Patent Application
20140136585
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.
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.
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.
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.
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 is 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.
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.
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.
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.
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.
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 problematic the independent observation of the sources of these multisource signals. 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.
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.
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.
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.
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.
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.
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.
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.
In another aspect of the present invention, feature points of signal waves are evaluated to detect clinically significant events and signal morphology characteristics.
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.
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.
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.
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.
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:
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.
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.
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, atrial 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.
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.
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.
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.
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 and 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.
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.
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.
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.
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 signal that includes 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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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)=(s1 (t), . . . , sP(t)) is a vector of source signals that are mixed together by an unknown mixing matrix A of size M×P, where M is a number of observed signals and P is a number of source signals and additive noise n(t)=(n1(t), . . . , nP(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)=(s1(t), . . . , sP(t)) from the observed signals x(t)=(x1(t), . . . , xM(t)) according to the formula s(t)=A−1x(t)−A−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 AM×P is underdetermined. In addition, the sources that make up multisource physiological signals are often not independent. In this situation the matrix AM×P is not invertible, rendering it difficult to estimate an inverse of the M×P mixing matrix and recover unmixed sources from mixed observations, as is the case when using ICA techniques.
Turning now to the figures,
Referring to
If the dimension of the first domain is greater than 1, then SSF, πCA, 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.
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
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,
where dik and cik are decomposition coefficients or subcomponents. In some embodiments, the elements φk 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 dik and cik in equation (2) are close to or equal to zero. A combination of basis functions or elements of the dictionary forms the second domain.
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).
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:
where ∥.∥0 is a l0 norm, counting the nonzero entries of a vector, T is a constraint on the maximum number of non-zero elements allowed, and ∥.∥F2 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 built 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.
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 xi (t),i=1:M, can be represented as a sum of subband signals
x
i(t)=xi,1(t)+ . . . +xi,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]
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 Wj with each iteration step j making up a decomposition level. In this embodiment, the operator matrix Wj is a Toeplitz matrix with the entries formed by a wavelet impulse response. The size of the matrix operator Wj is N×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].
Referring again to
W
x
j
=A
j
W
s
j
+W
n
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.
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 Aj 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 (πCA) techniques and spatially selective filtering as shown in blocks 104 and 105 of
When using the PCA technique to identify subcomponents associated with signal sources, an inverse of the mixing matrix Aj 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.
In various embodiments involving the use of an ICA technique to identify subcomponents associated with signal sources, an inverse of the mixing matrix Aj 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
In embodiments involving the use of πCA technique to identify signal sources, an inverse of the mixing matrix Aj 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 πCA 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.
In one embodiment, a quasi-periodic signal can be phase-wrapped by mapping the period length to a linear phase φ(t) ranging from −π to +π 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.
In another embodiment, referring to
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.
Referring to
In some embodiments the spatially selective filtering is combined with one or more of PCA, ICA, or πCA. 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.
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 πCA may be applied, as in blocks 105 and 205 in
In some embodiments, step three as shown in blocks 106 and 206 of
Referring to
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
Where Psignal and Pnoise are respective signal and noise energy and Asignal and Anoise are respective signal and noise amplitude.
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 πCA 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).
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.
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.
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.
In the following examples, performance aspects are discussed as may be relevant to various embodiments 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.
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 xcl and denoised signal xden calculated sample-by-sample as a percentage of the original signal variance.
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.
where σsignal and σnoise are respective clean-signal and noise standard deviations.
Referring to the plots 710 and 711 of
Referring to the ECG tracings 707, 708, and 709 of
The embodiments shown in
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. 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 shown in
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.
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, πCA, or their combination, as part of step 2 (process 204 and 205) as shown in
In
In
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.
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
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.
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.
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.
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.
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.
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
Referring to
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.
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.
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.
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.
Referring again to
In another embodiment (e.g., relative to
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.
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.
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
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
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.
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.
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
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
Referring to
The system shown in
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.
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.
Flow chart 1800 in
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 some 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.
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.
In some embodiments, referring to
Referring to
Referring to the right side of
Referring back to
Referring to the right side of
Referring to the data flow diagram on the right side of
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
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.
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.
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
In one embodiment, and referring again to
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.
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
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.
In another example embodiment, referring to
In an alternate embodiment as shown in
In one embodiment, and referring to
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.
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.
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
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.
The identification of the characteristic frequency bands for time window is performed by evaluating a database of representative ECG recordings for a species (input 2107). 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.
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.
Referring again to
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.
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.
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
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).
Referring again to
Referring to process 1907 in
In another embodiment, referring to
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.
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.
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 encoded 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.
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.
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.
This patent document is a continuation under 35 U.S.C. §120 of U.S. patent application Ser. No. 12/938,995 filed on Nov. 3, 2010 (U.S. Pat. No. 8,632,465), which claims the benefit under 35 U.S.C. §119 of U.S. Provisional Patent Application Ser. No. 61/257,718 filed on Nov. 3, 2009, and of U.S. Provisional Patent Application Ser. No. 61/366,052 filed on Jul. 20, 2010; each of these patent documents is fully incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| 61257718 | Nov 2009 | US | |
| 61366052 | Jul 2010 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 12938995 | Nov 2010 | US |
| Child | 14156219 | US |
