Patent Grant
9209782
This invention relates to the field of computer-implemented systems and methods, and more specifically a software-, embedded-circuits- or firmware-implemented system and method to decompose signals having quasi-periodic wave properties, store such signals in a data structure, analyze such signals, and reconstruct such signals from the data structure, and to optionally transmit such data structure over a communications channel.
The digital representation of waveforms is a technology that is central to various sectors of industry where the detection of periodic and non-periodic waveforms can be critical to determining whether an erratic heartbeat, electrical short circuit, or some other problem exists. A digital representation must clearly and accurately represent the analog source of a waveform, but at the same time be able to accomplish such things as, compressing the incoming data into some manageable size, and maintain the integrity of the incoming data (i.e., making sure that the digital representation has enough fidelity to the original signal to be useful). Of additional import is the ability to have a digital representation that can consistently allow one to identify the presence and location of certain wave features, and/or that lends itself to certain types of automated analyses.
High-fidelity digital representations are problematic for a number of reasons. First, they require relatively large amounts of space within which to store the digitized data. Put another way, the higher the fidelity of the digitized data, the larger the amount of storage needed. Another problem with high-fidelity digital representations is that they can result in large amounts of digital data that has little or no import in terms of conveying meaning. For example, a periodic wave signal that merely repeats the same waveform does not convey much meaning to the person analyzing the waveform, and may in fact just take up storage space with unremarkable data. An additional problem is the repeated sampling, over sampling of such high-fidelity data even though it is otherwise unremarkable. Such over sampling results in wasted processing bandwidth (i.e., processor cycles, and/or power) as well as data bandwidth (data storage space and/or transmission bandwidth).
One solution to the above-cited problems associated with a high-fidelity digital representation is to devise a method for filtering and concentrating this high-fidelity waveform data in such a manner that the integrity and significant features of the data are maintained, but while utilizing lower data bandwidth and lower processing bandwidth. Thus, one needs a method and structure that can efficiently and accurately capture the underlying waveform, with little or no degradation of the value and meaning of that waveform data.
Some embodiments of the present invention provide a system and method for representing a quasi-periodic (qp) waveform including: representing a plurality of limited decompositions of the qp waveform, wherein each decomposition includes a first and second amplitude value and at least one time value associated with the first amplitude value. In some embodiments, at least one time value is associated with a second amplitude value, and each of the decompositions is phase adjusted such that the arithmetic sum of the plurality of limited decompositions reconstructs the qp waveform.
In some embodiments, the present invention provides a method for analyzing qp waveforms including an input/analysis process that takes an original qp waveform signal and produces from it a stream of descriptive objects. Within the input/analysis process, a signal-decomposition process decomposes the original quasi-periodic signal into a set of corresponding component signals (herein these are also called “component waves,” “decomposition components,” or simply “components”). The component signals are then processed into a fractional-phase representation to identify the object boundaries and measure basic attributes. The corresponding object-related information is then passed to an object-construction-and-linking process that constructs a stream of objects, filling out whatever additional object-related and/or data-structure-related information is needed for a particular application. An interpretive process is then applied that takes the stream of objects and produces an interpretation of the original qp waveform signal. Within the interpretive process, a state constructor that takes the stream of objects and constructs states from them is applied. Further, an organized mapping process is utilized, where the information is mapped in state space that includes both ad hoc and structured mapping utilizing Hidden Markov Models and Neural Networks. Additionally, an organized mapping process is applied, where the information is mapped in a vector space along one or more object attributes and mapping utilizing vector-space methods, including Hidden Markov Models or Neural Networks. Furthermore, a pattern-recognition discrimination process to transform the mapped information into a human-interpretable form is applied. Then a storage process takes the object stream and samples of this object stream and stores some or all of the samples in a memory. Optionally, a storage process takes the object stream and transmits some or all of the objects over a communications link. In some embodiments, a re-synthesis/output process takes the object stream from a communications link and/or storage and reconstructs an estimate of the original signal(s).
In some embodiments, the method further includes placing the first and second amplitude values, the first time value, and the second time value in a data structure.
Some embodiments further include a computer-readable medium having stored thereon a data structure optionally three or more fields including: a first field containing an object ID, a second field containing a T4 value, a third field containing a phase label, a fourth field containing an amplitude measure, a fifth field containing the total time elapsed, a sixth field containing a relative amplitude value, a seventh field containing a relative T4 value, and an eighth field containing an atomic period value, a ninth field containing a memory address of a left neighbor, a tenth field containing a memory address of a right neighbor, an eleventh field containing a memory address of a left object, a twelfth field containing a memory address of a right object, a thirteenth field containing a left harmonic, a fourteenth field containing a right harmonic, a fifteenth field containing a memory address for an upper neighbor, a sixteenth field containing a memory address for a lower neighbor, a seventeenth field containing data relating to cyclic regularity, an eighteenth field containing data related to a left local period time, a nineteenth field containing data related to a right local period time, a twentieth field containing data related to a left harmonic, a twenty-first field containing data related to a right harmonic, and a twenty-second field related to amplitude modulation.
Other embodiments include a computer-readable medium having executable instructions stored thereon for causing a suitable programmed central processor to perform a method including: taking input in the form of a binary representation of a base waveform, passing this binary representation through a series of bandpass filters, taking the component-signal output of the bandpass filters and extracting data from these components waves, and storing the data into a plurality of data structures. The method further including organizing the data structure. The method further including outputting the data in the plurality of data structures in the form of a graphical representation. The method further including processing the data in the data structure using a Hidden Markov model. The method further including processing the data in the data structure using a Neural Network.
In other embodiments, the invention provides a system for analyzing qp waves including: a database containing data related to qp waves, a CPU operatively coupled to the database. A computer-readable medium operatively coupled to the CPU containing instructions in some CPU-readable format or language, which allows the CPU to: take input in the form of a binary representation of a base waveform, passing this binary representation through a series of bandpass filters, taking a component wave output of the bandpass filters, extracting data from these components waves, and storing the data into a plurality of data structures. Additionally, an input device is operatively coupled to the CPU, an output device is operatively coupled to the CPU, the CPU operatively coupled to an internet, and a terminal for receiving data from the data structures via the internet.
In some embodiments, the invention provides a system for analyzing qp waves. This system includes a set of electrodes each with one end operatively connected to the skin surface or disposed below the skin surface of the body, an optional set of electrodes at least one end of which is operatively connected to the heart itself, a second end of a set of electrodes operatively coupled to an analog-to-digital converter, the analog-to-digital converter operatively coupled to a microprocessor unit, a computer-readable medium containing program instructions operatively coupled to the microprocessors unit, a memory storage operatively coupled to the microprocessor unit, a communications module operatively coupled to the microprocessor unit, and a power supply operatively coupled to the analog-to-digital converter, microprocessor unit and communications module. The system further including: a communications module, a microprocessor unit, a computer-readable medium containing a set of application instructions, an associated storage module containing data values representing the original signal, a Human-Computer Interface by which an operator may interact with the microprocessor unit, an input device, and an output device.
In some embodiments, the invention is implemented using a multiplierless architecture in conjunction with embedded circuits. In still other embodiments, the invention is implemented using firmware, and in still other embodiments the invention is implemented using software.
Although the following detailed description contains many specifics for the purpose of illustration, a person of ordinary skill in the art will appreciate that many variations and alterations to the following details are within the scope of the invention. Accordingly, the following preferred embodiments of the invention are set forth without any loss of generality to, and without imposing limitations upon the claimed invention. Further, in the following detailed description of the preferred embodiments, reference is made to the accompanying drawings that form a part hereof, and in which are shown by way of illustration specific embodiments in which the invention may be practiced. It is understood that other embodiments may be utilized and structural changes may be made without departing from the scope of the present invention.
The leading digit(s) of reference numbers appearing in the Figures generally corresponds to the Figure number in which that component is first introduced, such that the same reference number is used throughout to refer to an identical component that appears in multiple Figures. Signals and connections may be referred to by the same reference number or label, and the actual meaning will be clear from its use in the context of the description.
In some embodiments, the present invention provides a high-fidelity digital representation by using a method for filtering and concentrating high-fidelity waveform data in such a manner that the integrity and significant features of the data are maintained, while utilizing less storage space and fewer processor cycles than conventional techniques. In some embodiments, the present method takes a high-fidelity wave signal (i.e., a qp or the base wave signal), passes this signal through a filter bank that decomposes or separates out this signal into series of locally periodic component signals, and then stores these component signals and their associated characteristics into a data structure. This data structure can then be used to reconstruct the base wave signal, to search for certain features or patterns within the wave signal, or a variety of other purposes. One advantage of the present method is that insignificant or undesirable data can be filtered out and prevented from taking up storage space. In some embodiments, the filters can be adjusted to allow certain wave features to be stored, while other features are disregarded. This ability to adjust filters has the added advantage that when a sample of wave data is taken, it is more likely that the sampling will be of significant wave data, rather than insignificant wave data, thus resulting in fewer processor cycles needed for wave sampling.
Another advantage of this method is that by storing the component wave signals into a data structure, these signals can be easily utilized and manipulated for a variety of purposes. For example, a component wave signal can be analyzed for certain characteristics or patterns. These characteristics or patterns of the component wave signal can then be compared to the other components or the base waveform itself to give a better idea of the significance of the characteristic or pattern.
To begin to understand the use of filter banks in the decomposition of a base wave or original signal it is necessary to understand a base wave signal (i.e., an original signal) as being quasi-periodic (qp). Quasi-periodic means that an original signal may be considered as the linear combination of a series of locally periodic component signals having different periods and/or shapes. Locality here is taken to be on the order of a few cycles of the component in question, though in certain cases can be as small as one cycle, such as, for example, if a certain transient condition to be detected or analyzed sporadically elicits such a component.
Starting with the assumption of a qp waveform, decomposition of the original signal involves the local estimation of the qp components of the original signal. Many methods have been proposed, developed, and are well established in the art. Examples of these are the Short-Time Fourier Transform (STFT), Wavelet Transform, and generally any method that may be cast as a bank of filters (Filter Bank). (See Wavelets and Filter Banks, by Gilbert Strang and Truong Nguyen, Wellesley College, 1996.) The individual filters in the filter bank, herein called component filters or component bands, are designed to emphasize (i.e., correlate to) a locally periodic component of interest in the original signal, and so will tend to have impulse responses that oscillate over a local region of time. The frequency responses of the component filters will tend to have relatively confined regions (passbands) where the magnitude is substantially higher than in other regions. The width of the passband (the bandwidth) is typically measured at the half-power points (i.e., the difference of the two frequencies where the magnitude response is a factor of sqrt(2)/2 below the maximum value in between). A single component filter may have more than one passband, for example if the component is represented by a harmonic series. One may denote the outputs of these component filters as the component outputs, or the component signals.
The design of such filter banks is well understood and established in the art. (See Multirate Digital Signal Processing: Multirate Systems—Filter Banks—Wavelets, by N. J. Fliege, John Wiley & Sons, 2000.) Of specific interest here is that the component filters match, in some useful sense, the underlying local behavior of the original signal. Many approaches are readily available. For example, a high-resolution time-frequency analysis of the original signal could be performed to simultaneously identify frequency bands and time-widths of interest, (i.e., by studying local peaks of signal energy on the time-frequency plane). Alternatively, or in conjunction, an autoregressive moving average (“ARMA”) model could be generated of the original signal. The ARMA model, itself a filter, may be decomposed into component filters using pole and zero separation techniques, such as partial fraction expansion, or any established transformation from cascade to parallel form. Multiple ARMA models could be formed by ARMA modeling each output of a relatively coarse filter bank. Alternatively, or in conjunction, a Karhunen-Loeve Transform (KLT) or a Singular-Value Decomposition (SVD) could be performed on the convolution matrix or correlation matrix estimate formed from a finite time window of the original signal to either directly produce a bank of filters, or to identify filters that correlate to the principal components of the original signal (i.e., a Principal Components Analysis, or PCA). (See Matrix Computations, 3rd Edition, by G. H. Golub, and C. F. Van Loan, Johns Hopkins University Press, 1996.) The identified components of interest could then be used directly or approximated with simpler wavelets. Generally, the decomposition of the original signal may employ any linear or nonlinear method that produces a moving local estimate of the underlying locally periodic components of the original signal.
In some embodiments, prior knowledge of the original wave signal may also be used to define component bands. For example, in some embodiments, an electrocardiogram (“ECG”) signal is expected to have a strong component at the fundamental heart rate. Therefore, for example, in some embodiments, a component is designed having a passband from roughly 0.5 Hz to 5 Hz, to represent a range of expected heart rates.
Generally, the time widths of the impulse responses of the filters in the bank are inversely proportional to the bandwidths of their frequency responses, thus representing a tradeoff in time resolution and frequency resolution. Of primary interest is that the time width be narrow enough to be substantially representative of the duration over which a corresponding underlying locally periodic component will be locally stable, while still maintaining sufficient resolution in frequency. Such tradeoffs are well understood and established in the art, and fall under the class of Spectral Estimation. (See Spectral Analysis of Signals, by Petre Stoica and R. L. Moses, Prentice-Hall, 2005.)
One may define a reconstructed signal, or reconstruction, to be the signal resulting from taking the sum of the component signals at each point in time (i.e., the linear combination of the component signals at each point in time is understood here as a “generalized sum”). In some embodiments, the component filters are designed such that the reconstructed signal is a satisfactory approximation of the original signal. The type and quality of the approximation, and therefore the quality of the signal that can be reconstructed, will depend strongly upon the application, but may be measured, for example, using well-established error-distance measures, such as, for example, mean-square error (MSE). (See Some comments on the minimum mean square error as a criterion of estimation, by C. Radhakrishna Rao, Technical report/University of Pittsburgh-Department of Mathematics and Statistics-Institute for Statistics and Applications, 1980.) The approximation would need qualification, for example, in the case where the original signal was comprised of qp components plus wideband noise, and the reconstruction estimated the components while exhibiting substantially no noise.
Note that in the above sense, the frequency responses of the component bands, when taken as a whole, need not cover the entire spectrum to the highest frequency of interest. Frequency bands may be excluded in which relatively little signal energy exists and/or where a poor signal-to-noise ratio (SNR) exists. Additionally, in some embodiments, the time and/or phase alignment of the component filter impulse responses and/or the component signals is assumed to be understood as part of this invention, for the purpose of producing or representing an adequate reconstruction, including the various well-understood conditions for perfect and approximate reconstruction filter banks (see Wavelets and Subband Coding, M. Vetterli and J. Kovacevic, Prentice-Hall, 1995). The alignment may, for example, take into account the various delays inherent among the impulse responses in the filter bank. Such alignment may also be accounted for in the timing representations embedded in the processing of the component signals, such as in the data structures, by recording any necessary time and/or phase offsets in conjunction with, or added to, any timing and/or phase information for the components. If signal reconstruction is necessary, such timing information may thus be applied within the reconstruction process.
Moreover, in some embodiments, the impulse responses may be linear phase (symmetric/antisymmetric), minimum phase, or dispersive (having asymmetric response but not linear phase), provided that they meet the above signal-reconstruction criteria.
In all of the above filter-bank embodiments, an analytic form may be included, wherein both the above component signals and their Hilbert transforms are produced. This may be accomplished by any of the means well understood and established in the art. (See The Fourier Transform and Its Applications 3rd Edition, by R. Bracewell, McGraw-Hill, 1999.) For example, in some embodiments, the output of each component filter in the bank is provided to the input of a Hilbert transformer, where the component filter output is taken as the real part and the Hilbert transformer output is taken as the imaginary part of a resulting analytic signal. In another embodiment, a second bank of component filters may be provided, each component filter having an impulse response that is the Hilbert transform of its corresponding component filter from the first bank. In either case, an analytic form of each component signal (e.g., an analytic component signal) is formed by taking the component signal of each component filter as the real part and the Hilbert-transformed version of that component signal as the imaginary part. One may note that the STFT inherently produces an analytic form for the components.
The Hilbert transform need not be ideal, only that the frequency-domain phase response of the effective component filter for the imaginary part be offset by substantially 90 degrees from the phase response of the corresponding component filter for the real part, and only over frequency ranges where the frequency response magnitudes are sufficiently high to be of interest (e.g., above some arbitrary threshold, say 20 dB below the peak magnitude). Of course, the frequency response magnitudes of the real-imaginary filter pair should be substantially equal over these frequency ranges, or at least equivalent up to some predetermined level of equivalence, based, for example, upon an error-distance measure or maximum allowable deviation in decibels.
These analytic component signals make available the instantaneous analytic magnitude (i.e., the “magnitude” or “amplitude”) and the instantaneous analytic phase of the component signal. These may be computed, for each point in time of interest, from the magnitude and principal argument of the angle in the complex plane.
The following are example embodiments of the method of decomposition and the filter banks utilized in this invention. In terms of notation, capital letters used for the transfer functions signify frequency-domain transfer functions, but imply their time-domain counterparts as well. The converse also holds true (i.e., lower-case letters signifying the impulse responses of the blocks also imply their corresponding frequency-domain transfer-function counterparts). Additionally, cascading of blocks implies multiplication of their corresponding transfer functions in the frequency domain and/or convolution of their corresponding impulse responses in the time domain, as is well established and understood in the art. (See Discrete-Time Signal Processing, by A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Prentice-Hall, 1999.)
Blocks are considered to be linear, time-invariant systems, and so in a given implementation the blocks include a certain cascade of blocks and may be rearranged in any order within the cascade. These filter banks may be applied as a bank of filter sections arranged in a parallel form known here as a Parallel-Form Kovtun-Ricci Wavelet Transform, and/or these filter banks may be applied as a bank of filter sections arranged in a cascade form known here in as a Cascade-Form Kovtun-Ricci Wavelet Transform. For the purpose of this invention the below-described method of decomposition, filter banks and algorithms utilized in these filter banks, may be implemented in software, embedded circuits (hardware) or firmware.
An example embodiment of basic derivative kernel block Kdp 301 is given by the following discrete-time impulse response description:
kdp(n)=(δ(n)−δ(n−kp))/2
Here the normalization by the factor 2 is a choice for this example to provide a unity-gain passband, but may, in fact, be any arbitrary factor, and where there is defined the discrete-time delta function sometimes called the unit delta or unit impulse:
In some embodiments, kdp is comprised of, besides delay operations, the operations of signed addition and scaling by a factor of two, allowing for, in some embodiments, implementation using a multiplierless architecture. Specifically, it is well understood in the art of digital-signal-processing architectures, that scaling a signal sample by any integer power P of 2 may be readily implemented in digital form with appropriate shifting by P bits the digital bit representation of the signal sample. Thus, by extension through the cascade, scaled-derivative kernel block Hdp may itself, in some embodiments, be implemented using a multiplierless architecture. Further, it is worth noting that the time-domain impulse response hdp of scaled-derivative kernel block Hdp 201 has generally a bell-shaped envelope, particularly evident for Nd greater than or equal to about 4, with alternating-sign coefficients. All nonzero coefficients of hdp have (kp−1) zeros interpolated between them. Even values of Nd result in hdp symmetric about their time centers, while odd values of Nd result in hdp antisymmetric about their time centers. The intrinsic delay of the impulse response hdp, equivalent to the time delay to the time center of hdp, is Ddp=Ndkp/2 samples, and this delay is integer-valued for Nd and/or kp being even. The z-domain transfer function of the scaled-derivative kernel block is:
where, when evaluated on the unit circle (z=ejω), the frequency response is:
The frequency response Hdp(ω) exhibits a series of substantially bell-shaped passbands, particularly for Nd greater than or equal to about 4, centered at ωpq=π(1+2q)/kp radians, or fpq=Fs(1+2q)/(2kp) Hz, where Fs is the sample rate in Hz, and where q=0, 1, 2, . . . , such that the q=0 case corresponds to the fundamental passband centered at ωp0≡ωp (fp0≡fp). The passbands have bandwidths that are inversely proportional to kp and monotonically inversely related to Nd. In the fundamental passband, the frequency response has constant (“flat”) phase equal to Ndπ/2, after accounting for the linear phase term due to the intrinsic delay Ddp.
kip(n)=(u(n)−u(n−w))/wp
where the normalization by the factor wp is a choice for this example to provide a unity-gain passband, but may in fact be any arbitrary factor, and where there is defined the discrete-time unit step function:
In some embodiments, kip is comprised of, besides delay operations, the operations of signed addition and scaling by a factor of wp. In still other embodiments, choosing wp as an integer power of 2 allows for implementation of kip using a multiplierless architecture, and by extension through the cascade, allows scaled-integral kernel block hip to be implemented using a multiplierless architecture.
In some embodiments, the time-domain impulse response of basic integrator-kernel block kip is rectangular in shape, with a width of wp samples and with height wp−1, so that the gain at dc of this block, the sum of the coefficients, is unity. One may further note that, in some embodiments, the time-domain impulse response hip of scaled-integrator kernel block Hip 202 is generally bell-shaped, particularly evident for Ni greater than or equal to about 3. In practice, in some embodiments, wp is set to a value of at least 2 samples, since both wp=0 and wp=1 are degenerate cases resulting in all-zero coefficients and a unit impulse for hip, respectively, regardless of Ni. In some embodiments, impulse responses hip are symmetric about their time center regardless of Ni or wp. The intrinsic delay of the impulse response hip, equivalent to the time delay to the time center of hip, is Dip=Ni(wp−1)/2 samples and this delay is integer-valued for Ni being even and/or wp being odd. The z-domain transfer function of the scaled-integral kernel block is:
where, when evaluated on the unit circle (z=ejω), the frequency response is:
In some embodiments, the amplitude of Hip(ω) exhibits a classic periodic-sinc shaped lowpass response having a mainlobe centered at dc (ω=0) with a peak amplitude of unity, sidelobes of generally decreasing amplitude towards Nyquist (ω=π), and amplitude zeros (nulls) at ωp=2π·r/wp radians, or fp=Fs·r/wp Hz, where Fs is the sample rate in Hz, and where r=1, 2, . . . N. The lowpass mainlobe passband has bandwidth inversely proportional to wp and monotonically inversely related to Ni. With increasing Ni the sidelobes become progressively more attenuated and the overall amplitude response becomes progressively more bell-shaped. The phase response of Hip(ω), after accounting for the linear phase term due to the intrinsic delay Dip, is flat and equal to zero radians in the lowpass mainlobe, and continues to be flat in the sidelobes, alternating between 0 and π radians every other sidelobe for odd values of Ni (i.e., the phase is flat and zero radians everywhere for even values of Ni).
In some embodiments, component filter Hp 106 is composed of the cascade of scaled-derivative kernel block Hdp 201 and scaled-integral kernel block Hip 202, and the frequency response of component filter Hp 106 is the product of the frequency responses of Hdp 201 and Hip 202. In practice, the band-tuning parameters kp and Nd of Hdp 201 are tuned to set the center frequency and bandwidth of the fundamental passband, and by consequence, the harmonic passbands, of component filter Hp 106. In some embodiments, the band-tuning parameters wp and Ni of Hip 202 are tuned to set the degree of suppression needed, if any, in the harmonic passbands. This is typically accomplished by disposing the width of the mainlobe of Hip 202 so that it covers the fundamental passband of Hdp 201, along with the desired number of successive harmonic passbands, if any, and so that the undesired harmonic passbands reside in the region of the nulls and sidelobes of Hip 202. If no harmonic suppression is needed, the block Hip 202 may be eliminated, or equivalently, replaced with a unit impulse (e.g., by setting wp=1).
In some embodiments, only the fundamental passband is kept, and all harmonic passbands are suppressed without loss of generality (WLOG). In some embodiments, a degree of optimality in overall harmonic suppression may be obtained by placing the first null of Hip 202 as close as possible to the peak of the first harmonic passband. In some embodiments, the relation:
where the operator Q[x] quantizes x to an integer, satisfies this constraint. The operator Q[•] may take the form of the ceiling, floor, or rounding operation, and, in some embodiments, in order to satisfy the requirements of a multiplierless architecture, may be made to round to the nearest integer power of 2.
In some embodiments, the family of component filters Hp 106, p=1 . . . N forms a class of wavelets. The degree of orthogonality among the various wavelet scales may be controlled in the frequency domain by determination of the degree of overlap between band responses. Overlap may be reduced by increasing the frequency spacing between the associated passbands, and/or by increasing Nd, and/or by generally suppressing harmonic passbands. Note that in the above examples, the band-tuning parameters kp, Nd, wp, and Ni were implicitly held constant with respect to band index p. These band-tuning parameters may, in fact, be made individually tunable band-by-band, for example in order to optimize orthogonality in a particular filter bank implementation. One may now define the values of delay mp for the delay blocks Dp 107, p=1 . . . N in the parallel filter bank 100 of
mN=0
mp=(DdN+DiN)−(Ddp+Dip),p=1 . . . N−1
where delay block Dp 107 is implemented with impulse response:
dp(n)=δ(n−mp)
In some embodiments, the input to the pth filter section in the bank is applied to the inputs of both block Hp 507 and first delay block Dp1 502. In some embodiments, the output of Hp 507 is applied to the input of second delay block Dp2 503, forming pth component output yp 506. In some embodiments, the output of Hp 507 is also subtracted from the output of Dp1 502, forming pth signal gp 504, called a “residual” or “feed-forward” output signal associated with the pth band. The filter sections are then connected such that the input to the pth filter section is gp-1 505, (i.e., the residual output signal of the previous band, for p=2 . . . N). For p=1, the input of the filter section is the input signal x 501. Hence, in some embodiments, the filter bank features a cascade topology.
In some embodiments, the component filters Hp 507 may be implemented according to the description of the Hp 106 corresponding to the above example parallel filter bank of this invention, with the condition that the coefficients of Hp 507 be normalized so that the corresponding frequency response magnitude peaks are unity, and with the condition that Hp 507 exhibit zero phase in the fundamental passband (after accounting for the linear phase term due to the intrinsic delay of Hp 507). The zero-phase condition may be met by having Nd be a multiple of four, or an appropriate multiple of two with corresponding polarity adjustment of the coefficients for zero phase in the passband of interest. In some embodiments, delay block Dp1 502 would be preferably assigned delay value mp1=Ddp+Dip as per Hp 106, and specifically the intrinsic delay of block Hp 507, where delay block Dp1 502 is implemented with impulse response:
dp1(n)=δ(n−mp1)
With regard to the pth transfer function Tgp defined from gp-1 to gp, the time alignment between delay element Dp1 502 and component filter Hp 507 produces cancellation in the frequency response of Tgp at the frequency-response-magnitude peaks of Hp 507 having zero phase (e.g., the fundamental passband peak). As such, the residual output signal gp contains the information of residual output signal gp-1, except with the significant frequency content (if any) associated with component filter Hp 507 having been substantially removed. (In this sense, original signal x 501 may be considered as residual g0, as it has had no information removed.) With regard to the pth transfer function Typ defined from input x 501 to output yp 506, this arrangement of cascaded residual output signals results in a set of transfer functions Typ, where p=1 . . . N, that are mutually approximately orthogonal.
In some embodiments, by choosing the wavelet scale factors kp so that they are monotonically decreasing with increasing p, the residuals gp and the transfer functions Typ become progressively more lowpass (i.e., have progressively less high-frequency content) due to the information-removal property of cascade filter bank 500 as discussed above. This has the effect of suppressing harmonic passbands present in the derivative kernel block Hdp, reducing or eliminating the need for the harmonic suppression from the scaled integrator kernel Hip in component filter Hp 507. In some embodiments, the final residual output signal gN would thus contain substantially lowpass-only, near-DC information, such as baseline artifacts, and particularly no periodic or qp information, and so would not be used for creation of objects in the context of this invention. The lowpass residual output may, however, be useful for general signal-processing tasks and so is left as part of this invention. One may now define the values of delay mp2 for the second delay elements Dp2 503, p=1 . . . N, in cascade filter bank 500 of
and is generally different for every component filter. As with the parallel filter bank 100 of
where delay block Dp2 503 is implemented with impulse response:
dp2(n)=δ(n−mp2)
An important property of the cascade filter bank is that the original signal x 501 is perfectly recovered from the sum of the outputs and the final lowpass residual, that is:
Partial sums of outputs over a subset of p=1 . . . N may also be used to define wider bands from a combination of narrower ones (“composite bands”), and linear combinations of the outputs of narrower bands may more generally be used to define composite bands of arbitrary shape, for example using the band centers of each transfer function Typ to approximately define a value of frequency-domain magnitude.
In some embodiments, the component signal yp 601 of the corresponding of pth filter section, is processed through a delay element Dhp 605 to form the real part of the analytic component signal yrep 604. In some embodiments, delay element Dhp 605 is set for a delay substantially equal to the intrinsic delay mhp of Hilbert-transformer block Hhp 602 over at least the bandwidth of interest for the pth component. This is useful for systems requiring the intrinsic delay from yp 601 to real party yrep 604 to be the same as that from yp 601 to imaginary party yimp 603. In some embodiments delay element Dhp 605 is implemented with the following impulse response:
dhp(n)=δ(n−mhp)
In some embodiments, basic derivative kernel block Kdp 301 is used for Hilbert-transformer block Hhp 602. In this case, the delay of block Dhp 605 may be set with mhp=kp/2, equivalent to the intrinsic delay of the basic derivative kernel block Kdp 301. This delay is integer-valued for kp even.
The outputs of pth analytic filter bank 600 may then be used to form the pth complex analytic signal:
ycp=yrepjyimp
such that ycp may be generally applied for spectral analysis, instantaneous frequency and amplitude measurement, and related applications well understood and established in the art of analytic signals and communications systems. In particular, regarding the current invention context, in some embodiments, the analytic component signals may be used for defining quarter-phase or fractional-phase objects and attributes. With regard to the pth transfer function Typ defined from input x to output yp 601 of any of the above filter banks, Typ may thus be extended using this example arrangement to analytic transfer functions Tyrep, Tyimp and Tycp, and applied as analytic filters in either the time or frequency domain, again as well understood and established in the art.
In some embodiments the intrinsic delay mhp from component signal yp 601 to imaginary part yimp 603 may depend upon the scale (indexed by p), particularly for Hilbert-transformer block Hhp 602 implemented in real-time form using basic derivative kernel block Kdp 301. In such cases, the intrinsic delay of transfer function Tyimp (and by extension, where applicable, that of Tyrep) would be additively increased over the intrinsic delay of transfer function Typ by the delay mhp. As with the above filter-bank examples, some systems may accommodate the relative delay of the various real and imaginary signal paths of the various bands of the analytic filter bank by accounting for each known delay, for example in the values of the object timestamps in the context of the current invention. Some systems may require the real-time output of each analytic component band to be aligned in time. As is well known, for such systems, the delay of each delay block for each component band may be adjusted so that the intrinsic delay from original signal x to each real party yrep and imaginary party yimp in the analytic bank are substantially the same for all p=1 . . . N. (See Discrete-Time Signal Processing, by A. V. Oppenheim, R. W. Schafer, and J. R. Buck, Prentice-Hall, 1999.) This may be accomplished in the above examples with minimum overall delay by adding the delay-adjustment terms:
mN3=0
mp3=mhN−mhp,p=1 . . . N−1
to the respective delay terms mp, p=1 . . . N for the parallel bank, or respectively to mp2, p=1 . . . N for the cascade bank.
The above-described development of the filter-bank examples does not strictly mention the notions of completeness, including overcompleteness and/or undercompleteness, and sampling, including over sampling and/or sub-sampling, as understood and established in the sense of wavelet transform considerations. With regard to completeness, in some embodiments, decomposition need only substantially represent and/or support the components in the signal that are meaningful to its representation in the sense that the objects produced correspond to desirable feature points on the original signal morphology. In the case of reconstruction of an estimate of the original signal from objects, meaningful components may be those in which the reconstruction substantially represents the original signal morphology to some predetermined degree of accuracy, such as, for example, clinical equivalence for diagnosis by a physician in the ECG example, or as a more general example, some predetermined maximum allowable amount of mean-square error. Thus, undercomplete transforms as well as overcomplete transforms may be readily accommodated in some embodiments.
With regard to sampling, some embodiments of the wavelet transform sub-sample (i.e., they decimate) the individual inputs and/or outputs of the component bands, taking advantage of the fact that the component signals and band filters may carry significantly less information content than the original signal. The resulting sub-sampled filter bank is also called a multirate filter bank. Systems that do not sub-sample are also called oversampled systems. The current invention may be considered a multirate filter bank, in the sense that the fractional-phase objects are sub-samples of their corresponding component signals. However, in this case the sample timings of objects depend on characteristics of the component signals and not simply upon some proportion of the sample period of the original signal. In some embodiments, sub-sampling of one or more of the component bands may take place. In addition, the sub-sampled component signals may be carried forward in the processing system in place of and/or in addition to the non-sub-sampled component signals. In some embodiments, objects based upon such sub-sampled component signals may be formed. One may present an example filter bank employing a single sample rate throughout (i.e., an oversampled filter bank) WLOG.
As to reconstruction, sub-sampled systems rely on interpolation to reconstruct the component signal at the original sample rate, typically using zero-interpolation by the corresponding decimation factor, and then filtering the zero-interpolated result using a resynthesis filter, which is, in some embodiments, defined according to well-established wavelet transform methods. In some embodiments, interpolation and resynthesis may be employed, utilizing the object stream from each component to define impulses, or narrow interpolation kernels, on the original sampling grid. In some embodiments, each impulse would be weighted by its corresponding object amplitude, and centered at the time of the corresponding component peak. Overlap between interpolation kernels, if any, would be summed, sample by sample. The remaining samples would be set to zero. The resulting sequence would then be passed through a reconstruction filter for the corresponding component band, thereby reconstructing the component signal.
In the sense that completeness and sampling fit into the context of a frame in the wavelet transform, frames of any width and height in the time-frequency plane may thus be considered as part of this invention, provided that they meet the criteria given in this invention for signal representation and/or reconstruction.
The output of decomposition is generally a set of component signals having a simple morphology, substantially resembling sine waves, though not necessarily purely sinusoidal. In some embodiments, the various component filters outlined above will be excited to a varying degree by underlying components in the signal that correlate to the impulse response shape of the corresponding component filter, producing a component output (e.g., yp 108, yp 506) that is active. The relative amplitude of an active component will be affected by the degree of correlation and the overall strength of the signal. The states of active or inactive may be discriminated according to amplitude of the component, for example, comparing it to a fixed threshold or adaptive threshold based upon the total time-local energy of the signal.
Some embodiments provide a method to discretely quantify meaningful events in active components, in that these events tend to exhibit strong correspondence to desired feature points along the original signal. Active component signals should generally exhibit regular undulations with readily distinguishable cycles and corresponding cycle features.
In practice, these unique features such as the zero crossings may be readily detected by a change in sign of the signal, while the peaks may be detected by any of various means well known and established in the art, including a corresponding change in sign of the derivative of the signal. Alternatively, in some embodiments, given the availability of analytic signal components, the positive-going and negative-going zero crossings of the imaginary part of the component signal are used to represent the positive and negative peaks of the real part and vice versa. The feature points bound the four quarter-phases and therefore define the transitions of the component signal between and among the four quarter-phases A, B, C, and D. In some embodiments, the component signal is defined as being “in” a particular quarter-phase at any point in time that the component signal satisfies the starting and ending criteria for that quarter-phase. Under those conditions the quarter-phase is defined as “active” for the corresponding component.
These unique features, quarter-phase transitions of the component signals tend to correspond strongly to points of interest (i.e., feature points) along the original signal, such as, for example, local peaks, valleys, inflections, curves, and changes of shape. In some embodiments, this property is utilized for various types of robust analysis. One example application of this property is that of robust delineation (or segmentation) of an ECG waveform into the various waves (e.g., P, Q, R, S, T, U, etc.) and segments between these waves, along with the various applications of such delineation. This property can be utilized in the analysis of most any qp waveform.
In some embodiments, by following the quarter-phase structure, the imaginary part of the analytic signal for a component may be approximated at any given point by evaluating that component signal at a time T4 to either side from that point. The value used for T4 in this evaluation may be the most recent value for that component, or it may be a constant derived from a quarter of the period at the center frequency of the component band.
In some embodiments, T4 may itself be estimated by taking the time between that last two zero-crossings or, if appropriate, the current and prior zero crossings and dividing that value by two.
In some embodiments, these parameters (L, a, T4) are collected, where L is the label of the quarter-phase (i.e., one of A, B, C, or D), and this collection is referred to as a “quarter-phase object” (or “object”) Oij occurring at the ith time index for the jth component. A corresponding notation for the continuous-time case would be Oj(t), where t is continuous time. This object representation therefore creates a compact packet of concentrated information about the component signal and hence, the original signal, precisely tied to a point of interest along the original signal.
In some embodiments, these quarter-phase objects will then be represented as structures or objects in a computer program. In some embodiments, these objects will have data attributes that correspond to the parameters (L, a, T4). The representation of data through the use of objects is well known in the arts of software engineering and computer science. (See Data Structures and Other Objects Using C++ 2nd Edition, by M. Main and W. Savitch, Addison-Wesley, 2001.) The use of the term object implies a meaning as understood by these arts.
In some embodiments, the absolute timing of an object (i.e., the time of occurrence of the object) may be referenced to the beginning time or ending time of its associated quarter-phase. However, in a given system implementation one reference should be chosen and used consistently throughout.
In still other embodiments, over the course of a typical component signal, a stream of such objects will be generated, which may be stored, transmitted, and/or analyzed. The index i may be indexed according to some underlying constant sample rate, in which case i may be considered as a form of time-stamp for the generated object. In some embodiments, index i would be incremented upon generation of each corresponding object. In this case, the absolute time of occurrence for a particular Oij may be stored along with the object, or it may be reconstructed by summing the values T4 for that component from some arbitrary reference time.
In some embodiments, the values T4 may be measured, if necessary, with arbitrary resolution using interpolation methods to determine the zero crossings and/or peaks of the component signals. Interpolation methods to accomplish this are well known and established in the art, including methods such as polynomial fitting and cubic splines in a neighborhood of the zero crossing of the curve or a zero crossing of the derivative of the curve.
Also depicted in
While the component signal in
In some embodiments, the four points of transition between quarter-phases also correspond to the quarter-wave points along the forward transition of the analytic phase of the component (i.e., where the analytic phase substantially equals 0, 90, 180 and 270 degrees, or 0, π/2, π, and 3π/2 radians, respectively). From this perspective, further wave subdivisions are possible, say, defining fractional-phase regions bounded by N fractional phase increments nπ/N, n=0, 1, . . . , N−1 along with their associated objects, labels and parameters. In some embodiments, using N=16, quarter-phase A could be subdivided into four fractional-phase regions A1, A2, A3, and A4, and likewise for quarter-phases B, C, and D, resulting in sixteen fractional-phase regions (i.e., one-sixteenth-phase regions). The particular region the component signal is residing in at any given time would be determined by comparing the analytic phase to the bounding phase values. As before with parameter T4, a parameter TN would be the time length between fractional-phase transitions into and out of the corresponding fractional phase. The amplitude parameter a of the corresponding object may be determined as before, for example by the maximum value, mean value, rms value, or related amplitude measure over the fractional phase, with the added ability to evaluate these based upon a complex-amplitude measure (i.e., the analytic amplitude).
In some embodiments, for a set of Nc components, for quarter-phase object resolution the total number of potential states is 4Nc. For the fractional-phase case with total possible fractional phases F (such that for quarter-phase, F=4), the total number of potential states is thus FNc. For example, with quarter-phases, the system of two components may take on any of the 16 possible states (42), depending upon the relative timing and phase offset of the two components. Again, a system of three components may take on any of the 64 possible states (43). This may extended to the case of any number of components and fractional phases.
In some embodiments, these finer fractional-phase transition points, fractional-phase regions and resulting objects may be suitable for certain applications requiring a finer underlying phase resolution per component, such as situations where the component signal substantially deviates from a sinusoid. However, it can be noted that resolving the analytic phase, in some embodiments, requires evaluation or approximation of the inverse tangent function, which may be prohibitive in certain applications. Further, it can be noted that, in some embodiments, such finer resolution may also be obtained by simply adding and utilizing an appropriate higher-frequency component and its associated quarter-phase objects. Thus, the fractional-phase representation is provided as an embodiment of this current invention, and a direct extension of the quarter-phase representation.
One may further note that, in some embodiments, the complete analytic signal may be utilized. In some embodiments, the fractional phases and transitions of the imaginary part of the analytic signal may be detected and measured, thereby generating objects for the imaginary part in substantially the same manner as already described for the real part. Additionally, timing and amplitude information from the real-part objects may be used for the imaginary-part objects, and vice versa. For example, the real and imaginary parts of the amplitudes may be recorded at all fractional-phase transitions for either or both real and imaginary parts, to establish a moving measure of local energy in the component. Adding the imaginary-part information serves to increase the amount of objects and/or the amount of information per object. This in turn increases the object storage and/or transmission requirements, but can be useful for processing steps involving the objects.
In some embodiments, a special case exists having quarter-phases based upon an analytic component. In this case, zero crossings from both the real part and the imaginary part of the component are used to define the quarter-phase regions, and the complex amplitude information at these zero crossings is used to establish an amplitude measure. This case has particular applicability to multiplierless architectures for creation of quarter-phase objects.
In some embodiments, if the amplitude measure of an object falls below a certain threshold, the object could be flagged, or a special object could be generated, to indicate the absence of the component, whereupon the component and/or object are considered “quiescent.” This is useful for analysis of signals having relatively long quiescent periods in some or all components, such as, for example, ECG signals in the higher-frequency components centered roughly 20 Hz and above.
As the original signal varies or undulates over time, various objects are generated from the corresponding components, and the signal takes on various states, acquiring a new state with each object generated. The sequence of states may then be analyzed as a collection for characterizing the original signal. For example, a purely periodic original signal would produce a perfectly repeating pattern of states. Provided that the fundamental period falls within the passband of one of the components, the substantially lowest-frequency component that is active would produce the same object, for example, an object with the “A1” label, once per fundamental period. If needed, more precision in time would be afforded in this example by detecting particular labels for higher-frequency components synchronous with the “A1” label. Quasi-periodic patterns may also be followed by identifying local cycles of repetition among the states and objects. For example, a certain sequence of states and/or object labels may be identified with an original signal associated with a certain condition. The entire sequence may be given a label according to that condition, and stored and processed simply according to this predetermined label. Next is described how to extend this general idea of linking objects and states, and how to utilize these links.
In still other embodiments, an object will contain a single field value and a single link (i.e., memory address) linking this single-field object to another object. For example, a first such single-field object might contain an Object ID field and a link that links the first single-field object to a second single-field object, which contains only a single T4 value and a link that links the second single-field object to a third single-field object, which contains a phase label and a link that links the third single-field object to yet another single-field object and so on, each single-field object linking to another until some end-of-link. In some embodiments, these objects may contain two data fields (e.g., one for the Object ID and T4 value) and a single link. In some embodiments, these objects may contain three data fields (e.g., one for the Object ID, T4 value, and phase label) and a link to other objects. In some embodiments, these objects might contain four data fields (e.g., one for the Object ID, T4 value, phase label, and atomic period) and a link to another object. In some embodiments, these objects might contain five data fields (e.g., one each for the Object ID, T4 value, phase label, atomic period, and amplitude measure) and a link to another object. In some embodiments, these objects might contain six data fields (e.g., one each for the Object ID, T4 value, phase label, atomic period, amplitude measure, and total time elapsed) and a link to another object. In some embodiments, these objects might contain seven data fields (e.g., one each for the Object ID, T4 value, phase label, atomic period, amplitude measure, total time elapsed and relative amplitude value) and a link to another object.
In still other embodiments, an object will contain a single field value and a single link (i.e., memory address) linking this single-field object to another object. For example, such an object might contain a left object field and a link linking this object to a second single-field object, which in turn contains only a single right object value and a link that links the second single-field object to another single-field object, which in turn contains a left harmonic and a link that links to yet another single-field object, and so on (i.e., a linked list of single-field objects). In some embodiments, these objects may contain two data fields (e.g., left object and right object) and a single link. In some embodiments, these objects may contain three data fields (e.g., left object, right object and left harmonic) and a link to other objects. In some embodiments, these objects might contain four data fields (e.g., left object, right object, left harmonic, and right harmonic) and a link to another object. In some embodiments, these objects might contain five data fields (e.g., left object, right object, left harmonic, right harmonic, and left neighbor) and a link to another object. In some embodiments, these objects might contain six data fields (e.g., left object, right object, left harmonic, right harmonic, left neighbor, right neighbor, and lower neighbors) and a link to another object. Thus, the linked lists may contain like-length objects all having the same number of data fields, or different-length objects having different numbers of data fields.
In some embodiments, the various data-containing attributes and fields would be filled as objects are generated and information becomes available. Information pertaining to objects not yet generated would be filled in as those objects are generated, or at some suitable time after those objects are generated. Because objects from another component may be generated while an object from one component is still active, it is possible that the links to upper neighbors and lower neighbors may require a list of appropriate links to handle the multiplicity of object pairings. In some embodiments, the length of such lists of links is expected to be on the order of the ratio of the periods of the associated paired components, and so should be readily manageable.
In some embodiments, an object has two different types of links. Horizontal links deal with timing relationships and order of occurrence of objects and/or attributes of objects all arising from a certain component. Link types 809, 810, 813, and 814 fall into this category. Vertical links deal with relationships between objects arising from different components, and so link types 815, 816 fall into this category.
Horizontal links are useful for arriving at measures of the local periodic structure of component signals and, by extension, that of the original signal. In some embodiments, such measures arise from the analysis of object attributes in the context of horizontal links. One may call such measures “linked attributes”.
In still other embodiments, an object will contain a single field value and a single link (i.e., memory address) linking this single-field object to another object. For example, a first such single-field object might contain an LLPL field and a link that links the first object to a second single-field object, which in turn contains only a single RLPL value and a link that links to yet another single-field object, which in turn contains an LLPT field and a link that links to still another single-field object and so on. In some embodiments, these objects may contain two data fields (e.g., LLPL, and RLPL) and a single link. In some embodiments, these objects may contain three data fields (e.g., LLPL, RLPL, and LLPT) and a link to other objects. In some embodiments, these objects might contain four data fields (e.g., LLPL, RLPL, LLPT and RLPT) and a link to another object.
Vertically linked objects (VLO) are defined as those objects that are collectively, simultaneously active and linked vertically. In some embodiments, by using the attributes contained in objects and/or by performing calculations using these attributes it is possible to generate or recreate various representations of the component wave signals that served as the basis for the VLO. One such representation is the state matrix and/or the state-transition matrix. Such state-related matrices are well understood in the art, for example, in the area of Hidden Markov Models.
In some embodiments, when representing VLO in a transition matrix it is important to define “a state” as consisting of a vector of objects derived from one or more original signals, and a “state space” as the discrete vector space indexed along each unique dimension by object labels corresponding to a particular component. It is then possible to refer to a trajectory of states as traversing such a state space.
In
In these given example tables or state-transition matrices (i.e.,
As stated above, in some embodiments, a periodic or qp original signal will produce a repeating (or near-repeating) pattern of states. Tracing out the sequence of states with dots and arrows on the table would tend to produce a signature pattern discernable to the human eye. Once such a pattern is understood, deviations from such a pattern, even relatively subtle ones, would also be discernable. Simple sequence detectors may be constructed to match a particular expected sequence of states, or expected set of such sequences, particularly cyclic ones originating (and/or ending) at a particular state or fundamental component object (i.e., fundamental in the sense of the lowest-frequency active component). The object/state representation affords the application of a variety of ad hoc methods for identifying morphologic features or sequences of such features. Any of the object data could be utilized in conjunction with state-pattern data to enhance the robustness of the pattern recognition. For example, the vector of object amplitudes associated with a particular state could be used with a threshold condition to qualify a certain state or pattern identification.
Methods for automatically training on and identifying such patterns and pattern deviations in state space are well understood and established in the art, such as Neural Networks (NN) and Hidden Markov Models (HMM). (See Neural Networks: A Comprehensive Foundation, S. Haykin, Prentice Hall, 1999; Learning Hidden Markov Model Structure for Information Extraction, by Kristie Seymore, Andrew McCallum, and Roni Rosenfeld, American Association for Artificial Intelligence 99, Workshop on Machine Learning for Information Extraction, 1999.) In any case, a training dataset would be presented to the system, for example, in the form of original signals representative of the condition(s) to be modeled, thereby producing a corresponding repository of objects and states. In the case of training with supervised learning algorithms, the object repository would be labeled according to the condition(s). In the case of training with unsupervised learning, no condition labels would be provided and models would be formed by identifying emergent clusters, such as for example with Self-Organizing Maps and Generative Topographic Maps.
In some embodiments, training is established, in the NN example, by converging on a set of weights that correspond to a particular state trajectory pattern (or cluster of trajectory patterns that are “close” in state-space). In the HMM example, training is established by building statistics on states and transitions between states, such as estimating the probability of occurrence of a particular state, and the conditional probability of the transition from one particular state to the next. The state and state-transition matrices may provide the framework for these statistics. While exercising the training dataset, each time a certain state is encountered, the corresponding element of the state matrix would be incremented. The length of time spent in a particular state may also be accumulated, and/or the time from the previous similar transition may be recorded. Accordingly, each time a certain state transition is encountered, the corresponding element of the state matrix would also be incremented. If necessary, these accumulated values may be split into separate accumulation bins, one for each condition label. The statistics can be derived by normalizing by the appropriate count, for example the total number of trials, the total number of trials for a certain condition, the total number of state transitions, and/or the total number of state transitions for a certain trial.
Regardless of the particular method, a test trajectory may be compared for degree-of-match to a trained trajectory pattern or set of patterns. As is known in the art, the degree-of-match measure may be based upon a distance measure in state space, or comparison of the probability of tracing that particular path based upon accumulated statistics.
In some embodiments, the structure of objects, states, and state space in this invention provides a solid framework for building up state models such as those found in HMM. Traditionally, given an original signal, translating points along this signal into useful states for building an HMM has been difficult. Such translation has been typically based on ad hoc methods of delineation of the original signal waveform. The method of this invention provides a natural mapping between discrete states and significant points in the original signal.
For the ECG example, such networks and/or models may be trained using nominal (or normal) data, and then tested with data that may be normal or correspond to a pathologic condition such as ischemia or arrhythmia. The match or absence of match would act as a discriminator of normal and abnormal conditions. Further, training on a set of one or more different abnormal conditions would allow distinguishing among them as well.
Each state naturally carries all of the underlying information associated with the VLO from which the state is comprised. Thus, at each state, all of the attributes, links, and linked attributes including the VLO are available for use. For example, a particular state trajectory or path may map out an ordered set of states {Sk, k=1 . . . NS}, where NS is the total number of states in the path. One may thus construct an NO×NS matrix A of a particular attribute, with the kth column being the NO values of the attribute for the objects including the state Sk. This matrix may then be applied to an augmented HMM or NN for training and testing purposes. One example embodiment would train an HMM directly on states and state transitions only, while training an NN on corresponding attribute matrices for each training trajectory of the HMM. Another example embodiment would use only the information in the attribute matrices, for example to directly train an NN. Such augmented or alternate training could be advantageous in identification or discrimination applications that are more sensitive to information contained in the ordered attributes.
Note that all embodiments may also use more than one attribute in the training and testing, whereupon multiple attribute matrices would be constructed and utilized. One may add to the list of possible attributes the time spent in a particular state, which may be expressed as an absolute time or a percentage of the time for the whole state trajectory, and/or the period or relative frequency of a particular state transition.
A state trajectory may be characterized in some cases by certain states that are inhabited a relatively long time, connected by intermediary states that are inhabited for a relatively short amount of time. Similarly, a state trajectory will generally move more slowly (change less often) along some object dimensions, and at a faster rate along others, according to the fundamental frequencies of the associated component bands. One may thus consider states in their time-context, and so reject states, and/or transitions to and from such states, that represent relatively small time contributions to the state trajectory. This can serve to reduce excess complexity in state trajectories due to relatively rapid chatter between states.
In some embodiments, the base waveform is substantially reconstructed from a linear combination of its associated components. A particular set of quarter-phase objects and their associated attributes are used to reconstruct each component wave, and so in turn each reconstructed component wave is used to reconstruct the base waveform. As described above, in some embodiments, a quarter-phase object contains the attributes of amplitude (a), time (T4), and an associated label (e.g., A, B, C, or D). In some embodiments, there is an object ID associated with each object. If, for example, one considers four objects (A, a1, T1), (B, a2, T2), (C, a3, T3), and (D, a4, T4), one can reconstruct a component waveform along time variable t, over an interval from 0 to T, such that:
Using the above values, one cycle of the original signal of the component form is estimated. The time adjustment among reconstructed components is maintained by sampling time and applying all component reconstructions with appropriate time alignments, deriving an absolute time reference for each component from an accumulation of the T4 values, or using the “total time elapsed” attribute if available to establish absolute timing. The reconstructed components provide the ability to reconstruct the base quasi-periodic waveform by a linear combination of these reconstructed components.
In some embodiments, higher-frequency components will generate more objects within a local time period. For example, the ECG wave will only be generated substantially in the area of the QRS complex over a relatively narrow range of time, as a percentage of time per beat. However, other high-frequency components may be discarded to the degree that the loss of high-frequency information can be tolerated in the reconstruction. Components may be discarded based upon redundancy of the object and the data stored therein, or other features that render the data unremarkable relative to other data sets contained in other objects and components. In some embodiments, a determination of whether or not to discard a component may be based upon a data sampling of these objects and components. The ability to discard some high-frequency components has a number of advantages. First, it allows for less storage space to be used to store the objects. Next, the existence of fewer objects means quicker retrieval times of the requisite data used in reconstruction. Moreover, the existence of fewer objects means that less processor cycles need to be devoted to a particular data set. Once a redundant data is discarded the remaining data may then, in some embodiments, be transferred in some type of compressed format know in the art. (See Data Compression 3rd Edition, by David Salomon, Springer, 2004.)
One example implementation if the present invention is reflected in the below-described calling tree and the associated functions written in pseudo code appearing at each level of the calling tree. For the purpose of brevity and conciseness, below we will provide pseudo-code examples pertaining to the generating of a filter bank (i.e., the GenerateFilter Bank node and associated child nodes), and pseudo-code relating to the generation of quarter-phase representations in object form (i.e., the GetState node and associated child nodes). The calling tree is described by the cells and by the successive top-down connections between cells in the following table:
One knowledgeable in the art would be able to use this calling tree as an architectural blueprint to implement some embodiments of the present invention.
An example pseudo-code representation of a System_Driver function for a software implementation of the present invention is described below. Central to this System_Driver function is a GenerateFilter Bank function (“genFilterBank”), in conjunction with a Time-Frequency Analysis function (“tfAn”), and a GetState (“getState”) function. The genFilterBank implements the analytic wavelets for the above-described Parallel-Form Kovtun-Ricci Wavelet Transform, while the Time-Frequency Analysis performs the transform in block form using the precomputed wavelets from genFilterBank. The getState implements the quarter-phase object and state representation described above. The pseudo-code representation for the System_Driver function is as follows:
One knowledgeable in the art would be able to implement this pseudo-code representation of the System_Driver function and those functions described below on a general-purpose computer using an object-oriented programming language such as C++, C#™, Java™ or Delphi™. Similarly, one knowledgeable in the art would be able to implement this pseudo code on a general-purpose computer using a structured programming language such as Matlab™ or C. Alternatively, one knowledgeable in the art could design an embedded circuit/hardware implementation of this pseudo-code representation.
As outlined above in the calling tree, the System_Driver function, in some embodiments, calls a Generate Filter Bank function (“genFilterBank”). Essential to this genFilterBank function is a Filter Band function (“difilter”) to create individual Kovtun-Ricci wavelets and a Returns Hilbert Transform function (“hilbananorm”) to convert real wavelets into normalized analytic form using a Hilbert transform coupled with normalization in the frequency domain. The pseudo-code representation of the genFilterBank is as follows:
Called within this genFilterBank function are the difilter and hilbananorm functions. The difilter, among other things, returns an “h” value from the Parallel-Form Kovtun-Ricci Wavelet Transform representing a real wavelet, or component filter impulse response. This wavelet is then passed to the hilbananorm function implementing a Hilbert transformer. The following pseudo-code implements the difilter:
Once difilter is implemented, a filter output value or “y” value and intrinsic delay value or “D” stored in a data structure, such as an array, are returned. Function difilter implements an input-output model, such that an input signal “x” is passed in to produce filtered output “y” according to wavelet parameters k, Nd, w, and Ni. On calling difilter, passing in a “1” for “x”, as does genFilterBank, thus returns in output “y” the impulse response of the filter, or the wavelet itself. This output is returned into “h” in genFilterBank, then passed to the hilbananorm function, the pseudo-code for which is described as follows:
In some embodiments, a complex analytic form signal or impulse response “hc” of input signal or impulse response “hr” is returned. In this example, value “hc” is the analytic form of the wavelet. This wavelet is then passed back up to the genFilterBank function and becomes one member of the cell array “cH” (array of wavelets), and then ultimately up to the System_Driver function.
In some embodiments, a Multiple Convolve function (“mconv”) is implemented. This mconv function convolves the input signal (i.e., “x”) with an impulse response “h” multiple times, where the number of times is set by the value “n”. This function implements a multiple cascade of kernels, where kernel “h” would in this example be the impulse response of a basic derivative kernel Kdp 301 or the impulse response of a basic integrator kernel Kip 401 with the order of the kernel set by “Nd” or “Ni”, respectively, depending upon the respective call from difilter, and is reflected in the following code:
The output signal “y” resulting from the multiple convolution is returned, in this example, to difilter.
The System_Driver function, in some embodiments, calls a Time-Frequency Analysis function (“tfAn”). This function performs a wavelet transform of input signal “data” using array of wavelets “cH”, by convolving each wavelet with input signal “data”. Convolution is performed in this example by calling the Linear-Phase convolution function for each wavelet, returning for each wavelet the resulting analytic component signal. In this example the convolution result is truncated to the central samples of the same length as the input “data”, such that a zero-phase FIR filter is effectively realized with the wavelet. The function tfAn takes the convolution results and constructs output matrix “X”. This matrix can be described as an N×M matrix, where N is along the time dimension and M is along the components dimension, such that each column of a matrix “X” is an analytic component signal (i.e., a band output). Matrix “X” is then returned to the System_Driver function.
In addition to the above-described functions and supporting pseudo-code used to generate a filter bank, functionality is implemented to allow for constructing quarter-phase objects or structures from components. As a threshold matter, in some embodiments, a Get State function (“getState”) is called by a System_Driver function. This getState function takes an analytic component signal matrix “X” as a parameter or argument, in this example, as would be returned by the function tfAn. A pseudo-code implementation of getState is as follows:
GetState returns the following outputs: “S”, “cS”, “sObjStrm”, and “sObj”. Output “S” is a matrix representing a stream of state arrays constructed from an object stream obtained from a matrix “X”. Each column of “S” corresponds to a component (i.e., the corresponding column of “X”). Each row of “S” is a state array, here comprised of quarter-phase labels (see discussion of “getObj” below) which are set to zero if the associated component is inactive. A new state array is formed upon the change of active quarter-phase of a component (i.e., when a new object is formed). If more that one object is formed simultaneously, then the state array reflects the change of quarter-phase labels across all affected components. Output “cS” is an array equal to the length of “S” (i.e., the total number of states). Each element of “cS” is an array containing the indices of the individual data structures in the “sObj Strm” array of data structures that result in a corresponding state. As will be discussed below, output “sObj” is an array of structs, with each struct containing a plurality of arrays. Arrays of structures are commonly understood in the art. (See C++ Primer Plus 3rd Edition, by Stephen Prata, Sams Publishing, 1998.) The output “sObjStrm” is an array of structs.
The states obtained via the getState function are defined by objects. In some embodiments, the objects are an array of structs, indexed simply by object index. In some embodiments, the objects are embedded into an array of structs, the array indexed according to the originating component dimension, with each struct of the array of structs containing a plurality of arrays indexed by object index. In some embodiments, a GetObjects function (“getObj”) is described that accepts a matrix “X” as described above, and returns the array of structs, indexed simply by object index, in the above-described “sObjStrm”, and also returns the array of structs, indexed according to originating component dimension, with each struct of the array of structs containing a plurality of arrays indexed by object index, in the above described “sObj”. Pseudo-code representing this getObj function is outlined below:
As described above, the getObj function returns both “sObjStrm” and “sObj”. These return values are then returned to System_Driver by the getState function where they are used in the display of data, in some embodiments, graphically, or used for the purpose of analysis of component signals, original input signals or the like.
Object indices and quarter-phase labels for specific structs or objects are derived from a Find QP function (“findQP”). This function takes an analytic component signal “x” and from this component signal extracts arrays of quarter-phase values labeled as: 1, 2, 3, and 4 (corresponding to the above-described labels A, B, C, and D). The getObj function retrieves this information for each analytic component by calling findQP passing in for “x” each corresponding column of “X”. The code for this findqp function is as follows:
There are three outputs for findQP: “siqp”, “iqps”, and “Lqps”. Output “siqp” is a structure of arrays containing quarter-phase point indices, members arranged by label (i.e., A, B, C, D), each member an array of indices down a column of “X” at which occur qp points of the type corresponding to that label. Output “iqps” is an array of quarter-phase transition indices, for all label types, sorted in order of increasing value (i.e., order of appearance), and output “Lqps” is an array of labels for quarter-phases being entered at corresponding quarter-phase transition points indicated by array “iqps”.
The above-described functions (i.e., System_Driver, GetState, GetObject) use structs and arrays of structs whose fields contain a plurality of arrays to store and manipulate data such as, for example, components, states and the like. As has been described elsewhere these structs, common to many structured programming regimes such as Matlab™ or the C programming language, can be replaced with objects common to an object-oriented programming paradigm such as C++, C#™, Java™ or Delphi™. Additionally, as is discussed above, these structs and/or objects can be organized into not only arrays, but list, stack, queue, tree, heap, hash table or some other structure known in the art. (See Algorithms in C++3rd Edition, by Robert Sedgewick, Addison-Wesley, 1998.) Additionally, any portion of these functions that utilize iterative methods to traverse a data structure, could, in the alternative, use a recursive method. Again, a determination of exact manner in which these structs and/or objects would be organized would be implementation-specific, and hence left up to the particular software developer's efficiency, and speed requirements.
Finally,
Any of the above-described methods, filter banks, or data structures maybe described and implemented using a computer programming language. In some embodiments, the Parallel-Form Kovtun-Ricci Wavelet Transform, and/or Cascade-Form Kovtun-Ricci Wavelet Transform described above can be implemented using the C++ programming language in conjunction with certain programming language libraries such as Microsoft Foundation Class (“MFC”). C++ and MFC are well known in the art of computer science and software engineering. (See Introduction to MFC Programming with Visual C++, by Richard M. Jones, Prentice Hall PTR, 1999.)
In some embodiments, as described below, ECG data from a file serves as input for a program written in C++ and MFC (“the Program”). This Program (a copy of which is filed with U.S. Provisional Patent Application No. 60/656,630 referenced above) utilizes a cascade-form filter bank to divide up the base wave signal or original signal into its component parts. Once divided, each component wave is broken up into quarter-phases, and the data reflecting each quarter-phase is stored into an object, with each object horizontally and vertically linked to other objects. These objects are organized in a linked list structure, a structure well known in the art. Once linked, the data from these objects is retrieved and outputted into a wave representation.
The below matrix depicts the formation of a set of seven composite bands formed from certain sums of band outputs yp taken over contiguous ranges of p. In the Program, a bank of N=250 filter sections is implemented. Within each filter section, the component filter Hp is comprised of only the scaled derivative kernel Hdp (i.e., the scaled integrator kernel Hip is not used, or equivalently, it is set so that wp=1). The Hdp are tuned so that Nd=16 and kp=p for p=1 . . . N.
The component bands for this example implementation are thus the seven composite bands formed from the respective sums of filter sections. The band designations for these component bands are based upon the approximate center frequency of the composite bands resulting from the corresponding ranges of p and assuming a sample rate of 500 Hz. One may note that this Program example does not use all filter section outputs yp, and, in theory, leaves small holes in the continuous spectrum. This is not problematic provided the component band outputs or filter section outputs, or composite band outputs substantially cover the support of the power spectrum of the original signal. An alternate embodiment would be that the sum of the component outputs represents an acceptable approximation of the original signal.
With regard to the structure of the Program itself, the Program contains the following modules and classes in addition to those supplied with MFC: A Readcse.cpp is implemented that reads in ECG from a file. A Leadband class is utilized that performs ECG signal band decomposition for each of the six (6) frequency bands (i.e., designated 1.25, 2.5, 5, 10, 20 and 40 Hz; see
As to the specifics of each of these modules and classes, Readcse.cpp typically reads in binary data from a file. The Leadband class then decomposes this binary data (i.e., a base waveform) into various component waveforms using the cascade-form filter described in
In the present example of the Program, the input file of the ECG data (“D—00004.DCD”) contains 10 seconds of the ECG records for a 12-leads standard ECG. Only one lead V2 has been used in the example. Two output files are generated by the program: d0004—7.wbd and d0004—7.qat. D0004—7.wbd displays the results of the ECG decomposition into six frequency bands from 1.25 to 40 Hz, while d0004—7.gat displays the results of the ECG quarter-phase analysis (i.e., states).
In some embodiments, the above-described filter banks (i.e., parallel and cascading filter banks) can be implemented using a multiplierless architecture. A multiplerless architecture is well known in the art and typically uses a series of adders, such as ripple-carry adders (RCA) or carry-save adders (CSA), to perform multiplication operations. (See Design of Multiplerless, High-Performance, Wavelet Filter Banks with Image Compression Applications, by K. Kotteri, A. Bell, and J. E. Carletta, IEEE Transactions on Circuits and Systems, Vol. 51 No. 3, March 2004; Multiplierless Approximation of Transform with Adder Constraints, by Ying-Jui Chen, IEEE Signal Processing Letters, Vol. 9 No. 11, November 2002.) The advantage of such an architecture is that it is relatively small and as such can be implemented in small devices needing relatively smaller power supplies. In some embodiments, such an architecture would use embedded circuits to implement the functionality of the filter banks, whereas in other embodiments some type of firmware implementation would be utilized to represent the functionality of the filter banks
Such a multiplierless architecture could be designed using VHSIC Hardware Description Language (VHDL). By using VHDL the logic of both the filter and the multiplierless architecture used to create a hardware implementation of a filter could be determined. This logic could then be used to implement an embedded hardware example of the VHDL code. The use of VHDL to design and implement an embedded hardware configuration including filter banks is well known in the art. (See VHDL: Programming By Example 4th Edition, by Douglas L. Perry, McGraw-Hill Professional, 2002.)
As a threshold matter, the above-described method, data structure or program can be used in conjunction with any qp waveform. In some embodiments, this is a qp wave in the form of ECG data, while in other embodiments it might be another type of qp waveform. With regard to qp waveforms derived from ECG data, any system that obtains ECG data from, for example, a set of electrodes could employ the above-outlined method, data structure or program.
The leads 2101 are connected to a sensing device 2103 (“Device”) which is either implanted into the human body or resides external to the human body. The Device 2103 in this example contains a set of units or modules, including a difference-amplifier module, an analog-to-digital converter 2104 (“ADC”), a microprocessor unit 2105 (“MPU”), associated computer-readable medium (“CRM”) 2106 containing program instructions, memory storage 2107 (“Storage”), and a power source for supplying power 2108 (“Power Supply”) to the contained modules. The Device may also contain a communications module 2109 (“Comm”) and an associated communications link 2110. This communication module and link, in some embodiments, may have a wireless means of providing ECG data.
The differential voltage existing across the leads is measured and/or sensed by presenting the leads to a difference amplifier 2102. In some embodiments, the difference amplifier is presumed to contain analog filtering sufficient to substantially meet predetermined requirements for signal-to-noise ratio (“SNR”) optimization and/or anti-aliasing. In this example, the output of the difference amplifier is digitized using an ADC 2104. The output samples from the ADC 2104 may then be used as the original signal for this invention example. This original signal is passed to the MPU 2105 for processing.
Multiple sensing vectors may be accommodated with multiple lead arrangements, as is well understood and established in the art. (See An Introduction to Electrocardiography 7th Edition, by Leo Schamroth, Blackwell Scientific, 1990.) In this case, each differential lead pair could connect to its own processing chain and can be defined here to include a difference amplifier followed by an ADC 2104, producing a corresponding original signal for each sensing vector so processed. Each resulting original signal would then be passed to the MPU 2105 for processing.
The MPU 2105 executes software and/or firmware stored on computer-readable medium 2106 to perform the various tasks necessary for basic function of the Device, and/or to generally fulfill feature and functionality requirements of the Device. The storage of data values representing the original signal and/or the associated objects can, in some embodiments, be in the storage 2107 or the computer-readable medium 2106.
In some embodiments, the communications module 2109 is used to send outgoing transmissions from the Device 2103 to a Host 2201 for further processing, evaluation, storage and/or transmission to other similar or different Hosts on a network. The communications module 2109 is also used to receive incoming transmissions from a Host 2201, which may be the same as or distinct from the Host 2201 used for outgoing transmissions. In some embodiments, incoming transmissions may include commands, data, and/or firmware/software updates as necessary to perform the basic function of the Device 2103, and/or to generally fulfill feature requirements of the Device 2103. For an implanted Device 2103, the connection between the communications module 2109 and a Host 2201 would be typically implemented in wireless fashion, using methods including but not limited to magnetic, capacitive, and/or RF, all well understood and established in the art. (See Principles of Wireless Networks: A Unified Approach 1st Edition, by Kaveh Pahlavan, Prashant Krishnamurthy, Prentice Hall, 2001.) For an external Device, the connection may be wireless or wired using methods also well understood and established in the art.
The Interpretive Process block 2305 takes the object stream and produces an interpretation of the original signal(s). In some embodiments, the State Construction block 2306 takes the object stream and constructs states from it. The resulting series of states, along with the underlying objects by which they are defined, then form the input to the Organized Mapping block 2307, where the information is mapped in state space and/or a vector space along one or more object attributes. As stated in the context of this invention, the information may be mapped directly in some ad hoc manner, for example using a sequence detector on the states and/or some nonlinear, neural and/or fuzzy map formed on the object attributes. In some embodiments, the information may also be used for training a model such as an HMM and/or NN. Once trained, the information may be applied to the model to produce a mapped output. In some embodiments, the mapped information is then passed to a Pattern Recognition, Discrimination and/or Display block 2308 to transform the mapped information into a human-interpretable form, such as for example an automated identification and/or diagnosis of a certain condition, and/or visualization of relevant mapped information. Automated identification could involve simple thresholding on the mapped information, or could use more sophisticated detection and discrimination techniques such as Novelty Detection and Support Vector Machines. (See The Nature of Statistical Learning Theory 2th Edition, by V. Vapnik, Springer 1995; Support-Vector Learning, by C. Cortes and V. Vapnik, 20 Machine Learning 1995.)
The Storage/Transmission block 2309 takes the object stream and/or the original signal samples and/or samples of the component signals (or a predetermined select subset of the component signals) and stores some or all of them in memory and/or transmits some or all of them over a communications link. The Resynthesis/Output Process block 2310 takes the object stream from a communications link and/or storage and reconstructs an estimate of the original signal(s). The object stream corresponding to a desired original signal is recovered from storage and/or received from a communications link via the Retrieval/Reception block 2311. Estimates of the component signals for the desired original signal are then produced by the Component Reconstruction block 2312, and the component signal estimates are then combined in the Signal Reconstruction block 2313 to produce an estimate of the desired original signal. If desired, the individual component signals may be output from the Resynthesis/Output Process block 2310 as well. Multiple original signals may be reconstructed using multiple instances of this Process, once for each desired original signal. The reconstructed signal(s) may then be displayed, for example, on a plot trace (or series of plot traces) for human interpretation, if desired, along with the output of the Interpretive Process block 2305.
In some embodiments, device 2103 performs no signal-analysis processes except for the process of Storage/Transmission. The original signal samples are merely passed to the communications module 2109 of the Device, either directly or perhaps via the MPU 2105, perhaps with some formatting of the data for optimal transmission. The Host 2201 then performs some or all of the remainder of the signal-analysis processes.
In some embodiments, device 2103 performs the Input analysis and the Storage/Transmission Processes, and the Host 2201 performs some or all of the remainder of the processes. In the case that the component signals and/or the original signal are not transmitted to the Host 2201, they may be estimated on the Host end from the transmitted objects via the Resynthesis/Output Process block 2310.
In some embodiments, the signal-analysis processes are performed, in whole or in part, by execution of software or firmware on the MPU 2105 and/or on the MPU 2203, using a program stored in the appropriate respective computer-readable medium 2204 and/or storage unit 2205. Such processing may also be performed, in whole or in part, in dedicated hardware which in this example resides entirely within the MPU 2105 and/or on the MPU 2203, for example in the form of microcode, in the form of a coprocessor, and/or in the form of peripheral hardware.
In some embodiments, the invention provides a method that includes: receiving a series of digital values representing a quasi-periodic waveform, decomposing the series of digital values into a plurality of component signals, wherein each component signal includes a series of digital values that are obtained using a cascade of additive linear combinations of substantially equal-magnitude versions of at least two of the digital values of the quasi-periodic waveform, and storing the component signals.
In some embodiments, the linear combinations use coefficients with equal magnitudes but opposite sign, applied to each of two of the digital values.
In some embodiments, the linear combinations use coefficients with equal magnitudes and signs applied to each of the two or more of the digital values of the quasi-periodic waveform.
In some embodiments, the coefficients are applied to sequential digital values separated from one another by a scaling factor.
In some embodiments, the sequential digital values enter into the linear combination, the number of values determined by a scaling factor.
In some embodiments, the series of values for each component signal represents a digitally filtered frequency-limited band of the quasi-periodic waveform.
In some embodiments, the digitally filtered frequency-limited band is selected based on frequency content of the quasi-periodic waveform.
In some embodiments, the series of values for each component signal include complex numbers and represent an analytic signal.
In some embodiments, the quasi-periodic waveform includes a time-varying physiological signal obtained from a mammal, and the series of digital values for each component signal represents a digitally filtered frequency-limited band of the quasi-periodic waveform.
In some embodiments, the quasi-periodic waveform includes an electrocardiogram (ECG) signal from a human, and the series of digital values for each component signal represents a digitally filtered frequency-limited band of the ECG signal.
Some embodiments further include transforming each of the component signals to obtain: a plurality of independent ordinate values, and a plurality of abscissa values each associated with a corresponding ordinate value. Each ordinate value is based on an amplitude metric. Each abscissa value is based on a time metric.
In some embodiments, transforming of each component signal further includes: a phase adjustment value such that a linear combination of the component signals substantially reconstructs the quasi-periodic waveform.
Some embodiments further include transforming of each component signal further includes: an amplitude adjustment value such that a linear combination of the component signals substantially reconstructs the quasi-periodic waveform.
Some embodiments further include transforming of each component signal uses a multiplierless architecture.
In some embodiments, the component signals are obtained by applying a Parallel-Form Kovtun-Ricci Wavelet Transform to the series of digital values.
In some embodiments, the component signals are obtained by applying a Cascade-Form Kovtun-Ricci Wavelet Transform to the series of digital values.
In some embodiments of the invention, a computer-readable medium is employed having executable instructions stored thereon for causing a suitable programmed central processing unit to perform a method that includes: obtaining a series of digital values representing a quasi-periodic waveform, decomposing the series of digital values into a plurality of component signals, wherein each component signal includes a series of digital values that are obtained using a cascade of additive linear combinations of substantially equal-magnitude versions of at least two of the digital values, and storing the component signals.
In some embodiments, the computer-readable medium includes further instructions configured for the linear combination to use coefficients with equal magnitudes but opposite sign, applied to each of two of the digital values.
In some embodiments, the computer-readable medium includes further instructions configured for the linear combination to use coefficients with equal magnitude and sign applied to each of the two or more of the digital values.
In some embodiments, the computer-readable medium includes further instructions configured for the coefficients to be applied to sequential digital values separated from one another by a scaling factor.
In some embodiments, the computer-readable medium includes further instructions configured for the sequential digital values to be entered into the linear combination, with the number of values determined by a scaling factor.
In some embodiments, the computer-readable medium includes further instructions configured for the series of values for each component signal to represent a digitally filtered frequency-limited band of the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for the digitally filtered frequency-limited band to be chosen corresponding to a frequency content of the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for the series of values for each component signal to represent an analytic signal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform to be a time-varying physiological signal obtained from a mammal, and the series of digital values for each component signal represent a digitally filtered frequency-limited band of the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform to be an electrocardiogram (ECG) signal from a human, and the series of digital values for each component signal represent a digitally filtered frequency-limited band of the ECG signal.
In some embodiments, the computer-readable medium includes further instructions for transforming each of the component signals to obtain: a plurality of independent ordinate values, and a plurality of abscissa values each associated with a corresponding ordinate value.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to be based on an amplitude metric.
In some embodiments, the computer-readable medium includes further instructions configured for each abscissa value to be based on a time metric.
In some embodiments, the computer-readable medium includes further instructions configured for the transforming of each component signal including: a phase adjustment value such that a linear combination of the component signals substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for the transforming of each component signal including: an amplitude adjustment value such that a linear combination of the component signals substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions for transforming of each component signals using a multiplierless architecture.
In some embodiments, the computer-readable medium includes further instructions configured for the component signals to be obtained by applying a Parallel-Form Kovtun-Ricci Wavelet Transform to the series of digital values.
In some embodiments, the computer-readable medium includes further instructions configured for the component signals to be obtained by applying a Cascade-Form Kovtun-Ricci Wavelet Transform to the series of digital values.
In some embodiments, the invention provides a method that includes obtaining a plurality of component signals representative of a quasi-periodic waveform, wherein each component signal includes a plurality of component values that represent a digitally filtered frequency-limited band of the quasi-periodic waveform, and transforming each of the plurality of component values into a fractional-phase representation including values representing at least one of an abscissa and an ordinate for each of a plurality of fractional phases per local cycle of the component signal, and storing the fractional-phase representation. In some embodiments, each ordinate value is based on an amplitude metric. In some embodiments, each abscissa value is based on a time metric.
In some embodiments, each fractional-phase representation includes a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, each fractional-phase representation includes an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, wherein: the fractional-phase representation for the first fractional phase includes a value representing an abscissa associated with a zero crossing of the associated component signal, the fractional-phase representation for the second fractional phase includes a value representing an ordinate and a value representing an abscissa both values representing an abscissa and values representing an ordinate associated with a maximum of the associated component signal, the fractional-phase representation for the third fractional phase includes a value representing an abscissa associated with another zero crossing of the associated component signal, and the fractional-phase representation for the fourth fractional phase includes a value representing an ordinate and a value representing an abscissa both associated with a minimum of the associated component signal. Each ordinate value is based on an amplitude metric. Each abscissa value is based on a time metric. Each fractional-phase representation includes a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform. Each fractional-phase representation includes an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the plurality of component values for each component signal are complex-valued, wherein the fractional-phase representation for each fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with an angular range of a complex argument of an associated analytic component signal.
In some embodiments, each ordinate value includes an amplitude metric selected from a group consisting of: a complex magnitude, a real part, and an imaginary part of the analytic component signal.
In some embodiments, each ordinate value includes an amplitude metric based upon: the complex magnitude of the analytic component signal, the real part of the analytic component signal, and the imaginary part of the analytic component signal.
In some embodiments, each abscissa value is based on a time metric.
In some embodiments, each fractional phase representation includes a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, each fractional-phase representation includes an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the number of fractional phases available per local cycle of each component signal is a predetermined number for that component, and wherein the angular ranges for the phases correspond to subdivisions of the angular range of the complex argument.
In some embodiments, each ordinate value is based upon an amplitude metric selected from a group consisting of: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, each ordinate value is based upon an amplitude metric including: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal. In some embodiments, each abscissa value is based upon a time metric.
In some embodiments, each fractional-phase representation includes a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, each fractional-phase representation includes an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the number of fractional phases available per local cycle of the component signal is four including: a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein: the fractional-phase representation for the first fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with a zero crossing of a real part of the associated analytic component signal, the fractional-phase representation for the second fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with a zero crossing of an imaginary part of the associated analytic component signal, the fractional-phase representation for the third fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with another zero crossing of the real part of the associated analytic component signal, and the fractional-phase representation for the fourth fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with another zero crossing of the imaginary part of the associated analytic component signal.
In some embodiments, each ordinate value includes an amplitude metric selected from a group consisting of: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, each ordinate value includes an amplitude metric consisting of: the complex magnitude of the analytic component signal, the real part of the analytic component signal, and the imaginary part of the analytic component signal.
In some embodiments, each abscissa value is based upon a time metric.
In some embodiments, each fractional-phase representation includes a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, each fractional-phase representation includes an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the component signal is an electrocardiogram (ECG) signal from a human.
In some embodiments, the zero crossing is evaluated by finding a change in sign.
In some embodiments, the abscissa for each zero crossing is evaluated by performing an interpolation in a neighborhood of the sign change.
In some embodiments, the interpolation is a curve fit and the zero crossing is evaluated by solving for a root.
In some embodiments, the amplitude metric is obtained from a quarter-phase.
In some embodiments, the invention provides a computer-readable medium having executable instructions stored thereon for causing a suitable programmed central processing unit to perform a method that includes: obtaining a plurality of component signals representative of a quasi-periodic waveform, wherein each component signal includes a plurality of component values that represent a digitally filtered frequency-limited band of the quasi-periodic waveform, and transforming each of the plurality of component values into a fractional-phase representation including values representing an abscissa, an ordinate, or both values representing an abscissa and values representing an ordinate for each of a plurality of fractional phases per local cycle of the component signal, and storing the fractional-phase representation in a storage device. In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to be based on an amplitude metric. In some embodiments, the computer-readable media allowing each abscissa value to be based on a time metric.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for a fractional-phase representation to include an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments of the invention, the computer-readable medium includes further instructions configured for the component signal to include a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and includes further instructions configured such that: the fractional-phase representation for the first fractional phase includes a value representing an abscissa associated with a zero crossing of the associated component signal, the fractional-phase representation for the second fractional phase includes a value representing an ordinate and a value representing an abscissa both values representing an abscissa and values representing an ordinate associated with a maximum of the associated component signal, the fractional-phase representation for the third fractional phase includes a value representing an abscissa associated with another zero crossing of the associated component signal, and the fractional-phase representation for the fourth fractional phase includes a value representing an ordinate and a value representing an abscissa both associated with a minimum of the associated component signal.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to be based on an amplitude metric.
In some embodiments, the computer-readable medium includes further instructions configured for each abscissa value to be based on a time metric.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for component signals that are analytic such that the plurality of component values for each component signal are complex-valued, and wherein the fractional-phase representation for each fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with an angular range of a complex argument of an associated analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to comprise an amplitude metric selected from a group consisting of: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to comprise an amplitude metric based upon: the complex magnitude of the analytic component signal, the real part of the analytic component signal, and the imaginary part of the analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for each abscissa value to be based on a time metric.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for the number of fractional phases available per local cycle of each component signal to be a predetermined number for that component, and wherein the angular ranges for the fractional phases correspond to subdivisions of the angular range of the complex argument.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to be based upon an amplitude metric selected from a group consisting of: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for ordinate value to be based upon an amplitude metric including: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for each abscissa value to be based upon a time metric.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for the component signal to include a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and also includes further instructions configured for: the fractional-phase representation for the first fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with a zero crossing of a real part of the associated analytic component signal, the fractional-phase representation for the second fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with a zero crossing of an imaginary part of the associated analytic component signal, the fractional-phase representation for the third fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with another zero crossing of the real part of the associated analytic component signal, and the fractional-phase representation for the fourth fractional phase includes a value representing an ordinate, a value representing an abscissa, or both values representing an abscissa and values representing an ordinate associated with another zero crossing of the imaginary part of the associated analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to include an amplitude metric selected from a group consisting of: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value to include an amplitude metric consisting of: the complex magnitude of the analytic component signal, the real part of the analytic component signal, and the imaginary part of the analytic component signal,
In some embodiments, the computer-readable medium includes further instructions configured for each abscissa value to be based upon a time metric.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include a phase adjustment value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for each fractional-phase representation to include an amplitude normalization value such that a linear combination of the plurality of component values substantially reconstructs the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform to be an electrocardiogram (ECG) signal from a human.
In some embodiments, the computer-readable media is configured for the zero crossing to be evaluated by finding a change in sign.
In some embodiments, the computer-readable medium includes further instructions configured for the zero crossing to be evaluated by finding a change in sign.
In some embodiments, the computer-readable medium includes further instructions configured for the zero crossing to be evaluated by performing interpolation in the neighborhood of the sign change.
In some embodiments, the computer-readable medium includes further instructions configured for the interpolation to be a curve fit and the zero crossing is evaluated by solving for a root.
In some embodiments, the computer-readable medium includes further instructions configured for the amplitude metric to be obtained from a quarter-phase.
In some embodiments, the invention provides a computer-readable medium having executable instructions stored thereon for causing a suitable programmed information processor to perform a method that includes: obtaining a plurality of fractional-phase representations of a signal, wherein each fractional-phase representation includes values representing an abscissa, an ordinate, or both values representing an abscissa and values representing an ordinate for each of a plurality of fractional phases of a component of a plurality of components that each represent a digitally filtered frequency-limited band of a quasi-periodic waveform, and assigning to each fractional-phase representation a phase label associated with an active fractional phase of the fractional-phase representation, generating a plurality of data structures, each data structure arising from each component associated with a different abscissa value and including a phase label, and storing the plurality of data structures in a storage device.
In some embodiments, the computer-readable medium includes further instructions configured for each ordinate value is based upon an amplitude metric.
In some embodiments, the computer-readable medium includes further instructions configured for the associated component to be an analytic signal and each ordinate value is derived from an amplitude metric selected from a group consisting of: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, the computer-readable medium includes further instructions configured for the associated component to be an analytic signal and each ordinate value is derived from an amplitude metric based upon: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments, the computer-readable media allows the ordinate to be based upon a statistical measure.
In some embodiments, the computer-readable medium includes further instructions configured for each of the abscissa values to be based upon a time metric.
Some embodiments of the invention include a computer-readable media having stored thereon a first data structure that includes: a descriptor that includes an abscissa value from an associated fractional-phase representation, an optional descriptor that includes an ordinate value from the associated fractional-phase representation, an optional descriptor that includes an abscissa interval value representing an interval of the abscissa over which an associated fractional phase is active, an optional descriptor that includes a label identifying a particular data structure from the plurality of data structures, an optional descriptor that includes a label identifying an associated component from the plurality of components, an optional descriptor that includes a label identifying the quasi-periodic waveform, an optional descriptor that includes a value representing a local period of the associated component, an optional descriptor that includes a value representing an interval from a reference abscissa value to an associated abscissa value, an optional descriptor that includes a value representing a relative measure between the ordinate value associated with the fractional phase, and an ordinate value associated with the component, the quasi-periodic waveform, or both the component and the quasi-periodic waveform, and an optional descriptor that includes a label indicating that the component is active for the active fractional phase associated with the particular data structure, based upon comparing the ordinate value to a predetermined threshold.
In some embodiments of the computer-readable medium having the data structure stored thereon, each ordinate value is based upon an amplitude metric.
In some embodiments of the computer-readable medium having the data structure stored thereon, the associated component is an analytic signal and each ordinate value is derived from an amplitude metric selected from a group consisting of: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments of the computer-readable medium having the data structure stored thereon, the associated component is an analytic signal and each ordinate value is derived from an amplitude metric based upon: a complex magnitude of the analytic component signal, a real part of the analytic component signal, and an imaginary part of the analytic component signal.
In some embodiments of the computer-readable medium having the data structure stored thereon, the abscissa values are based upon a time metric.
Some embodiments of the computer-readable medium having the data structure stored thereon include an atomic period value, wherein the atomic period value is the abscissa interval value multiplied by a number of fractional phases available in the fractional-phase representation.
In some embodiments of the computer-readable medium having the data structure stored thereon, the data structure further includes links to other data structures, including: a descriptor that includes a link to another particular data structure associated with the component, wherein the associated abscissa value of the other data structure is less than the abscissa value associated in the particular data structure, an optional descriptor that includes a link to another particular data structure associated with the component, wherein the abscissa value of the other data structure is greater than the abscissa value in the particular data structure, an optional descriptor that includes a link to another particular data structure associated with the component, wherein the abscissa value of the other data structure is less than the abscissa value in the particular data structure, and wherein the phase label of the other data structure is of the same value as the phase label of the particular data structure, and an optional descriptor that includes a link to another particular data structure associated with the component, wherein the abscissa value of the other data structure is greater than the abscissa value in the particular data structure, and wherein the phase label of the other data structure is of the same value as the phase label of the particular data structure.
In some embodiments of the computer-readable medium having the data structure stored thereon, the data structure further includes: a descriptor that includes a difference of abscissa values between the abscissa value included in the particular data structure and the abscissa value included in a first linked data structure relative to this particular data structure, an optional descriptor that includes a difference of abscissa values between the abscissa value included in the particular data structure and the abscissa value included in a second linked data structure relative to this particular data structure, an optional descriptor that includes an indication of deviation from an expected sequence of phase labels, and an optional descriptor that includes a moving average of abscissa values for a group of data structures surrounding the particular data structure.
In some embodiments of the computer-readable medium having the data structure stored thereon, the data structure is organized within a plurality of other data structures into an array based upon a common associated component.
In some embodiments of the computer-readable medium having the data structure stored thereon, the data structure further includes a label to the data structure, wherein a label describes a transition of the component to and from an active condition.
In some embodiments of the computer-readable medium having the data structure stored thereon, the data structure further includes a first phase label associated with a first fractional phase of a given component, a second phase label associated with a second fractional phase of the given component, a third phase label associated with a third fractional phase of the given component, and a fourth phase label associated with a fourth fractional phase of the given component.
Some embodiments of the invention include a computer-readable medium having the data structure stored thereon, wherein the data structure further includes: a descriptor including an abscissa interval value representing the interval of the abscissa over which the associated fractional phase is active to the first data structure, and a plurality of optional descriptors including at least three selected from the set comprising: an optional descriptor that identifies the first data structure from the plurality of data structures, an optional descriptor identifying the associated component from the plurality of components, an optional descriptor assigning a label identifying the quasi-periodic waveform, an optional descriptor assigning a value to representing the local period of the associated component, an optional descriptor assigning a referenced abscissa value representing the interval from a reference abscissa value to the associated abscissa value, an optional descriptor assigning a link to another data structure associated with the component, wherein the included abscissa value of the other data structure is less than the abscissa value, an optional descriptor assigning a link to another data structure associated with the component, wherein the included abscissa value of the other data structure is greater than the abscissa value included in the first data structure, an optional descriptor assigning to the first data structure a link to a second data structure associated with the component, wherein the second data structure is an upper neighbor, an optional descriptor assigning to the first data structure a link to a third data structure associated with the component, wherein the third data structure is a lower neighbor, an optional descriptor assigning to the first data structure a relative T value, an optional descriptor assigning to the first data structure a left harmonic value, an optional descriptor assigning to the first data structure a right harmonic value, an optional descriptor assigning to the first data structure an amplitude measurement, an optional descriptor assigning to the first data structure a phase label, an optional descriptor assigning to the first data structure a left object, an optional descriptor assigning to the first data structure a right object, and an optional descriptor assigning to the first data structure a relative amplitude value.
Some embodiments of the computer-readable medium further include a plurality of data structures organized as a list based upon a common associated component.
Some embodiments of the computer-readable medium further include a plurality of data structures organized as a tree based upon a common associated component.
Some embodiments of the computer-readable medium further include a plurality of data structures organized as a hash table based upon a common associated component.
Some embodiments of the computer-readable medium further include a plurality of data structures organized as a heap based upon a common associated component.
Some embodiments of the computer-readable medium further include a plurality of data structures organized as an array based upon a common associated component.
Some embodiments of the invention provide a method that includes: obtaining one or more data structures, organizing the one or more data structures into an arranged data structure, wherein the one or more arranged data structures are arranged according to their associated component, and the fractional phases associated with the one or more data structures within each arranged data structure are co-active over a range of abscissa values for an associated quasi-periodic waveform, generating a plurality of states, each state including a group of one or more arranged data structures, and storing the states in a storage device.
In some embodiments, the plurality of states include a state with a reference abscissa value associated with the abscissa values of the arranged data structure, and assigned to the state, and the plurality of states are arranged into a state sequence, according to the associated abscissa values.
In some embodiments, the reference abscissa value is the greatest of the abscissa values of the arranged data structure assigned to the corresponding state.
In some embodiments of the method, a new state is formed upon a change of an active fractional phase of one or more of the arranged data structures in the state, and forming a sequence of states.
In some embodiments, the method further includes: analyzing a state trajectory in state space, and identifying a set of features within a quasi-periodic waveform, a component signal, or both the quasi-periodic waveform and the component signals based upon the state space.
In some embodiments of the method, the quasi-periodic waveform signal is an electrocardiogram (ECG) signal from a mammal, with features including: a P wave, a Q wave, an R wave, an S wave, a T wave, a U wave, and any ECG segment delineated by any pair of these waves.
In some embodiments of the method, the quasi-periodic waveform signal is an electrocardiogram (ECG) signal from a mammal, with features selected from a group consisting of: a P wave, a Q wave, an R wave, an S wave, a T wave, a U wave, and any ECG segment delineated by any pair of these waves.
In some embodiments of the method, the quasi-periodic waveform signal is a heart-sound signal from a mammal, and the features include: an S1, an S2, and an S3 sound.
In some embodiments of the method, the quasi-periodic waveform signal is a heart-sound signal from a mammal and includes features selected from a group consisting of: an S1, an S2, and an S3 sound.
In some embodiments, the quasi-periodic waveform signal is a blood-pressure signal from a mammal, and the features include a dicrotic notch.
In some embodiments, the quasi-periodic waveform signal is a blood-flow signal from a mammal. In some embodiments, the quasi-periodic waveform signal is an electroencephalogram (EEG) signal from a mammal. In some embodiments, the quasi-periodic waveform signal is an electromyogram (EMG) signal from a mammal. In some embodiments, the quasi-periodic waveform signal is an electrooculogram (EOG) signal from a mammal. In some embodiments, the quasi-periodic waveform signal is a respiratory signal from a mammal, wherein the respiratory signal an optional descriptor includes: pressure, flow, and volume information.
Some embodiments of the method further include: analyzing a state trajectory in state space, and identifying a condition subset from a predetermined set of conditions associated with a quasi-periodic waveform, a component signal, or both the quasi-periodic waveform and the component signals based upon the state space analysis.
In some embodiments, the quasi-periodic waveform signal is an electrocardiogram (ECG) signal from a mammal. In some embodiments, the quasi-periodic waveform signal is a heart-sound signal from a mammal. In some embodiments, the quasi-periodic waveform signal is a blood-pressure signal from a mammal. In some embodiments, the quasi-periodic waveform signal is a blood-flow signal from a mammal. In some embodiments, the quasi-periodic waveform signal is an EEG, EMG, or EOG signal from a mammal. In some embodiments, the quasi-periodic waveform signal is a respiratory signal from a mammal, wherein the respiratory signal an optional descriptor includes pressure, flow, and volume information. In some embodiments, the quasi-periodic waveform signal is derived from a mammal and a condition set include an optional descriptor the condition normal, and an optional descriptor the condition abnormal.
In some embodiments of the method, the condition set includes cardiac conditions, the normal condition is normal sinus rhythm, and the abnormal condition is a set of rhythm abnormalities including: an optional descriptor bradycardia and tachycardia abnormality, and an optional descriptor conduction abnormality, wherein bradycardia is an optional descriptor set including sick sinus syndrome, and tachycardia is an optional descriptor set including atrial flutter, atrial fibrillation, ventricular tachycardia, and ventricular fibrillation, and conduction abnormality is an optional descriptor set to include bundle branch block, AV and node block.
In some embodiments, the method wherein: the condition set includes cardiac conditions, the normal condition is normal sinus rhythm, and the abnormal condition is a set of structural abnormalities including: optional ischemia, optional myocardial infarction, optional left ventricular hypertrophy, optional right ventricular hypertrophy, optional bi-ventricular hypertrophy, optional long QT syndrome, and optional tall T syndrome.
In some embodiments, the set conditions are normal blood pressure and abnormal blood pressure. In some embodiments, the method wherein the set conditions include hypotension and hypertension.
Some embodiments of the invention provide a method that includes: obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, wherein each quasi-periodic waveform corresponds to a predetermined set of feature types and assigned by feature type in a known manner to a set of states forming a state subsequence of the associated state sequence, constructing a sequence detector for each state subsequence resulting from a training set, such that detection of the state subsequence is an indication of the assigned feature type, and identifying the feature type associated with each feature of an arbitrary test quasi-periodic waveform, based upon a pattern-matching operation.
In some embodiments, the pattern-matching operation includes: establishing a degree of match measure between the sequences associated with each detector and the state sequence resulting from the test waveform, and forming a fuzzy inference upon these measures through a fuzzy logic mapping.
In some embodiments, the pattern-matching operation includes: establishing a degree of match measure between the state sequence associated with each detector and the state sequence resulting from the test waveform, and applying a neural network mapping to these measures, wherein the neural network is trained on the training set.
In some embodiments, the method further includes: obtaining the state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, each waveform assigned in a known manner to a subset of a set of one or more predetermined conditions, constructing a sequence detector for each state sequence resulting from the training set, such that detection of the state sequence is an indication of the assigned condition subset, and identifying the condition subset associated with an arbitrary test quasi-periodic waveform, based upon a pattern-matching operation.
In some embodiments, the pattern-matching operation includes: establishing a degree of match measure between the sequences associated with each detector and the state sequence resulting from the test waveform, and forming a fuzzy inference upon these measures through a fuzzy logic mapping.
In some embodiments, the pattern-matching operation includes: establishing a degree of match measure between the sequence associated with each detector and the state sequence resulting from the test waveform, and applying a neural network mapping to these measures, wherein the neural network is trained on the training set.
In some embodiments, the method further comprising: obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, each feature of a set of features in each waveform corresponding to a predetermined set of feature types and assigned by feature type in a known manner to a set of states forming a subsequence of the associated state sequence, constructing a model according to a span of a state vector or a subspace of that span in state space, training the model based upon the state subsequences resulting from a training data set, including associated feature sets as labels for supervised learning, and identifying a feature type associated with each feature of an arbitrary test quasi-periodic waveform, based upon statistics in the trained model in relation to the state subsequences resulting from test waveform features.
In some embodiments, the model has multiple modalities. In some embodiments, the model includes a plurality of submodels. In some embodiments, the submodel is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping. In some embodiments, the model is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping.
Some embodiments further include obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, each waveform assigned in a known manner to a subset of a set of one or more predetermined conditions, constructing a model according to a span of a state vector or a subspace of that span in state space, training the model based upon the state sequences resulting from a training data set, including associated condition sets as labels for supervised learning, and identifying the condition subset associated with an arbitrary test quasi-periodic waveform, based upon statistics in the trained model in relation to the state sequence resulting from a test waveform.
In some embodiments, the model has multiple modalities. In some embodiments, the model includes a plurality of submodels. In some embodiments, the submodel is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping. In some embodiments, the model is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the invention includes obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, waveforms in aggregation containing an unknown set of features generally falling into groups, these features each having an associated set of states forming a subsequence of the associated state sequence, constructing a model according to a span of a state vector or a subspace of that span in state space, training the model based upon the state subsequences resulting from the training data set, allowing for self-organization of the clusters in state-space, assigning labels to the emergent clusters and associating them with corresponding feature types on the quasi-periodic waveform, and identifying the feature type associated with each feature of an arbitrary test quasi-periodic waveform, based upon statistics in the trained model in relation to the state subsequences resulting from test waveform features.
In some embodiments, the model has multiple modalities. In some embodiments, the model includes a plurality of submodels. In some embodiments, the submodel is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping. In some embodiments, the model is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the invention includes obtaining the state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, the waveforms in aggregation pertaining to one or more unknown conditions of a set of conditions of unknown size, these waveforms each having an associated state sequence, constructing a model according to the span of the state vector or a subspace of that span in state space, training the model based upon the state sequences resulting from the training data set, allowing for self-organization of the clusters in state-space, and assigning labels to the emergent clusters and associating them with condition types, and identifying the set of condition types associated with an arbitrary test quasi-periodic waveform, based upon the statistics in the trained model in relation to the state sequence resulting from the test waveform. In some embodiments, the model has multiple modalities. In some embodiments, the model includes a plurality of submodels. In some embodiments, the submodel is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping. In some embodiments, the model is a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping. In some embodiments, the states are pruned from the state trajectory based upon the relative time of habitation within the states. In some embodiments, the states are pruned from the state trajectory based upon the relative time of habitation within the states.
Some embodiments of the invention provide a computer-readable medium having executable instructions stored thereon for causing a suitable programmed central processing unit to perform a method that includes: obtaining one or more data structures, organizing the one or more data structures into an arranged data structure, wherein the one or more arranged data structures are arranged according to their associated component, and the fractional phases associated with the one or more data structures within each arranged data structure are co-active over a range of abscissa values for an associated quasi-periodic waveform, generating a plurality of states, each state including a group of one or more arranged data structures, and storing the states in a storage device.
In some embodiments, the computer-readable medium includes further instructions configured for the plurality of states to contain a state with a reference abscissa value associated with the abscissa values of the arranged data structure, and assigned to the state, and the plurality of states are arranged into a state sequence, according to the associated abscissa values.
In some embodiments, the computer-readable medium includes further instructions configured for the reference abscissa value to be the greatest of the abscissa values of the arranged data structure assigned to the corresponding state.
In some embodiments, the computer-readable medium includes further instructions configured for a new state to be formed upon a change of an active fractional phase of one or more of the arranged data structures in the state, and forming a sequence of states.
In some embodiments, the computer-readable medium includes further instructions configured for analyzing a state trajectory in state space, and identifying a set of features within a quasi-periodic waveform, a component signals, or both the quasi-periodic waveform and the component signals based upon the state space.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be an electrocardiogram (ECG) signal from a mammal, with features including: a P wave, a Q wave, an R wave, an S wave, a T wave, a U wave, and any ECG segment delineated by any pair of these waves.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be an electrocardiogram (ECG) signal from a mammal, with features selected from a group consisting of: a P wave, a Q wave, an R wave, an S wave, a T wave, a U wave, and any ECG segment delineated by any pair of these waves.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a heart sound signal from a mammal, and the features include: an S1, an S2, and an S3 sound.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a heart sound signal from a mammal and includes features selected from a group consisting of: an S1, an S2, and an S3 sound.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a blood pressure signal from a mammal, and the features include a dicrotic notch.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a blood flow signal from a mammal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be an EEG, an EMG, or a EOG signal from a mammal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a respiratory signal from a mammal, further comprising allowing the respiratory signal an optional descriptor includes: pressure, flow, and volume information.
In some embodiments, the computer-readable medium includes further instructions configured for analyzing a state trajectory in state space, and identifying a condition subset from a predetermined set of conditions associated with a quasi-periodic waveform, a component signal, or both the quasi-periodic waveform and the component signals based upon the state space analysis.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be an electrocardiogram (ECG) signal from a mammal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a heart-sound signal from a mammal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a blood-pressure signal from a mammal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a blood-flow signal from a mammal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be an EEG, EMG or an EOG signal from a mammal.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be a respiratory signal from a mammal, further comprising allowing the respiratory signal an optional descriptor includes: pressure, flow, and volume information.
In some embodiments, the computer-readable medium includes further instructions configured for the quasi-periodic waveform signal to be derived from a mammal and a condition set including: an optional descriptor the condition normal, and an optional descriptor the condition abnormal.
In some embodiments, the computer-readable medium includes further instructions configured such that the condition set includes cardiac conditions, the normal condition is normal sinus rhythm, and the abnormal condition is a set of rhythm abnormalities including an optional descriptor bradycardia and tachycardia abnormality, and an optional descriptor conduction abnormality, further comprising allowing bradycardia is an optional descriptor set including sick sinus syndrome, and tachycardia is an optional descriptor set including atrial flutter, atrial fibrillation, ventricular tachycardia, and ventricular fibrillation, and conduction abnormality is an optional descriptor set to include bundle branch block, AV and node block.
In some embodiments, the computer-readable medium includes further instructions configured such that the condition set includes cardiac conditions, the normal condition is normal sinus rhythm, and the abnormal condition is a set of structural abnormalities including: optional ischemia, optional myocardial infarction, optional left ventricular hypertrophy, optional right ventricular hypertrophy, optional bi-ventricular hypertrophy, optional long QT syndrome, and optional tall T syndrome.
In some embodiments, the computer-readable medium includes further instructions configured for the set conditions to be normal blood pressure and abnormal blood pressure.
In some embodiments, the computer-readable medium includes further instructions configured for the set conditions to include hypotension and hypertension.
In some embodiments, the computer-readable medium includes further instructions configured for obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, further comprising allowing each quasi-periodic waveform corresponds to a predetermined set of feature types and assigned by feature type in a known manner to a set of states forming a state subsequence of the associated state sequence, constructing a sequence detector for each state subsequence resulting from a training set, such that detection of the state subsequence is an indication of the assigned feature type, and identifying the feature type associated with each feature of an arbitrary test quasi-periodic waveform, based upon a pattern-matching operation.
In some embodiments, the computer-readable medium includes further instructions configured for the pattern-matching operation includes: establishing a degree of match measure between the sequences associated with each detector and the state sequence resulting from the test waveform, and forming a fuzzy inference upon these measures through a fuzzy logic mapping.
In some embodiments, the computer-readable medium includes further instructions configured for the pattern-matching operation includes: establishing a degree of match measure between the state sequences associated with each detector and the state sequence resulting from the test waveform, and applying a neural network mapping to these measures, further comprising allowing the neural network is trained on the training set.
In some embodiments, the computer-readable medium includes further instructions configured for obtaining the state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, each waveform assigned in a known manner to a subset of a set of one or more predetermined conditions, constructing a sequence detector for each state sequence resulting from the training set, such that detection of the state sequence is an indication of the assigned condition subset, and identifying the condition subset associated with an arbitrary test quasi-periodic waveform, based upon a pattern-matching operation.
In some embodiments, the computer-readable medium includes further instructions configured for the pattern-matching operation includes: establishing a degree of match measure between the sequences associated with each detector and the state sequence resulting from the test waveform, and forming a fuzzy inference upon these measures through a fuzzy logic mapping.
In some embodiments, the computer-readable medium includes further instructions configured for the pattern-matching operation includes: establishing a degree of match measure between the sequences associated with each detector and the state sequence resulting from the test waveform, and applying a neural network mapping to these measures, further comprising allowing the neural network is trained on the training set.
In some embodiments, the computer-readable medium includes further instructions configured for obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, each feature of a set of features in each waveform corresponding to a predetermined set of feature types and assigned by feature type in a known manner to a set of states forming a subsequence of the associated state sequence, constructing a model according to a span of a state vector or a subspace of that span in state space, training the model based upon the state subsequences resulting from a training data set, including associated feature sets as labels for supervised learning, and identifying a feature type associated with each feature of an arbitrary test quasi-periodic waveform, based upon statistics in the trained model in relation to the state subsequences resulting from test waveform features. In some embodiments, the computer-readable medium includes further instructions configured for the model to have multiple modalities.
In some embodiments, the computer-readable medium includes further instructions configured for producing the model to include a plurality of submodels.
In some embodiments, the computer-readable medium includes further instructions configured for producing the submodel as a Hidden Markov Model, Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the computer-readable medium includes further instructions configured for the model to be a Hidden Markov Model, Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the computer-readable medium includes further instructions configured for obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, each waveform assigned in a known manner to a subset of a set of one or more predetermined conditions, constructing a model according to a span of a state vector or a subspace of that span in state space, training the model based upon the state sequences resulting from a training data set, including associated condition sets as labels for supervised learning, and identifying the condition subset associated with an arbitrary test quasi-periodic waveform, based upon statistics in the trained model in relation to the state sequence resulting from a test waveform. In some embodiments, the computer-readable medium includes further instructions configured for the model to have multiple modalities.
In some embodiments, the computer-readable medium includes further instructions configured for the model to have a plurality of submodels.
In some embodiments, the computer-readable medium includes further instructions configured for the submodel to be a Hidden Markov Model, Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the computer-readable medium includes further instructions configured for the model to be a Hidden Markov Model, Neural Network, or Fuzzy Logic Mapping.
Some embodiments of the invention provide a computer-readable medium that includes instructions configured to perform a method that includes: obtaining a state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, waveforms in aggregation containing an unknown set of features generally falling into groups, these features each having an associated set of states forming a subsequence of the associated state sequence, constructing a model according to a span of a state vector or a subspace of that span in state space, training the model based upon the state subsequences resulting from the training data set, allowing for self-organization of the clusters in state-space, assigning labels to the emergent clusters and associating them with corresponding feature types on the quasi-periodic waveform, and identifying the feature type associated with each feature of an arbitrary test quasi-periodic waveform, based upon statistics in the trained model in relation to the state subsequences resulting from test waveform features. In some embodiments, the computer-readable medium includes further instructions configured for the model to have multiple modalities. In some embodiments, the computer-readable medium includes further instructions configured for the model to include a plurality of submodels.
In some embodiments, the computer-readable medium includes further instructions configured for the submodel to be a Hidden Markov Model, Neural Network, or Fuzzy Logic Mapping.
In some embodiments, the computer-readable medium includes further instructions configured for the model to be a Hidden Markov Model, Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the computer-readable medium includes further instructions configured for obtaining the state sequence resulting from each of one or more quasi-periodic waveforms comprising a training set, the waveforms in aggregation pertaining to one or more unknown conditions of a set of conditions of unknown size, these waveforms each having an associated state sequence, constructing a model according to the span of the state vector or a subspace of that span in state space, training the model based upon the state sequences resulting from the training data set, allowing for self-organization of the clusters in state-space, and assigning labels to the emergent clusters and associating them with condition types, and identifying the set of condition types associated with an arbitrary test quasi-periodic waveform, based upon the statistics in the trained model in relation to the state sequence resulting from the test waveform. In some embodiments, the computer-readable medium includes further instructions configured for the model to have multiple modalities. In some embodiments, the computer-readable medium includes further instructions configured for the model to include a plurality of submodels.
In some embodiments, the computer-readable medium includes further instructions configured for the submodel to be a Hidden Markov Model, a Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the computer-readable medium includes further instructions configured for the model to be a Hidden Markov Model, Neural Network, or a Fuzzy Logic Mapping.
In some embodiments, the computer-readable medium includes further instructions configured for allowing the states to be pruned from the state trajectory based upon the relative time of habitation within the states.
In some embodiments, the computer-readable medium includes further instructions configured for allowing the states to be pruned from the state trajectory based upon the relative time of habitation within the states.
Some embodiments of the invention provide a method that includes: obtaining a plurality of data structures, assigning the data structure attributes to a reconstruction formula to substantially reconstruct the associated component signal over the associated abscissa interval where the associated fractional phase is active, and generating a plurality of one or more reconstructed component signals.
In some embodiments, the method allows the plurality of reconstructed component signals to be linearly combined to substantially reconstruct the quasi-periodic waveform.
In some embodiments, the method allows the subset of the one or more reconstructed component signals to be linearly combined to substantially reconstruct a desired portion of the quasi-periodic waveform.
In some embodiments, the method further includes: obtaining a plurality of data structures, performing data compression on the data structures, and storing the compressed data structures in a storage device.
In some embodiments, the method provides for compression to be performed through selective removal of data structures containing redundant/unremarkable information.
In some embodiments, the method provides for compression to be performed through selective removal of redundant/unremarkable information in the data structures.
In some embodiments, the method provides for compression to be performed through data compression operations performed upon the attribute information in the data structures.
Some embodiments of the invention provide a computer-readable medium having executable instructions stored thereon for causing a suitable programmed processing unit to perform a method that includes: obtaining a plurality of data structures, assigning the data structure attributes to a reconstruction formula to substantially reconstruct the associated component signal over the associated abscissa interval where the associated fractional phase is active, and generating a plurality of one or more reconstructed component signals.
In some embodiments, the computer-readable medium includes further instructions configured for the plurality of reconstructed component signals to be linearly combined to substantially reconstruct the quasi-periodic waveform.
In some embodiments, the computer-readable medium includes further instructions configured for a subset of one or more reconstructed component signals to be linearly combined to substantially reconstruct a desired portion of the quasi-periodic waveform.
Some embodiments of the invention provide a computer-readable medium that includes instructions for causing a suitable programmed central processing unit to perform a method that includes: obtaining a plurality of data structures, performing data compression on the data structures, and storing the compressed data structures in a storage device.
In some embodiments, the computer-readable media is described wherein the compression is performed through selective removal of data structures containing redundant/unremarkable information.
In some embodiments, the computer-readable medium includes further instructions configured for the compression to be performed through selective removal of redundant/unremarkable information in the data structures.
In some embodiments, the computer-readable medium includes further instructions configured for the compression to be performed through data compression operations performed upon the attribute information in the data structures.
As used herein, a fractional phase is one of a plurality of fractional phases in a fractional phase representation (i.e., an interval within one or more cycles of a component signal representative of a digitally filtered frequency-limited band obtained from a quasi-periodic waveform).
As used herein, a quarter-phase is one of four fractional phases in a fractional phase representation (i.e., one of four intervals within one cycle of a component signal representative of a digitally filtered frequency-limited band obtained from a quasi-periodic waveform).
As used herein, a quasi-periodic waveform is a signal that may be considered as the linear combination of a series of locally periodic component signals, each having different periods and/or shapes, with locality exhibited on the order of a cycle to a few cycles. Examples of quasi-periodic waveforms that can be processed and analyzed by the present invention include but are not limited to:
general biological signals from humans or animals such as ECG waveforms, intracardiac electrogram (EGM) waveforms, EEG waveforms, EMG waveforms, EOG waveforms, blood pressure waveforms, plethysmography waveforms, heart sounds, ultrasound signals, MRI signals, bio-impedance waveforms, electro-dermal testing and polygraph signals;
natural phenomena such as solar sunspot cycles, climate patterns, patterns of El NiZo, drought cycles, seismic activity (i.e., earthquakes in certain areas occur approximately every N years, and major earthquakes can be preceded by relatively minor seismic activity that exhibit quasi-periodic characteristics), ocean currents and tides, and occurrence of tsunamis;
manufacturing process signals for feedback and control;
demographics and societal patterns such as population cycles, product market preferences, crowd noise, traffic patterns, traffic noise, and industrial noise;
vehicular signals pertaining to feedback, control, and detection and diagnosis of issues associated with operation and/or service of a vehicle, such as engine and transmission noise, temperature, information from any number of different types of sensors such as oxygen sensors, temperature sensors, pressure sensors, and flow sensors;
aircraft performance-related signals, such as from laminar-flow sensors, pressure sensors, windshear sensors and downdraft sensors;
financial time series, representing for example prices of stocks, equities, securities, commodities, bonds, mutual funds, and futures, and economic cycles including micro and macro-economic phenomena;
content-carrying signals, such as music, speech or video signals (including analysis and compression of such signals);
telemetry signals for applications such as radar (including, for example, Synthetic Aperture Radar (SAR), Terrain Following Radar (TFR), and Forward-Looking Infrared (FLIR) radar) and sonar;
communications signals such as transmitted over an optical fiber, cable or remote radio link (including spread-spectrum transmissions), such as weak signals from satellites and spacecraft or animal-tracking devices, cordless and cellular telephones, and wireless computer networks such as those using the IEEE802.11g standard;
analysis of data storage medium signals such as from a magnetic disk head or a laser optical receiver for optical storage media; and
other signals exhibiting quasi-periodic behavior, including seismic signals obtained for oil exploration.
Other embodiments involving various permutations and distributions of processes or blocks within the Device and Host are specifically contemplated to fall within the scope of this current invention. Some embodiments include combinations or subcombinations of a plurality of items listed individually or within one or more groups of other items.
Some embodiments of the invention provide a method that includes: obtaining a component signal representative of a digitally filtered frequency-limited band from a quasi-periodic waveform, wherein the component signal includes a plurality of component values, reducing the plurality of component values into a fractional-phase representation of N fractional phases where N is greater than one, and wherein each fractional phase is associated with a phase label and includes one or more values representing an abscissa, an ordinate, or both an abscissa and an associated ordinate for each of one or more local cycles of the component signal, and storing the fractional-phase representation in a data structure on a storage device.
In some embodiments of the method, the fractional-phase representation for one local cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein the reducing of the plurality of component values into the fractional-phase representation further includes: obtaining, for the first fractional phase, a value representing an abscissa of a zero crossing of the component signal, obtaining, for the second fractional phase, a value representing an ordinate and a value representing an abscissa of a maximum of the component signal, obtaining, for the third fractional phase, a value representing an abscissa of another zero crossing of the component signal, and obtaining, for the fourth fractional phase, a value representing an ordinate and a value representing an abscissa of a minimum of the component signal.
In some embodiments, each ordinate value is obtained based on an amplitude metric and each abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
In some embodiments, each zero crossing is determined by finding a change in sign and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, the maximum and the minimum of the component signal are each obtained by finding a change in sign in an estimate of the local derivative and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, the method further includes: calculating a fractional phase adjustment value for the fractional-phase representation such that a linear combination of values derived from the fractional-phase representation with values derived from one or more other fractional-phase representations obtained from other digitally filtered frequency-limited bands substantially reconstructs the quasi-periodic waveform.
In some embodiments, the component signal is analytic such that the plurality of component values for the component signal are complex numbers, and wherein the fractional-phase representation for each fractional phase is associated with an angular range of a complex argument of the analytic component signal.
In some embodiments, the fractional-phase representation for one local cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein the reducing of the plurality of component values into the fractional-phase representation further includes: obtaining, for the first fractional phase, a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of a real part of the component signal, obtaining, for the second fractional phase, a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of an imaginary part of the component signal, obtaining, for the third fractional phase, a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of the real part of the component signal, and obtaining, for the fourth fractional phase, a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of an imaginary part of the component signal.
In some embodiments, each zero crossing is determined by finding a change in sign and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, each complex ordinate value is based on an amplitude metric and each real abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
In some embodiments, the invention provides a computer-readable medium having instructions thereon for causing a suitably programmed information processor to execute a method comprising: obtaining a component signal representative a digitally filtered frequency-limited band from a quasi-periodic waveform, wherein the component signal includes a plurality of component values, reducing the plurality of component values into a fractional-phase representation of N fractional phases where N is greater than one, and wherein each fractional phase is associated with a phase label and includes one or more values representing an abscissa, an ordinate, or both an abscissa and an associated ordinate for each of one or more local cycles of the component signal, and storing the fractional-phase representation in a data structure on a storage device.
In some embodiments of the medium, the fractional-phase representation for one cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein: the fractional-phase representation for the first fractional phase includes a value representing an abscissa of a zero crossing of the component signal, the fractional-phase representation for the second fractional phase includes a value representing an ordinate and a value representing an abscissa of a maximum of the component signal, the fractional-phase representation for the third fractional phase includes a value representing an abscissa of another zero crossing of the component signal, and the fractional-phase representation for the fourth fractional phase includes a value representing an ordinate and a value representing an abscissa of a minimum of the component signal.
In some embodiments, each ordinate value is based on an amplitude metric and each abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
In some embodiments, each zero crossing is evaluated by finding a change in sign and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, the maximum and the minimum of the component signal are each evaluated by finding a change in sign in an estimate of the local derivative and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, the fractional-phase representation includes a fractional phase adjustment value such that a linear combination of values derived from the fractional-phase representation with values derived from one or more other fractional-phase representations obtained from other digitally filtered frequency-limited bands substantially reconstructs the quasi-periodic waveform.
In some embodiments, the component signal is analytic such that the plurality of component values for the component signal are complex numbers, and wherein the fractional-phase representation for each fractional phase is associated with an angular range of a complex argument of the analytic component signal.
Some embodiments of the invention provide an apparatus wherein the fractional-phase representation for one local cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein: the fractional-phase representation for the first fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of a real part of the component signal, the fractional-phase representation for the second fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of an imaginary part of the component signal, the fractional-phase representation for the third fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of the real part of the component signal, the fractional-phase representation for the fourth fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of an imaginary part of the component signal.
In some embodiments, each zero crossing is evaluated by finding a change in sign and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, each complex ordinate value is based on an amplitude metric and each real abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
Some embodiments of the invention provide an apparatus that includes a storage device, a source for a component signal representative of a digitally filtered frequency-limited band from a quasi-periodic waveform, wherein the component signal includes a plurality of component values, and means for reducing the plurality of component values into a fractional-phase representation of N fractional phases where N is greater than one, and wherein each fractional phase is associated with a phase label and includes one or more values representing an abscissa, an ordinate, or both an abscissa and an associated ordinate for each of one or more local cycles of the component signal, and for storing the fractional-phase representation on the storage device.
In some embodiments of the apparatus, the fractional-phase representation for one local cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein: the fractional-phase representation for the first fractional phase includes a value representing an abscissa of a zero crossing of the component signal, the fractional-phase representation for the second fractional phase includes a value representing an ordinate and a value representing an abscissa of a maximum of the component signal, the fractional-phase representation for the third fractional phase includes a value representing an abscissa of another zero crossing of the component signal, and the fractional-phase representation for the fourth fractional phase includes a value representing an ordinate and a value representing an abscissa of a minimum of the component signal.
In some embodiments, each ordinate value is based on an amplitude metric and each abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
In some embodiments, each zero crossing is evaluated by finding a change in sign and interpolating in the neighborhood of the sign change, e.g., in some embodiments, interpolating using a curve fit.
In some embodiments, the maximum and the minimum of the component signal are each evaluated by finding a change in sign in an estimate of the local derivative and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, the fractional-phase representation includes a fractional phase adjustment value such that a linear combination of values derived from the fractional-phase representation with values derived from one or more other fractional-phase representations obtained from other digitally filtered frequency-limited bands substantially reconstructs the quasi-periodic waveform.
In some embodiments, the component signal is analytic such that the plurality of component values for the component signal are complex numbers, and wherein the fractional-phase representation for each fractional phase is associated with an angular range of a complex argument of the analytic component signal.
In some embodiments, the fractional-phase representation for one local cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein: the fractional-phase representation for the first fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of a real part of the component signal, the fractional-phase representation for the second fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of an imaginary part of the component signal, the fractional-phase representation for the third fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of the real part of the component signal, the fractional-phase representation for the fourth fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of an imaginary part of the component signal.
In some embodiments, each zero crossing is evaluated by finding a change in sign and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, each complex ordinate value is based on an amplitude metric and each real abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
Some embodiments of the invention provide an apparatus that includes: a storage device, a source for a component signal representative, a digitally filtered frequency-limited band from a quasi-periodic waveform, wherein the component signal includes a plurality of component values, and a fractional-phase reducer that operates on the plurality of component values and generates a fractional-phase representation of N fractional phases where N is greater than one, and wherein each fractional phase is associated with a phase label and includes one or more values representing an abscissa, an ordinate, or both an abscissa and an associated ordinate for each of one or more local cycles of the component signal, and wherein the reducer is operatively coupled to store the fractional-phase representation on the storage device.
In some embodiments of the apparatus, the fractional-phase representation for one local cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein: the fractional-phase representation for the first fractional phase includes a value representing an abscissa of a zero crossing of the component signal, the fractional-phase representation for the second fractional phase includes a value representing an ordinate and a value representing an abscissa of a maximum of the component signal, the fractional-phase representation for the third fractional phase includes a value representing an abscissa of another zero crossing of the component signal, and the fractional-phase representation for the fourth fractional phase includes a value representing an ordinate and a value representing an abscissa of a minimum of the component signal.
In some embodiments, each ordinate value is based on an amplitude metric and each abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
In some embodiments, each zero crossing is evaluated by finding a change in sign and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, the maximum and the minimum of the component signal are each evaluated by finding a change in sign in an estimate of the local derivative and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, the fractional-phase representation includes a fractional phase adjustment value such that a linear combination of values derived from the fractional-phase representation with values derived from one or more other fractional-phase representations obtained from other digitally filtered frequency-limited bands substantially reconstructs the quasi-periodic waveform.
In some embodiments, the component signal is analytic such that the plurality of component values for the component signal are complex numbers, and wherein the fractional-phase representation for each fractional phase is associated with an angular range of a complex argument of the analytic component signal.
In some embodiments, the number of fractional phases is four and the fractional-phase representation for one local cycle of the component signal includes a first fractional phase, a second fractional phase, a third fractional phase, and a fourth fractional phase, and wherein: the fractional-phase representation for the first fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of a real part of the component signal, the fractional-phase representation for the second fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of a zero crossing of an imaginary part of the component signal, the fractional-phase representation for the third fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of the real part of the component signal, the fractional-phase representation for the fourth fractional phase includes a complex value representing an ordinate and a real value representing an abscissa of another zero crossing of an imaginary part of the component signal.
In some embodiments, each zero crossing is evaluated by finding a change in sign and interpolating in the neighborhood of the sign change using a curve fit.
In some embodiments, each complex ordinate value is based on an amplitude metric and each real abscissa value is based on a time metric, and wherein the component signal is an electrocardiogram (ECG) signal from a human.
Some embodiments of the invention provide a computer-readable medium having stored thereon a first data structure for a fractional-phase representation, the data structure including: a descriptor that includes an abscissa value from the fractional-phase representation, a descriptor that includes an ordinate value from the associated fractional-phase representation, and a descriptor that includes an abscissa interval value representing an interval over which an associated fractional phase is active.
In some embodiments of the computer-readable medium, the first data structure further includes three or more descriptors selected from a group consisting of: a descriptor that includes a label identifying a particular data structure from within a plurality of data structures, a descriptor that includes a label identifying the component from a plurality of components, a descriptor that includes a label identifying the quasi-periodic waveform, a descriptor that includes a value representing a local period of the associated component, a descriptor that includes a value representing an interval from a reference abscissa value to an associated abscissa value, a descriptor that includes a value representing a relative measure between the ordinate value associated with the fractional phase, and an ordinate value associated with the component, the quasi-periodic waveform, or both the component and the quasi-periodic waveform, and a descriptor that includes a label indicating that the component is active for the active fractional phase associated with the particular data structure, based upon comparing the ordinate value to a predetermined threshold.
In some embodiments, the first data structure further includes links to other data structures, including: a link to another particular data structure associated with the component, wherein the associated abscissa value of the other data structure is less than the abscissa value associated in the particular data structure, a link to another particular data structure associated with the component, wherein the abscissa value of the other data structure is greater than the abscissa value in the particular data structure, a link to another particular data structure associated with the component, wherein the abscissa value of the other data structure is less than the abscissa value in the particular data structure, and wherein the phase label of the other data structure is of the same value as the phase label of the particular data structure, and a link to another particular data structure associated with the component, wherein the abscissa value of the other data structure is greater than the abscissa value in the particular data structure, and wherein the phase label of the other data structure is of the same value as the phase label of the particular data structure.
In some embodiments, the first data structure further includes linked attributes, including: a descriptor that includes a difference of abscissa values between the abscissa value included in the particular data structure and the abscissa value included in a first linked data structure relative to this particular data structure, a descriptor that includes a difference of abscissa values between the abscissa value included in the particular data structure and the abscissa value included in a second linked data structure relative to this particular data structure, a descriptor that includes an indication of deviation from an expected sequence of phase labels, and a descriptor that includes a moving average of abscissa values for a group of data structures surrounding the particular data structure.
It is to be understood that the above description is intended to be illustrative, and not restrictive. Although numerous characteristics and advantages of various embodiments as described herein have been set forth in the foregoing description, together with details of the structure and function of various embodiments, many other embodiments and changes to details will be apparent to those of skill in the art upon reviewing the above description. The scope of the invention should, therefore, be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled. In the appended claims, the terms “including” and “in which” are used as the plain-English equivalents of the respective terms “comprising” and “wherein,” respectively. Moreover, the terms “first,” “second,” and “third,” etc., are used merely as labels, and are not intended to impose numerical requirements on their objects.
This application is a divisional of U.S. patent application Ser. No. 13/220,679, filed Aug. 29, 2011 (which issued as U.S. Pat. No. 8,386,244 on Feb. 26, 2013), titled “SIGNAL DECOMPOSITION, ANALYSIS AND RECONSTRUCTION,” which is a divisional of U.S. patent application Ser. No. 12/760,554, filed Apr. 15, 2010 (which issued as U.S. Pat. No. 8,010,347 on Aug. 30, 2011), titled “SIGNAL DECOMPOSITION, ANALYSIS AND RECONSTRUCTION APPARATUS AND METHOD,” which is a divisional of U.S. patent application Ser. No. 11/360,135, filed Feb. 23, 2006 (which issued as U.S. Pat. No. 7,702,502 on Apr. 20, 2010), titled “APPARATUS FOR SIGNAL DECOMPOSITION, ANALYSIS AND RECONSTRUCTION,” which claimed benefit of U.S. Provisional Patent Application 60/656,630, filed Feb. 23, 2005, titled “SYSTEM AND METHOD FOR SIGNAL DECOMPOSITION, ANALYSIS AND RECONSTRUCTION,” each of which is incorporated herein by reference in its entirety. This application is related to U.S. patent application Ser. No. 11/360,223, filed Feb. 23, 2006 (which issued as U.S. Pat. No. 7,706,992 on Apr. 27, 2010), titled “SYSTEM AND METHOD FOR SIGNAL DECOMPOSITION, ANALYSIS AND RECONSTRUCTION,” which is incorporated herein by reference in its entirety.
| Number | Name | Date | Kind |
|---|---|---|---|
| 3819920 | Goldfischer | Jun 1974 | A |
| 3993862 | Karr | Nov 1976 | A |
| 4193393 | Schlager | Mar 1980 | A |
| 4251831 | Kamath | Feb 1981 | A |
| 4314105 | Mozer | Feb 1982 | A |
| 4461022 | Slagley | Jul 1984 | A |
| 4633884 | Imai et al. | Jan 1987 | A |
| 4680797 | Benke | Jul 1987 | A |
| 4736295 | Lachiver et al. | Apr 1988 | A |
| 5115240 | Fujiwara et al. | May 1992 | A |
| 5230038 | Fielder et al. | Jul 1993 | A |
| 5392044 | Kotzin et al. | Feb 1995 | A |
| 5486867 | Hsu et al. | Jan 1996 | A |
| 5730142 | Sun et al. | Mar 1998 | A |
| 5749367 | Gamlyn et al. | May 1998 | A |
| 5778881 | Sun et al. | Jul 1998 | A |
| 5828995 | Satyamurti et al. | Oct 1998 | A |
| 6020840 | Ong | Feb 2000 | A |
| 6389308 | Shusterman | May 2002 | B1 |
| 6456651 | Pilozzi et al. | Sep 2002 | B1 |
| 6475245 | Gersho et al. | Nov 2002 | B2 |
| 6510339 | Kovtun et al. | Jan 2003 | B2 |
| 6512944 | Kovtun et al. | Jan 2003 | B1 |
| 6687674 | Suzuki et al. | Feb 2004 | B2 |
| 6718038 | Cusmario | Apr 2004 | B1 |
| 6782058 | Nayler | Aug 2004 | B1 |
| 6791482 | Koyanagi | Sep 2004 | B2 |
| 6914935 | Eklof | Jul 2005 | B2 |
| 6925324 | Shusterman | Aug 2005 | B2 |
| 7054792 | Frei et al. | May 2006 | B2 |
| 7058548 | Pupalaikis et al. | Jun 2006 | B2 |
| 7088276 | Wegener | Aug 2006 | B1 |
| 7254187 | Mohan et al. | Aug 2007 | B2 |
| 7640055 | Geva et al. | Dec 2009 | B2 |
| 7702502 | Ricci et al. | Apr 2010 | B2 |
| 7706992 | Ricci et al. | Apr 2010 | B2 |
| 8010347 | Ricci et al. | Aug 2011 | B2 |
| 8386244 | Ricci et al. | Feb 2013 | B2 |
| 20020178006 | Suzuki et al. | Nov 2002 | A1 |
| 20040181399 | Gao | Sep 2004 | A1 |
| 20140278382 | Ricci et al. | Sep 2014 | A1 |
| 20140289297 | Ricci et al. | Sep 2014 | A1 |
| Entry |
|---|
| Bracewell, Ronald N., “The Fourier Transform and its Applications (Table of Contents, Index Only)”, 2000, Publisher: McGraw-Hill. |
| Cappe, Olivier, et al., “Inference in Hidden Markov Models (Table of Contents, Index Only)”, 2005, Publisher: Springer. |
| Cetin, A. Enis, et al., “Compression of Digital Biomedical Signals”, “The Biomedical Engineering Handbook: Second Edition. Joseph D. Bonzino, Ed. CRC Press LLC”, 2000, vol. Chapter 54. |
| Chen, Ying-Jui, et al., “Multiplierless Approximations of Transforms with Adder Constraing”, “IEEE Signal Processing Letters ”, Nov. 2002, pp. 344-347, vol. 9, No. 11. |
| Duda, Richard O., et al., “Pattern Classification, Second Ed. (Table of Contents, Index Only)”, 2001, Publisher: John Wiley & Sons, Inc. |
| Elliott, Richard J., et al., “Hidden Markov Models (Table of Contents, Index Only)”, 1995, Publisher: Springer-Verlag. |
| Fliege, N. J., “Multirate Digital Signal Processing (Table of Contents, Index Only)”, 1994, Publisher: Wiley. |
| Golub, Gene H., et al., “Matrix Computations, Third Ed. (Table of Contents, Index Only)”, 1996, Publisher: Hopkins. |
| Kotteri, K., et al., “Design of Multiplierless, High-Performance, Wavelet Filter Banks with Image Compression Applications”, “IEEE Transactions on Circuits and Systems,”, Mar. 2004, pp. 483-494, vol. 51, No. 3. |
| MacDonald, Iain L., et al., “Hidden Markov and Other Models for Discrete-valued Time Series (Table of Contents, Index Only)”, 1997, Publisher: Chapman. |
| Openheim, Alan V., et al., “Discrete-Time Signal Processing, Second Ed. (Table of Contents, Index Only)”, 1999, Publisher: Wiley. |
| Ozaktas, Haldun M., et al., “The Fractional Fourier Transform (Table of Contents, Index Only)”, 2001, Publisher: Wiley. |
| Stoica, Petre, et al., “Spectral Analysis of Signals (Table of Contents, Index Only)”, 2005, Publisher: Prentice-Hall. |
| Strang, Gilbert, et al., “Wavelets and Filter Banks (Table of Contents, Index Only)”, 1997, Publisher: Wellesley. |
| Vapnik, Vladimir N., “The Nature of Statistical Learning Theory, Second Ed. (Table of Contents, Index Only)”, 2000, Publisher: Springer. |
| Vetterli, Martin, et al., “Wavelets and Subband Coding (Table of Contents, Index Only)”, 1995, Publisher: Prentice-Hall. |
| Xue, Qiuzhen, et al., “Late Potential Recognition by Artificial Neural Networks”, “IEEE Trans Bio Eng”, 1997, pp. 132-143, vol. 44. |
| Number | Date | Country | |
|---|---|---|---|
| 20130282339 A1 | Oct 2013 | US |
| Number | Date | Country | |
|---|---|---|---|
| 60656630 | Feb 2005 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 13220679 | Aug 2011 | US |
| Child | 13765636 | US | |
| Parent | 12760554 | Apr 2010 | US |
| Child | 13220679 | US | |
| Parent | 11360135 | Feb 2006 | US |
| Child | 12760554 | US |
