. 16, digitized signal
1609 is processed by computer instructions executed by the 16-bit
RISC MCU of MSP430 1603. The MCU is in communication with offboard
memory 1605 which may be used to provide additional data storage
and for storage of computer instructions. The MCU is additionally
in communication with wireless transmitter 1604 that can send
compressed data to wireless receiver 1606 located remote from
subject device 1801 of FIG. 18. Wireless receiver 1606 is further
in communication with a logic circuit or computer that is
configured to execute instructions to decompress the compressed
signal in process 1607. In some embodiments the wireless
transmitter 1604 and receiver 1606 may each be wireless
transceivers capable of both sending and receiving data. An example
of a wireless transceiver that may be used in connection with this
embodiment is the CC2540 low-energy Bluetooth chip available from
Texas Instruments (see above).
[0151] Referring to the data flow diagram on the right side of FIG.
16, digitized signal 1609 is decomposed in process 1610 into a
combination of basis functions (subcomponents) in a second domain
of higher dimension than the first domain. The dimension of the
first domain is equal to the number of sensed ECG signals.
Decomposition in process 1610 is performed using one of a discrete
cosine transform, a wavelet related transform, a Karhunen-Loeve
transform, a Fourier transform, a Gabor transform, and a filter
bank. The subcomponents are denoised in process 1611 using MDSP
techniques (e.g., one or more of SSF, PCA, ICA, and .pi.CA) and a
QRS emphasis signal is computed in process 1612 as a linear
combination of a subset of denoised subcomponents containing QRS
signal wave energy. The emphasis signal is evaluated to detect each
QRS complex, as described herein, and a feature signal containing a
series of feature points indicating the R-R interval of consecutive
cardiac cycles is constructed in process 1613. The feature signal
and morphology of the denoised ECG are evaluated in process 1613,
as described herein, to detect arrhythmic events, such as
bradycardia and tachycardia, and the denoised ECG strips containing
arrhythmic events reconstructed in process 1614. The denoised ECG
strips to be transmitted are compressed using processes 1615 and
1616. In process 1615, ECG strips are segmented by cardiac cycle
and adjacent strips are aligned using a feature point of the QRS
complex to form an image of a 3D plot similar to that shown in plot
1701 of FIG. 17. The image is subsequently encoded in process 1616
by, for example, applying transform encoding, subband encoding, or
wavelet based encoding, after which the encoded image is saved in
memory for later wireless transmission or it may be sent
immediately.
[0152] In an alternate embodiment, instead of forming the image by
aligning adjacent strips, the strips are arranged by cardiac cycle
length, as in plot 1702 of FIG. 17. This results in additional
redundancies in strips adjacent to each other in the image and
hence results in more efficient compression.
[0153] In another embodiment a template QRS complex is computed and
subtracted from detected normal QRS complexes. The residual signal
is compressed by a lossy compression technique and the QRS template
and QRS complex locations are compressed using lossless coding.
[0154] In order to meet power consumption requirements for
implementation within a subject device for denoising and wireless
communication, computer-executable instructions for carrying one an
MDSP-based approach as discussed herein, stored as embedded code
within the subject device, are configured to facilitate low-power
implementation. In one embodiment, the computer instructions are
optimized using integer or fixed point arithmetic and lifting or
B-spline wavelet implementation [41] of a signal decomposition
transform in order to reduce and/or minimize the number of clock
cycles or machine states required. In such an embodiment, a portion
of the computations required to analyze physiologic signals may be
implemented within the subject device while others may be
implemented in the data review system. In another embodiment, the
subject device captures, denoises, and compresses the ECG of the
subject and information is extracted from the ECG recording
off-line in the data review system. The data review system may
include a review function that facilitates human review of ECGs
that were classified as uncertain by the algorithm.
[0155] Other embodiments are directed to a computer-based system or
logic circuit, such as a computer operating using one or more
processor circuits, which operate using executable modules. Each
module (e.g., software-based module) is executable by a computer
circuit to carry out one or more functions or processes as
described herein. For instance, one such module may carry out MDSP
computations upon input data such as ECG data, and transform the
data into a denoised output. This transformation may involve, for
example, the use of a software module that, when executed by a
computer, carries out the steps as shown in FIGS. 1 and 2. Various
other embodiments are directed to similar transformations, which
may involve taking and processing one or more inputs to generate a
transformed output useful for one or more of a variety of purposes,
such as for detecting physiological conditions. Accordingly, the
embodiments discussed herein may be carried out using such a system
and/or computer circuit and related executable modules.
[0156] In one embodiment, and referring again to FIG. 19, a
denoising function is implemented within a computerized apparatus
configured to execute programming instructions. A digitized
physiological signal 1906 is input to the computerized apparatus.
The device computes a denoised output signal from the digitized
input signal resulting in an improvement in SNR.
[0157] In one implementation, an input physiological signal sensed
by ECG electrodes 1901 is a relatively noise-free single lead ECG
recording from a human being or other mammal with a resting heart
rate less than 150 BPM, contaminated with band-limited (0.5 to 100
Hz) white noise. Following denoising in processes 1907, 1908, and
1909 using an MDSP embodiment, the SNR is improved by at least 5
decibels and the mean QRS amplitude for any 10 second interval of
the recording is preserved within +/-10% of the mean QRS amplitude
of the input signal for the same 10 second interval.
[0158] In this embodiment, the input ECG can be described as
residing in a first domain having a dimension equal to the number
of leads (e.g., channels) in the recording. For example, a
recording consisting of a lead set (e.g., Leads I, II, and III),
would have three dimensions. Referring to FIG. 19, the digitized
input ECG 1906 is decomposed in process 1907 into a second domain
of higher dimension than the first domain. Decomposition into the
second domain results in generation of subcomponents that represent
the information contained in the ECG signal. The dimension of the
second domain is defined as the number of subcomponents
representing each lead of the ECG multiplied by the number of
leads. In one embodiment, decomposition performed in process 1907
is accomplished using a transform that largely achieves
independence of the subcomponents. Independence of the
subcomponents combined with qualification of the frequency content
of each subcomponent facilitates a more precise identification of
the association of a subcomponent or group of subcomponents with an
aspect (e.g., T-wave) of the ECG signal.
[0159] In another embodiment, the subcomponents in the second
domain are examined in process 1908, as described herein, based on
their spatial distribution in time, to identify noise and other
aspects of the ECG. In one embodiment, the spatial distribution is
examined using spatially selective filtering, whereby aspects or
waves of the ECG associated with wider frequency band, such as the
QRS complex, are identified and preserved across the subcomponents.
In this embodiment, subcomponents associated with high frequencies
are preserved for a time window corresponding to the QRS complex,
but are zeroed out in (or not used for) the time window
corresponding to the remainder of the cardiac cycle. This allows
the morphology of the QRS complex to be preserved, while removing
the high frequency noise from the remainder of the cardiac cycle,
where the primary signal content corresponds to lower frequencies.
When the input signal is a multi-lead ECG, principal component
analysis and independent component may be used in addition to
spatially selective filtering to further enhance the independence
of subcomponents.
[0160] In another example embodiment, referring to FIG. 20,
spatially selective filtering is used to remove noise from an ECG.
FIG. 20a (top) shows a cardiac cycle of an ECG waveform showing a
P-wave, QRS complex 2001, T-wave, and U-wave, as processed in
connection with this embodiment. The QRS complex is detected and,
referring to FIG. 20b (middle), the time spanning the beginning of
Q to end of S defines Time Window 2002. The remainder of the
cardiac cycle is defined as Time Window 2003. Time Window 2002,
containing the QRS complex, includes a wide range of frequencies
ranging from low-frequencies to high-frequencies. For the ECG of a
healthy human, this frequency range may span from 0.5 to 100 Hz. In
order to preserve the morphology of the QRS complex, subcomponents
in this frequency range are preserved and considered to be
associated with a desired signal wave. In one embodiment, the
beginning and end of Time Window 2002 are respectively defined by
the onset of the Q-wave and the offset of the S-wave. In another
embodiments, the beginning and end of Time Window 2002 are shifted
somewhat earlier or later. Because the signal waves occurring in
Time Window 2003 (e.g., P-wave, T-wave, U-wave) contain lower
frequency components, the subcomponents comprising the
low-frequencies, represented by region 2005, are preserved and the
subcomponents comprising high-frequencies are removed since they
are primarily are associated with noise.
[0161] In an alternate embodiment as shown in FIG. 20c,
subcomponents associated with higher frequencies are removed in
Time Windows 2002 and 2003 while retaining low-frequency components
represented by region 2007. Those containing the higher frequencies
are then added back during the time spanning Time Window 2002, as
represented by region 2006, to restore the higher frequency
components of the QRS complex. In some implementations, this
approach is carried out to mitigate or avoid morphology
distortion.
[0162] In one embodiment, and referring to FIG. 21, an input ECG
signal 2101 is decomposed into subcomponents, as described herein,
in process 2102. The QRS complex is identified in process 2103
using one of a number of techniques including thresholding,
adaptive thresholding, and spatially selective filtering, as
described herein. Following identification of the QRS complex the
ECG is segmented by cardiac cycle using the QRS complex as a
fiducial point. In one embodiment, each cardiac cycle is
partitioned into two time windows in process 2104. Referring to
FIG. 20, first time window 2002 begins at the onset of the Q-wave
of the cardiac cycle and ends at the offset of the S-wave. The
second time window 2003 begins at the offset of the S-wave and ends
at the onset of the Q-wave of the next cardiac cycle. In other
embodiments, the start and end times of windows 2002 and 2003 can
differ somewhat from the above relative to the location of feature
points of the QRS complex and some degree of overlap in the windows
is also acceptable. In other embodiments, the cardiac cycle is
partitioned into more than two time windows. For example, in an
alternate embodiment a first time window includes the QRS complex,
a second time window includes the T-wave, and a third time window
includes the remainder of the cardiac cycle. The set of
subcomponents present within time window 2002 are referred to as
QRSsub and the set of subcomponents present in time window 2003 are
referred to as PTsub.
[0163] In process 2105, subcomponents of the set PTsub are
evaluated to identify subcomponent subset PTsubnn that overlaps the
spectral content of (are associated with) a characteristic
frequency band, FBpt, of the representative ECG signal present in
window 2003. The value of a subcomponent represents the energy of
the input ECG signal contained in a narrow range of frequencies and
the subcomponent is said to be associated with this narrow range of
frequencies.
[0164] Process 2105 relies on knowledge of the characteristic
frequency band, FBpt, of the desired ECG signal in window 2003.
FBpt is pre-identified using a database of representative ECG
recordings for a species as described in process 2108.
[0165] In process 2106, subcomponents of the set QRSsub are
evaluated to identify subcomponent subset QRSsubnn that overlaps
the spectral content of (are associated with) a characteristic
frequency band, FBqrs, of the representative ECG signal present in
window 2002 of FIG. 20. In one embodiment, the input ECG signal is
preprocessed to remove out-of-band noise in a signal conditioning
circuit such as 1902 in FIG. 19, by a digital filter implemented in
a computing or logic circuit, or a combination thereof. In one
embodiment, where the energy of out-of-band noise in the input
signal is negligible, the energy of the QRS subcomponents QRSsub
and QRSsubnn essentially overlap.
[0166] Process 2106 relies on knowledge of the characteristic
frequency band, FBqrs, of the desired ECG signal in window 2002.
FBqrs is pre-identified using a database of representative ECG
recordings for a species as described in process 2108.
[0167] The identification of the characteristic frequency bands for
time window is performed by evaluating a database of representative
ECG recordings for a species (input 2017). The database of
representative ECG recordings is selected to represent a broad
scope of ECG morphologies, heart rate, anomalies, noise
characteristics, and pathologies. In the process 2108, the spectral
content of the QRS complex is characterized and the characteristic
frequency band of the QRS complex FBqrs is identified using
spectral analysis techniques. The time window outside of the QRS
complex in cardiac cycle is identified and the characteristic
frequency band of the cardiac cycle outside of the QRS complex,
FBpt, is identified using spectral analysis techniques. It should
be noted that computing the characteristic frequency band is
typically only performed once for the ECG sampled for a given
species. Once FBpt and FBqrs are determined for a given species,
they are used as parameters by the processes 2105 and 2106 and it
is not necessary to recompute them.
[0168] In the process 2111 the identified target subcomponents
contained in subset PTsubnn and subset QRSsubnn are combined and
subjected to the inverse of the transform used in process 2102 to
construct a denoised ECG signal.
[0169] FIG. 22 shows a system 2200 for computing a denoised ECG
signal from an input signal including a desired ECG signal and
noise, according to another example embodiment. The system includes
a logic circuit 2210 and a memory circuit 2212. The memory circuit
2212 stores instructions that, when executed by the logic circuit,
carry out the following steps. For illustration, FIG. 22 represents
steps as carried out in the logic circuit 2210 with various blocks;
however, various approaches for carrying out the steps may be
implemented in connection with other embodiments. Moreover, the
logic circuit 2210 may include two or more logic circuits, such as
two or more processors that carry out different aspects of the
respective steps. In addition, the various steps shown in FIG. 22
and described here may be implemented using one or more embodiments
as discussed above, in connection with the other figures and
otherwise.
[0170] Referring again to FIG. 22, an input ECG signal 2205, in a
first domain, is received at a communications input port 2214, and
is decomposed at block 2220 into subcomponents 2225 in a second
domain (e.g., of higher dimension than the first domain as
discussed hereinabove). QRS window-based identification is carried
out at block 2230, at which a location of the QRS complex of a
cardiac cycle in the ECG signal is identified. A first time window
in the cardiac cycle that includes the QRS complex is identified,
along with at least one time window in the cardiac cycle that does
not include the QRS complex. The first time window and the at least
one time window span the duration of the cardiac cycle.
[0171] At block 2240, and for each of the identified time windows,
target subcomponents 2245 are identified as subcomponents that
contain more desired ECG signal energy than noise energy. For
example, when a particular subcomponent exhibits more energy
associated with a desired ECG signal than energy not associated
with such a signal, that subcomponent may be identified as a target
subcomponent. The target subcomponents 2245 are used at block 2250
to construct a denoised physiological signal 2255.
[0172] The denoised ECG signal is then output via a communications
output port 2216. This output may, for example, involve a wired or
wireless communication, and may be carried out in accordance with
one or more embodiments as described herein. In some
implementations, some or all of the logic circuit 2210 is included
as part of an implantable device, and carries out some or all of
the steps in the implantable device, for generating the denoised
physiological signal 2255. Using this approach, and as consistent
with the above, the logic circuit 2210 can be implemented to
significantly reduce the size of the denoised and compressed
physiological signal 2255 (e.g., as relative to approaches that do
not denoise in this manner), and facilitate the communication of
the denoised and compressed physiological signal using less data
and, correspondingly, lower power.
[0173] Those skilled in the art will appreciate that various
alternative logic circuits or computing arrangements, including one
or more processors and a memory arrangement configured with program
code, would be suitable for carrying out the approaches as
discussed herein, including those discussed in connection with FIG.
22 above, along with data structures for organizing the required
data. Such computer code can be encoded in a processor executable
format and may be stored on and/or provided via a variety of
computer-readable storage media or delivery channels such as
magnetic or optical disks or tapes, electronic storage devices, or
as application services over a network. With specific reference to
FIG. 22, the logic circuit 2210 (and memory 2212, where
appropriate) may be implemented with separate components on a
circuit board or may be implemented internally within an integrated
circuit. When implemented internally within an integrated circuit,
the logic circuit 2210 can be implemented as a microcontroller.
[0174] The architecture of the logic circuits, processors and
computer type circuits as described herein depends on
implementation requirements as would be recognized by those skilled
in the art. In this context, these components may be one or more
general purpose processors, or a combination of one or more general
purpose processors and suitable co-processors, or one or more
specialized processors (e.g., RISC, CISC, and pipelined).
[0175] Referring again to FIG. 22, the memory circuit 2212 may
include multiple levels of cache memory and a main memory, and
local and/or remote persistent storage such as provided by magnetic
disks, flash, EPROM, or other non-volatile data storage. The memory
circuit 2212 may be read or read/write capable. The logic circuit
2210 may store instructions (e.g., software) in the memory circuit
2210, read data from and stores data to the memory circuit 2212,
and communicate with external devices through the input/output
ports 2214 and 2216. These functions may be synchronized by a clock
signal generator that may be part of the logic circuit 2210. The
resources of the logic circuit 2210 may be managed by either an
operating system, or a hardware control unit.
[0176] Referring to process 1907 in FIG. 19, example transforms
that can be used for decomposing the signal into a second domain of
higher dimension while enhancing, or maximizing, the independence
of the subcomponents include a discrete cosine transform, a wavelet
related transform, a Karhunen-Loeve transform, a Fourier transform,
a Gabor transform, or a filter bank.
[0177] In another embodiment, referring to FIG. 16, a denoising
function and a data compression function are implemented within a
computerized apparatus (e.g., a logic circuit) configured to
execute programming instructions. In this embodiment, denoising and
compression of an ECG results in a reduction in bit rate required
to retain the information in the signal. By compressing the ECG
signal prior to wireless transmission, fewer bits can be used to
achieve data transmission, and hence the power consumed in the
transmission (e.g., a telemetry link) is reduced. This approach may
also be implemented when storing ECG data (or other signal data) in
memory, to reduce the data storage space required.
[0178] In a more particular example embodiment, a denoising and
compression involves reducing the bit rate of a signal by a factor
of 15:1 to 20:1, relative to the bit rate of input signal. For
instance, such a compression factor can be achieved with a 16-bit
digitized single lead ECG having a bit rate of 4,000 bits per
second and sampled at 250 Hz, and with band-limited (0.5 to 100 Hz)
noise corresponding to a SNR of 4 dB. The compressed signal is
communicated to a receiving device where the compressed ECG signal
is reconstructed. The denoising approach facilitates reconstructing
QRS complex, from the compressed signal, having mean amplitude for
any 10-second interval therein that differs from the mean amplitude
of the input ECG by 10% or less. In one embodiment, communication
is performed using a wireless link such as a Bluetooth transceiver.
In some embodiments, events in the ECG, such as arrhythmias, are
detected and compression and transmission of the ECG is triggered
based on the presence of a detected event.
[0179] In this example embodiment, the input ECG is processed by
the computerized apparatus to first denoise the signal, then detect
cardiac cycles, segment the ECG by cardiac cycles and align them in
time according to an identifiable fiduciary (e.g., a QRS peak), and
form an image consisting of aligned cardiac cycles. The image is
then encoded using a lossy compression technique such as transform
encoding, subband encoding, or wavelet based encoding. The encoded
image contains nearly all of the information in the denoised ECG,
but does so with about 15-fold fewer bits. The encoded image is
reconstructed at a receiver using the inverse of the transform used
to encode the image, and ECG segments are reconnected to form a
denoised version of the input signal. In one embodiment, denoising
prior to compression is achieved in the same manner as described
herein for the apparatus used for denoising.
[0180] In another embodiment, cardiac cycles are detected by
computing a QRS emphasis signal and evaluating peaks, valleys, or
slopes to identify the QRS complex. In another embodiment, the QRS
complex is detected and a template QRS complex that is
representative of the average complex is subsequently computed. The
template is then subtracted from each QRS complex, creating a time
series consisting of the resulting difference. The time series of
residuals is subsequently encoded using a lossy compression
technique such as transform encoding, subband encoding, or wavelet
based encoding. The template QRS is encoding using a lossless
transform such as Huffman encoding. This approach may be useful,
for example, in connection with an embodiment in which a relatively
simpler and less computationally intense compression technique is
appropriate (e.g., where QRS morphology is not expected to change
rapidly). Accordingly, the computational intensity of denoising and
resulting signal quality/size can be weighed as a tradeoff, for
implementation in various embodiments.
REFERENCES CITED
[0181] For general information regarding a variety of fields that
may relate to one or more embodiments of the present invention, and
for specific information regarding the application of one or more
such embodiments, reference may be made to the following documents,
which are fully incorporated herein by reference. Various ones of
these references are further cited above via corresponding
numerals, and may be implemented as such. [0182] 1. A. Hyv{umlaut
over ( )}arinen and E. Oja. Independent component analysis:
Algorithms and applications. Neural Networks, 13(4-5):411-430,
2000. [0183] 2. K. S. Ball, L. Sirovich, L. R. Keefe, Dynamical
eigenfunction decomposition of turbulent channel flow,
International Journal for Numerical Methods in FluidsVolume 12,
Issue 6, Date: 5 Apr. 1991, Pages: 585-604 [0184] 3. S. Mallat, A
Wavelet Tour of Signal Processing, Academic Press, 1999. [0185] 4.
K. R. Rao and P. Yip, Discrete Cosine Transform: Algorithms,
Advantages, Applications San Diego, Calif.: Academic, 1990. [0186]
5. Mallat, S. G., and Zhang, Z., Matching Pursuits with
Time-Frequency Dictionaries, IEEE TSP(41), No. 12, December 1993,
pp. 3397-3415. [0187] 6. Donoho, D. L.; Huo, X.; Uncertainty
principles and ideal atomic decomposition., IEEE Transactions on
Information Theory Volume 47, Issue 7, November 2001
Page(s):2845-2862 [0188] 7. Jang, G.-J. et. al. Single-channel
signal separation using time-domain basis functions. IEEE Signal
Processing Letters, vol. 10, Issue: 6, pp 168-171 [0189] 8. Y. C.
Pati, R. Rezaiifar, and P. S. Krishnaprasad, "Orthogonal matching
pursuit: recursive function approximation with applications to
wavelet decomposition," in Asilomar Conference on Signals, Systems
and Computers, 1993, vol. 1, pp. 40-44. [0190] 9. S. S. Chen, D. L.
Donoho, and M. A. Saunders, "Atomic decomposition by basis
pursuit," SIAM J. Scientific Computing, vol. 20, no. 1, pp. 33-61,
1999. [0191] 10. Aharon, M.; Elad, M.; Bruckstein, A., K-SVD: An
Algorithm for Designing Overcomplete Dictionaries for Sparse
Representation, IEEE Transactions on Signal Processing, page(s):
4311-4322, Volume: 54, Issue: 11, November 2006 [0192] 11. M.
Alghoniemy, Ah. Tewfik, "Reduced complexity bounded error subset
selection," in IEEE Int. Conf. Acoustics, Speech and Signal
Processing (ICASSP), March. 2005, pp. 725-728.
[0193] 12. K. Zhang, L. W. Chan, "An Adaptive Method for Subband
Decomposition ICA", Neural Computation, vol. 18, no. 1, pp.
191-223, 2006. [0194] 13. Vaidyanathan, Multirate Systems and
Filter Banks, Prentice Hall, 1993 [0195] 14. L. Smith A tutorial on
Principal Components Analysis,
http://users.ecs.soton.ac.uk/hbr03r/pa037042.pdf [0196] 15.
Aminghafari, M.; Cheze, N.; Poggi, J-M. (2006), "Multivariate
de-noising using wavelets and principal component analysis,"
Computational Statistics & Data Analysis, 50, pp. 2381-2398
[0197] 16. P. Comon, "Independent component analysis, a new
concept?," Signal Process. Special Issue on Higher Order
Statistics, vol. 36, no. 3, pp. 287-314, 1994. [0198] 17. J.-F.
Cardoso, "High-order contrasts for independent component analysis,"
Neural Comput., vol. 11, no. 1, pp. 157-192, 1999. [0199] 18.
Blaschke, T.; Wiskott, L.; CuBICA: independent component analysis
by simultaneous third- and fourth-order cumulant diagonalization
IEEE Transactions on Signal Processing, Volume 52, Issue 5, May
2004 Page(s):1250-1256 [0200] 19. Hyvarinen, A. New approximations
of differential entropy for independent component analysis and
projection pursuit. In Advances in Neural Information Processing
Systems, volume 10, pages 273-279.MIT Press. [0201] 20. Bell, A.
and Sejnowski, T., An information-maximization approach to blind
separation and blind deconvolution, Neural Computation,
7:1129-1159. [0202] 21. Cardoso, J.-F. Infomax and maximum
likelihood for source separation. IEEE Letters on Signal
Processing, 1997, 4, 112-114. [0203] 22. M. Zibulevsky, B.
Pearlmutter, Blind source separation by sparse decomposition in a
signal dictionary, Neural Computation 2001, v13, pp 863-882. [0204]
23. Todros, K.; Tabrikian, J.; Blind Separation of Independent
Sources Using Gaussian Mixture Model IEEE Transactions on Signal
Processing Volume 55, Issue 7, Part 2, July 2007 Page(s):3645-3658
[0205] 24. Hyvarinen, A. Fast and robust fixed-point algorithms for
independent component analysis. IEEE Transactions on Neural
Networks, 10(3):626-634. [0206] 25. L. K. Saul and J. B. Allen,
"Periodic component analysis: An eigenvalue method for representing
periodic structure in speech.," in NIPS, [Online] 2000, pp.
807-813. Available:
http://www.cs.cmu.edu/Groups/NIPS/00papers-pub-on-web/SaulAllen.pdf
[0207] 26. R Sameni, et. al. Multichannel electrocardiogram
decomposition using periodic component analysis. IEEE Transactions
on Biomedical Engineering, 2008 vol 55, no 8 pp 1935-1940 [0208]
27. Mallat, S. and Zhong, S. 1992. Characterization of Signals from
Multiscale Edges. IEEE Trans. Pattern Anal. Mach. Intell. 14, 7
(July 1992), [0209] 28. Mallat, S. G., Hwang, W. L., Singularity
Detection and Processing with Wavelets, IEEE Transactions on
Information Technology (38), 1991, pp. 617-643. [0210] 29. Xu,
Yansun, et. al. Wavelet transform domain filters: a spatially
selective noise filtration technique, IEEE transactions on image
processing 1994, vol. 3, no6, pp. 747-758 [0211] 30. Pan Q., Zhang
L. Dai G. and et al. Two denoising methods by wavelet transform.
IEEE Trans. on SP., 1999, 47(12): 3401-3406 [0212] 31. Donoho, D.
L., "Denoising by soft-thresholding," IEEE Trans. on Inf. Theory,
42 3, pp. 613-627, 1995 [0213] 32.
http://www.physionet.org/physiobank/database/#ecg [0214] 33. M.
Hassan, J. Terrien, B. Karlsson, and C. Marque, "Spatial analysis
of uterine EMG signals: Evidence of increased in synchronization
with term," Conf Proc IEEE Eng Med Biol Soc, vol. 1, pp. 6296-9,
2009. [0215] 34. X. Li, X. Yao, J. Fox, and J. G. Jefferys,
"Interaction dynamics of neuronal oscillations analysed using
wavelet transforms," J Neurosci Methods, vol. 160, pp. 178-85, Feb.
15 2007. [0216] 35.
http://www.physionet.org/physiobank/database/mitdb/36. [0217] 36.
M. L. Hilton. Wavelet and wavelet packets compression of
electrocardiogram. IEEE Transactions on Biomedical Engineering,
44(5):394-402, May 1997. [0218] 37. Z. Lu, D. Y. Kim, and W. A.
Pearlman. Wavelet compression of ECG signals by the set
partitioning in hierarchical trees algorithm. IEEE Tran-sactions on
Biomedical Engineering, 47(7):849-856, July 2000. [0219] 38. S. C.
Tai, C. C. Sun, and W. C. Tan, "2-D ECG compression method based on
wavelet transform and modified SPIHT," IEEE Trans. Biomed. Eng.,
vol. 52, no. 6, pp. 999-1008, June 2005 [0220] 39. Marcellin M.,
et. al, An Overview of JPEG-2000, Proc. of IEEE Data Compression
Conference, pp. 523-541, 2000. [0221] 40. K. Sayood, Introduction
to Data Compression, Academic Press 2000. [0222] 41. K. G. Oweiss,
A. Mason, Y. Suhail, A. M. Kamboh and K. E. Thomson "A scalable
wavelet transform VLSI architecture for real-time signal processing
in high-density intra-cortical implants", IEEE Trans. Circuits
Syst. I, vol. 54, pp. 1266 2007. [0223] 42. Martinez J-P, et. al. A
wavelet-based ECG delineator: Evaluation on standard databases.
IEEE transactions on biomedical engineering 2004, vol. 51, no4, pp.
57 [0224] 43. R Sameni, et. al. A nonlinear Bayesian filtering
framework for ECG denoising. IEEE transactions on biomedical
engineering, vol 54, n012, pp 2172-2185, 2007 [0225] 44. S.
Paredes, T. Rocha, P. de Carvalho, J. Henriques, Atrial Activity
Detection through a Sparse Decomposition Technique," vol. 2, pp.
358-362, 2008 International Conference on BioMedical Engineering
and Informatics, 2008 [0226] 45. D A Clunie, Extension of an open
source DICOM toolkit to support SCP-ECG waveforms. 2nd OpenECG
Workshop 2004, Berlin, Germany
[0227] Based upon the above discussion and illustrations, those
skilled in the art will readily recognize that various
modifications and changes may be made to the present invention
without strictly following the exemplary embodiments and
applications illustrated and described herein. Such modifications
and changes may include, for example, incorporating one or more
aspects described in the above references and/or applying one or
more embodiments thereto, or combining embodiments. These and other
modifications do not depart from the true spirit and scope of the
present invention, including that set forth in the following
claims.
* * * * *
References