Impedance measurements

Abstract

A method of determining frequencies for use in performing impedance measurements. The method includes determining estimates for parameter values representing an impedance response for at least one subject, using the estimated parameter values to determine a design and using the design to determine frequencies for use in impedance measurements.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS

This application is a §371 national phase of International application PCT/AU2007/000726 filed May 25, 2007, which claims foreign priority to Australian Patent Application no. 2006-902907 filed May 30, 2006.

BACKGROUND OF THE INVENTION

The present invention relates to a method and apparatus for determining frequencies for use in performing impedance measurements on a subject, as well as to a method and apparatus for performing impedance measurements.

DESCRIPTION OF THE PRIOR ART

The reference in this specification to any prior publication (or information derived from it), or to any matter which is known, is not, and should not be taken as an acknowledgment or admission or any form of suggestion that the prior publication (or information derived from it) or known matter forms part of the common general knowledge in the field of endeavour to which this specification relates.

Bioelectrical Impedance Analysis (BIA) measures the impedance to flow of an alternating electrical current passed through biological tissue. Such impedance measurements are typically performed at a number of different frequencies, allowing a subject's impedance response to be modelled, using a mathematic relationship, such as the Cole model. This in turn enables the estimation of various parameters, which can in turn be used to derive information regarding a subject's health.

However, the frequencies used are typically selected randomly, or based on the ability of the measuring device used to perform the measurements. Consequently the effectiveness and accuracy of the measurements varies greatly.

SUMMARY OF THE PRESENT INVENTION

In a first broad form the present invention provides a method of determining frequencies for use in performing impedance measurements, the method including:

• • a) determining estimates for parameter values representing an impedance response for at least one subject;
  • b) using the estimated parameter values to determine a design; and,
  • c) using the design to determine frequencies for use in impedance measurements.

Typically the method includes, determining the estimates for the parameter values at least in part using a model of the subject's impedance response.

Typically the method includes:

• • a) determining a residual variance model having a predetermined distribution; and,
  • b) determining the estimates for the parameter values at least in part using the selecting variance model.

Typically the method of determining the residual variance model includes using an expectation maximization algorithm.

Typically the method includes determining a range of parameter values representing the impedance response of a number of subjects.

Typically the method includes:

• • a) determining a Fisher information matrix for a model of the impedance response of the at least one subject; and,
  • b) determining a design using the Fisher information matrix.

Typically the model includes at least one of:

• • a) A Cole model; and,
  • b) A residual variance model.

Typically the model includes determining the design by optimising the determinant of the Fisher information matrix.

Typically the design is a D-optimal design.

Typically the method includes determining a design space to take into account practical limitations.

Typically the method includes, modifying the design by at least one of:

• • a) restricting the frequencies; and,
  • b) providing an error structure as a function of frequency.

Typically the method is performed at least in part using a processing system.

Typically the method includes, in the processing system:

• • a) determining impedance data representing the impedance response of the at least one subject; and,
  • b) using the impedance data to determine the estimates for the parameter values.

Typically the method includes, in the processing system:

• • a) determining an impedance model; and,
  • b) using the impedance model and the estimated parameter values to determine the design.

Typically the method includes, in the processing system:

• • a) calculating a Fisher information matrix for the model; and,
  • b) using the Fisher information matrix to determine the design.

Typically the method includes, in the processing system, optimising a determinant of the Fisher information matrix for the model.

Typically the method includes, in the processing system:

• • a) receiving limits on applicable frequencies; and,
  • b) using the limits and the design to determine the frequencies.

Typically the method includes, in the processing system, determining frequencies for use in impedance measurements within the range:

• • a) 0-50 kHz;
  • b) 2-200 kHz;
  • c) 50-500 kHz; and,
  • d) 200-1000 kHz.

Typically the frequencies are within the ranges:

• • a) 0-24 kHz;
  • b) 5-74 kHz;
  • c) 77-200 kHz;
  • d) 530-1000 kHz.

Typically the frequencies are:

• • a) 14 kHz
  • b) 57 kHz
  • c) 188 kHz
  • d) 679 kHz

Typically the frequencies are:

• • a) 14.1844 kHz
  • b) 56.9201 kHz
  • c) 187.5397 kHz
  • d) 679.1141 kHz

In a second broad form the present invention provides apparatus for determining frequencies for use in performing impedance measurements, the apparatus including, a processing system for:

• • a) determining estimates for parameter values representing an impedance response for at least one subject;
  • b) using the estimated parameter values to determine a design; and,
  • c) using the design to determine frequencies for use in impedance measurements.

Typically the apparatus is for performing the first broad form of the invention.

In a third broad form the present invention provides a method of measuring the impedance of a subject the method including, the method including, in a measuring device:

• • a) causing one or more electrical signals to be applied to the subject using a first set of electrodes, the one or more electrical signals having four frequencies;
  • b) measuring electrical signals across a second set of electrodes applied to the subject in response to the applied one or more signals;
  • c) determining from the applied signals and the measured signals at least one measured impedance value at each of the four frequencies, wherein the four frequencies are in the ranges:
    • i) 0-50 kHz;
    • ii) 2-200 kHz;
    • iii) 50-500 kHz; and,
    • iv) 200-1000 kHz.

Typically a restricted range of frequencies can be defined by the ranges:

• • a) 0-24 kHz;
  • b) 5-74 kHz;
  • c) 77-200 kHz;
  • d) 530-1000 kHz.

Typically the frequencies are approximately:

• • a) 14 kHz
  • b) 57 kHz
  • c) 188 kHz
  • d) 679 kHz

Typically the frequencies are:

• • a) 14.1844 kHz
  • b) 56.9201 kHz
  • c) 187.5397 kHz
  • d) 679.1141 kHz

Typically the method includes determining one or more impedance parameter values based on the plurality of measured impedance values.

In a fourth broad form the present invention provides apparatus for measuring the impedance of a subject the apparatus including a measuring device for:

• • a) causing one or more electrical signals to be applied to the subject using a first set of electrodes, the one or more electrical signals having four frequencies;
  • b) measuring electrical signals across a second set of electrodes applied to the subject in response to the applied one or more signals;
  • c) determining from the applied signals and the measured signals at least one measured impedance value at each of the four frequencies, wherein the four frequencies are in the ranges:
    • i) 0-50 kHz;
    • ii) 2-200 kHz;
    • iii) 50-500 kHz; and,
    • iv) 200-1000 kHz.

Typically the measuring device includes:

• • a) a current source for applying current signals to the subject;
  • b) a voltage sensor for measuring voltages signals across the subject; and,
  • c) a processing system could to the current source and the voltage sensor for:
    • i) causing the current signals to be applied to the subject; and,
    • ii) determining the impedance using the applied current signals and measured voltage signals.

Typically the apparatus is for performing the method of the third broad form of the invention.

BRIEF DESCRIPTION OF THE DRAWINGS

Examples of the present invention will now be described with reference to the accompanying drawings, in which:—

FIG. 1 is a flow chart of an example of a process for determining frequencies for use in performing impedance measurements;

FIG. 2 is a schematic diagram of an example of a processing system for use in determining frequencies;

FIG. 3 is a flow chart of a specific example of a process for determining frequencies for use in performing impedance measurements;

FIG. 4 is a schematic diagram of an example of an “idealised” equivalent circuit for the Cole model;

FIG. 5 is a schematic diagram of an example of apparatus for measuring impedance;

FIG. 6 is an example of a residual plot for determining an initial model fit;

FIG. 7 is an example of a resistance versus reactance plot for example impedance measurements;

FIG. 8 is an example of plots of a variance function for various values of σ²; and,

FIG. 9 is an example of a plot of residuals from a model fit using an example extracted data set.

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

An example of a process for determining frequencies at which impedance measurements may be made will now be described with reference to FIG. 1.

At step 100 parameter values are determined representing the impedance response of one or more subjects. The parameter values may be determined in any one of a number of manners, such by modelling impedance data collected from a number of subjects from a prior study, as will be described in more detail below.

At step 110 the parameter values are used to determine an optimal design for studying the impedance response of the one or more subjects. This can be achieved, for example, by optimising designs representing frequencies that may be used in performing impedance measurements according to optimality criteria.

At step 120 the optimal designs may optionally be modified to take into account practical effects, such as the variants of subjects within a population to impedance measurement.

At step 130 the optimal designs are used to determine frequencies that may be used for performing impedance measurements.

The process may be performed manually, but typically requires advanced computation and therefore typically requires the use of a processing system.

An example processing system is shown in FIG. 2. In this example, the processing system 200 is formed from a processor 210, a memory 211, an input/output device 212 and an optional external interface 213 interconnected via a bus 214. The external interface 213 may be used to couple the processing system 200 to a database 220.

In use the processing system 200 executes applications software stored in the memory 211, to allow parts of the process to be performed, as will be described in more detail below. It will be appreciated from this that the processing system 200 may be any suitable form of processing system 200, such as a personal computer, desktop, laptop, super computer, Sparc station, or the like.

An example of the process will now be described in more detail with respect to FIG. 3.

In this example, at step 300 impedance data is collected from a sample population of subjects. The data is collected using a suitable measuring device that applies alternating electrical signals to a subject, and measures the electrical response of the subject. This is typically achieved by applying alternating currents to the subject at a number of different frequencies, and then measures voltage signals across the subject, to allow the impedance to be determined at each frequency. An example measuring device will be described in more detail below.

The impedance response of the subject can be modelled using a suitable model, such the Cole model, which is based on an equivalent circuit that effectively models the electrical behaviour of biological tissue, an example of which is shown in FIG. 4.

In this example, the equivalent circuit includes two branches representing current flow through extracellular fluid and intracellular fluid. The extracellular component of biological impedance is represented by a resistance R_e, whilst the intracellular component is represented by a resistance R_i and a capacitance C.

Accordingly, the impedance of the equivalent circuit of FIG. 4 at an angular frequency ω, where ω=2π*frequency, is given by:

$\begin{array}{cc}Z=R_{\infty }+\frac{R_0-R_{\infty }}{1+\left(j\mathrm{\omega \tau }\right)}& \left(A\right)\end{array}$

• • where:
    • R_∞=impedance at infinite applied frequency=R_iR_e/(R_i+R_e),
    • R₀=impedance at zero applied frequency=R_e and,
    • τ is the time constant of the capacitive circuit.

However, as an alternative to the equivalent circuit described above, an alternative equivalent circuit incorporating a Fricke constant phase element (CPE) can be used, as will be understood by persons skilled in the art.

In any event, the equation (A) can be modified to include an exponent α to account for the distribution of time constants observed in biological tissues (or systems), as follows:

$\begin{array}{cc}Z=R_{\infty }+\frac{R_0-R_{\infty }}{1+{\left(j\mathrm{\omega \tau }\right)}^{1-\alpha }}& \left(B\right)\end{array}$

In this equation, R₁, R_e, C and α are parameters in the model, Z is the predicted impedance, j=√{square root over (−1)} and f is the frequency of the current passed through the body. R₁ and R_e are the resistances due to the intracellular and extracellular fluids in the body, respectively.

However, this is a theoretical model and in practice a subject's response will vary. Accordingly, at step 310 a residual variance model is selected to allow variations from the predicted response to be taken into account. The residual variance model may be selected in any one of ways as will be described in more detail below.

At step 320, parameter estimates are determined that represent the measured responses of the subjects. The initial parameters are parameters which when inserted into the models provide an estimate of the response of the sample population.

It will be appreciated that as different subjects within the sample population will have a range of different responses to impedance measurements. Accordingly, it is typical to model the range of responses across the sample population to allow the mean and variability of the parameters between subjects to be determined. This is effectively used to define a range of parameter values representing a parameter space.

At step 330, a number of Cole models are constructed across the parameter space, using the range of parameter values. This effectively models the range of different impedance responses of the subjects within the population.

At step 340 a Fisher information matrix is determined for the residual and Cole models. This can be achieved in any one of a number of ways, but typically involves using suitable applications software, such as the software package POPT by S. B. Duffull, implemented by the processing system 200.

At step 350, the product determinant of the Fisher information matrices is optimised. Again, this may be achieved in any one of a number of manners, such as by using a simulated annealing approach to find maxima for the expression. This provides an optimised design which represents the theoretical preferred frequencies at which impedance measurements may be made.

However, certain frequency measurements may not be practically achievable, or desirable. Thus, for example, theory may predict that the application of an electrical signal 0 kHz frequency will result in an improved subject response and hence improved impedance measurements. However, as the application of such a frequency is not practical, such a frequency is usually excluded. Similarly, high frequency measurements whilst theoretically advantageous, can be difficult to measure from a practical point of view, thereby further limiting the range of available frequencies. Accordingly, at step 360, a design space is selected that can be used to exclude such impractical frequencies.

At step 370 the optimal design can be used to determine preferred practical frequencies at which impedance measurements may be performed, together with an indication of the relative efficiency of the measurement procedure.

Accordingly, the above-described process operates by utilising impedance measurements for a sample population to determine optimal design, which can in turn be used to determine preferred frequencies for performing impedance measurements.

In one example, described in more detail in the specific example below, the process is used to determine that in general at least four frequencies should be used for performing impedance measurements. This can also be used to determine preferred frequency ranges for the four frequencies.

In one example, the preferred frequency ranges are as follows:

• • 0-50 kHz;
  • 2-200 kHz;
  • 50-500 kHz; and,
  • 200-1000 kHz.

Limiting the design space based on practical constraints can lead to more specific frequency ranges as follows:

• • 0-24 kHz;
  • 5-74 kHz;
  • 77-200 kHz;
  • 530-1000 kHz.

Even more preferably, the range of frequencies can be further limited to specific values as follows:

• • 14 kHz
  • 57 kHz
  • 188 kHz
  • 679 kHz

These are based on theoretical preferred frequencies calculated to be as follows:

• • 14.1844 kHz
  • 56.9201 kHz
  • 187.5397 kHz
  • 679.1141 kHz

Determination of the frequencies allows an impedance measuring device to be provided, which is adapted to utilise the preferred frequencies, as shown for example in FIG. 5.

In this example the impedance measuring device typically includes a processing system 1 coupled to a current source 11 and a voltage sensor 12. The current source 11 is coupled via electrodes 13, 14 to a subject S with the voltage sensor being coupled to the subject S via electrodes 15, 16.

In use the processing system 1 causes the current source 11 to apply alternating current signals to the subject S via the electrodes 13, 14, at each of the preferred frequencies determined using the above described process. The response of subject S is then measured via the electrodes 15, 16 using the voltage sensor 12. The processing system 1 can then use details of the applied current and the measured voltage signals to determine an impedance value at each of the preferred frequencies.

By utilising the four measured frequencies, this allows parameters of the Cole model to be determined, such as the subject's impedance at a characteristic frequency Z_c, as well as values for the intracellular and extracellular impedance. This in turn allows information such as indicators of the subject's of intracellular or extracellular fluid, or the ratio there between, to be determined.

Specific Example

A specific example of the above process will now be described in more detail.

In this example, the Cole equation described in Cole, K S (1940) Permeability and impermeability of cell membranes for ions Cold Spring Harbor Symposia on Quantitative Biology 8 110-122, is used to model bioimpedance. The model includes nonlinear parameters, which are known to vary between individuals. Accordingly the process uses the theory of experimental design to find the frequencies at which measurements of bioimpedance are optimally made. This can also be used to determine how many frequencies are needed in each patient in order to obtain good estimates of the parameters.

Experimental Design

Much of experimental design focuses on parameter estimation. A design is optimized, that is, the best choice of covariate settings and experimental effort is chosen, through assessment of an optimality criterion. Criteria considered in this paper are D-optimality and product design optimality, which are typically used to gain good parameter estimates for one or more models. Such criteria are based on the expected Fisher information matrix described in more detail below.

A Design ξ is defined by:

$\xi =\left\{\begin{array}{cc}{x}_1& w_1\\ x_2& w_2\\ ⋮& ⋮\\ x_n& w_n\end{array}\right\}$

• • where w_i is the experimental effort given to the i-th vector of covariates x_i.

Use of the experimental effort factor is optional and included in this design for the purpose of example only and the following discussion will focus in examples in which the experimental effort is not taken into account, in which case the preferred Design ξ is defined by:

$\xi =\left\{\begin{array}{c}{x}_1\\ x_2\\ ⋮\\ x_n\end{array}\right\}$

If x_iεχ, then the design space can be written as

$\Xi =\left\{\xi {\mathrm{\varepsilon \chi }}^n\times {\left[0,1\right]}^n\text{:}\sum _{i=1}^n{\omega }_i=1\right\}.$ Fisher Information Matrix

For a model with response vector y=(y₁, . . . , y_n), dependent upon parameters θ and a design ξ, the expected Fisher information matrix can be defined as:

$\begin{array}{cc}M\left(\theta ,\xi \right)=-E\left[\frac{{\partial }^{2}l\left(\theta ;y\right)}{\partial \theta \partial {\theta }^{\prime }}\right]& \left(1\right)\end{array}$ where l(θ; y) is the log-likelihood.

The Cramer-Rao lower bound described in H. Cramer “Mathematical methods of statistics.” Princeton, N.J.: Princeton University Press, pages 474-477, 1946, states that the variance-covariance matrix of any unbiased estimator of θ is bounded below by M⁻¹ (θ,ξ). Thus, for given a design ξ, the expected Fisher information indicates how well to estimate the parameters in the model.

D-Optimality

A design ξ is D-optimal if it maximizes the determinant of the expected fisher information matrix. That is, arg max_ξ|M(θ,ξ)|. This design will minimize |M⁻¹(θ,ξ) which, minimizes the variance-covariance of {circumflex over (θ)}. Such D-optimal designs, therefore, in general provide good estimates of parameters since they minimise the ellipsoidal confidence regions around an estimate.

When comparing a designs ability to estimate model parameters, its D-efficiency is considered. The D-efficiency of the design ξ compared to the D-optimal design ξ* for a particular model is:

$\begin{array}{cc}\mathrm{Deff}=\left\{\frac{{M\left(\xi ,\theta \right)}^{1/p_1}}{{M\left({\xi }^{*},\theta \right)}^{1/p_1}}\right\}& \left(2\right)\end{array}$ where p_i is the number of parameters in the model.

A D-efficiency of 0.5 for ξ compared with ξ* means that twice as many samples need to be taken using ξ in order to obtain as accurate parameter estimates as given by ξ*.

Product Design Optimality

Product design is found when trying to obtain efficient parameter estimates for more than one model. Here the product of the determinants, scaled by the number of parameters, of each model is maximized giving a design ξ*_D1,D2 which should yield good parameter estimates under each model. That is,

$\begin{array}{cc}{\xi }_{D_1,D_2}^{*}=\arg \underset{\xi }{\max }\left\{{M_1\left({\theta }^1,\xi \right)}^{1/p_1}{M_2\left({\theta }^2,\xi \right)}^{1/p_2}\right\}& \left(3\right)\end{array}$ where M₁(θ₁, ξ) is the expected Fisher information matrix for model 1 with parameters θ₁ and p₁ is the number of model parameters. Similarly for model 2.Nonlinear Regression

A nonlinear model is specified by two main parts; a function expressing the predicted response and residual response variance structure. This, for the jth observation, can be written as follows:y_j=f(x_j,θ)+ε_j  (4)where:

• • y_j is the jth predicted response,
  • x_j are the covariates,
  • θ refers to the model parameters and
  • ε_j is error or uncertainty in the response for a given individual.

When modelling nonlinear data it is assumed that:

• • (i) E└ε_j┘=0; for j=1, . . . , m
  • (ii) Cor└ε_j,ε_l=0; for j=1, . . . , m and for l=1, . . . , m
  • (iii) Var[ε_j]=σ², and are identically distributed for all x_j for j=1, . . . , m;
  • (iv) ε_j˜N(μ,σ²) for j=1, . . . , m.

Often in practical applications, some of these assumptions do not hold. Accordingly, some generalizations or relaxations of this framework can be used so that the theory of nonlinear regression can be applied in these areas.

One relaxation is a generalization of the assumption of constant variance. Such a relaxation will allow flexibility in specifying intra-individual variance. This involves specifying a variance function h which can depend upon the predicted response E[y_j], covariates x_j and/or additional parameters δ. The model is then specified by:E└y_j┘=f(x_j,θ)Var[y_j]=σ²h²(μ_j,x_j,δ)μ_i=f(x_j,θ)

Such a specification generalizes assumption (iii) and allows for heteroscedasticity in the model.

The above defines a fixed effects model. However, to take into account variations between individuals, mixed effects models are typically used.

Fixed Effects Models

For a single individual, a fixed effects model (of the form y_i=f(x_i,θ)+ε_i) is considered. The expected fisher information matrix is defined as follows:

$\begin{array}{cc}-E\left[\frac{{\partial }^{2}l\left(\theta ;y\right)}{\partial \theta \partial {\theta }^{\prime }}\right]={\mathrm{JWJ}}^{\prime }\mathrm{where}\text{:}J=\left[\begin{array}{c}\frac{\partial f\left(\theta ,\xi \right)}{\partial {\theta }_1}\\ \frac{\partial f\left(\theta ,\xi \right)}{\partial {\theta }_2}\\ ⋮\\ \frac{\partial f\left(\theta ,\xi \right)}{\partial {\theta }_n}\end{array}\right]& \left(5\right)\end{array}$ and W is a diagonalized n×1 vector of weights and

$\frac{\partial f\left(\theta ,\xi \right)}{\partial {\theta }_i}$ is a row vector of derivatives of the model with respect to the ith parameter evaluated at the covariates in ξ.Mixed Effects Models

Mixed effects models allow for the analysis of individual data by incorporating fixed and random effects into a model. This gives two sources of variation; residuals within individuals and variation between individuals. Accordingly, the models have two types of coefficients; population-average and individual specific. One of the main applications of this theory is to repeated measures data.

It is known that where y_i is a n_i-vector of observations for the i-th individual, where i=1, . . . , N, and let the model be described by F(θ_i,ξ_i). Then:y_i=f(θ_i,ξ_i)+ε_i(s+tf(θ_i,ξ_i))where ξ_i=(x_i1, . . . , x_ini)′ is a n_i-vector of explanatory variables for the individual i, θ_i is the p-vector of individual parameters and ε_i is the n_i-vector of random errors.

Here ε is normally distributed with zero mean and a diagonal variance matrix characterized by parameters s and t so that s relates to the additive error component while t relates to the multiplicative error component.

Where θ_i is defined by β+b_i, β is the vector of fixed effects and b is the vector of random effects for individual i. Here b_i is normally distributed with mean zero and variance Ω.

The covariates for individual i is denoted by the vector x_i, and θ_i is defined by g(β,b_i,x_i), where g is a known function. Ψ is an estimate of a vector of all population parameters, let γ be the vector of all variance terms. Then, Ψ′=[β′,γ′]

The Fisher information matrix for a population design

$M\left(\Psi ,\Xi \right)=\sum _{i=1}^{N}M\left(\Psi ,{\xi }_i\right),$ where Ξ={ξ₁, . . . , ξ_N} and ξ_i is the ith individual design.

The Fisher information matrix for given values of x is given by:

$\begin{array}{cc}{E}_{y❘x}\left[-\frac{{\partial }^{2}l\left(\theta ;y\right)}{\partial \theta \partial {\theta }^{\prime }}\right]& \left(6\right)\end{array}$ where l(θ; y) is the log-likelihood of observations y for the population parameters θ. Due to the nonlinearity of f with respect to θ, there is no analytical expression for l(θ; y). That is, in general, for say N individuals where Y is a matrix of responses for the whole population and y_i is the vector of responses for the ith individual, the likelihood can be expressed as:

$\begin{array}{cc}L\left(\beta ;Y❘x,\Omega \right)=\prod _{i=1}^{N}L\left(\beta ;y_{\underset{\sim }{i}}❘x_i,{\omega }_i\right)\mathrm{where}& \left(7\right)\\ \begin{array}{c}L\left(\beta ;y_i❘x_i,{\omega }_i\right)=\int L\left(\beta ;y_i\underset{\sim }{;}{b}_i❘x_i,{\omega }_i\right)d{b}_i\\ =\int L\left(\beta ;y_i\underset{\sim }{❘}{b}_i,x_i\right)L\left(b_i;w_i\right)d{b}_i\end{array}& \left(8\right)\end{array}$

Given the nonlinearity of b_i in f, this integral is generally intractable. One solution for this integral is to approximate the nonlinear function with a first order Taylor expansion around the expected values of the random effects. The model is therefore:

$\begin{array}{cc}y\cong f\left(g\left(\theta ,\stackrel{\_}{v},x\right),\xi \right)+\frac{\partial f\left(\theta ,\stackrel{\_}{v},x\right),\xi }{\partial v^{\prime }}\left(v-\stackrel{\_}{v}\right)+\varepsilon \left(s+\mathrm{tf}\left(g\left(\theta ,\stackrel{\_}{v},x\right),\xi \right)\right)& \left(9\right)\end{array}$

Then the log-likelihood l is approximated by:

$\begin{array}{cc}-2l\left(\theta ;y\right)\cong n1n2\pi +1nV+\sum {\left(\underset{\sim }{y_i}-E_i\right)}^{\prime }{V}^{-1}\left(\underset{\sim }{y_i}-E_i\right)& \left(10\right)\end{array}$ where E and V are the marginal expectation and variance of y given by:

$E=E\left(y❘x\right)\cong f\left(g\left(\beta ,\stackrel{\_}{v},x\right),\xi \right)-\frac{\partial f\left(g\left(\beta ,\stackrel{\_}{v},x\right),\xi \right)}{\partial v^{\prime }}\stackrel{\_}{v}$ $\begin{array}{c}V=\mathrm{Var}\left(y❘x\right)\\ =\left[\frac{\partial f\left(g\left(\beta ,\stackrel{\_}{v},x\right),\xi \right)}{\partial v^{\prime }}\right]{\Omega \left[\frac{\partial f\left(g\left(\beta ,\stackrel{\_}{v},x\right),\xi \right)}{\partial v^{\prime }}\right]}^{\prime }+\\ {\mathrm{diag}\left(s+\mathrm{tf}\left(g\left(\beta ,\stackrel{\_}{v},x\right),\xi \right)\right)}^2\end{array}$

Then the Fisher information matrix can be expressed as:

$\begin{array}{cc}M\left(\Psi ,v,x,\xi \right)=\frac{1}{2}\left[\begin{array}{cc}A\left(E,V\right)& C\left(E,V\right)\\ C^{\prime }\left(E,V\right)& B\left(E,V\right)\end{array}\right]\mathrm{where}{A\left(E,V\right)}_{\mathrm{mn}}=2\frac{\partial E^{\prime }}{\partial {\beta }_m}{V}^{-1}\frac{\partial E}{\partial {\beta }_n}+\mathrm{tr}\left(\frac{\partial E^{\prime }}{\partial {\beta }_n}{V}^{-1}\frac{\partial E}{\partial {\beta }_m}{V}^{-1}\right),\mathrm{for}m\mathrm{and}n=1,\dots ,\dim \left(\beta \right){B\left(E,V\right)}_{\mathrm{mn}}=\mathrm{tr}\left(\frac{\partial V}{\partial {\gamma }_m}{V}^{-1}\frac{\partial V}{\partial {\gamma }_n}{V}^{-1}\right),\mathrm{for}m\mathrm{and}n=1,\dots ,\dim \left(\gamma \right){C\left(E,V\right)}_{\mathrm{mn}}=\mathrm{tr}\left(\frac{\partial V}{\partial {\gamma }_m}{V}^{-1}\frac{\partial V}{\partial {\beta }_n}\right),\mathrm{for}m=1,\dots ,\dim \left(\beta \right)\mathrm{and}n=1,\dots ,\dim \left(\gamma \right)& \left(11\right)\end{array}$

Thus, the approximate expected information matrix for a nonlinear mixed effects model is formed. This can now be used in a variety of optimality criteria to form designs with desired properties.

Bioelectrical Impedance Study

The application of these techniques to Bioelectrical impedance analysis (BIA), as modeled using the Cole equation, will now be explained. In particular D-optimal designs are derived based on different assumptions about the interaction between frequency and impedance and also the practical limitations of such a study.

This is achieved by obtaining initial parameter estimates by modeling data on individuals, using these initial estimates to form D-optimal designs and then extending these D-optimal designs so that they perform well in practice.

Optimal Designs for Studying Bioimpedance

In this example the Fisher information matrices are determined using the software package POPT written by S. B. Duffull. The search routine uses an adapted version of a simulated annealing algorithm for continuous variables. Together these techniques provide a means of finding D-optimal designs for nonlinear mixed and fixed effects models across a continuous search space.

Modeling the Data

For the Cole equation, initial parameter estimates were found by modeling paired frequency and bioimpedance data on 61 subjects. Whole-body, wrist to ankle, multifrequency bioimpedance data were recorded over the frequency range 5 to 1024 kHz using an Impedimed-SEAC SFB3 tetra-polar impedance instrument.

Modelling was performed using the software package MONOLIX, which is a MATLAB based package that uses stochastic approximation to the expectation maximization (SAEM) algorithm, in parametric maximum likelihood estimation for nonlinear mixed effects models.

The Expectation-Maximization (EM) algorithm is an iterative procedure used for likelihood function maximization given incomplete data sets formed from observable and non observable data. The E-step in this algorithm cannot be performed in closed formed due to the nonlinearity in f of the random effects, as described above.

The SAEM algorithm replaces this step with a stochastic procedure. Thus, the usual EM algorithm computes, at the kth iteration, the conditional expectation of the log-likelihood giving Q_k(Ψ)=E[log p(y,b_i;Ψ)|y,Ψ_k-1)], where p(y,b_i;Ψ) is the likelihood of (y,b_i).

The SAEM replaces this step by drawing b_i^(k) from the conditional distribution p(.|y;Ψ_k) and updates Q_k(Ψ) as follows:Q_k(Ψ)=Q_k-1(Ψ)+δ_k(log p(y,b_i^(k);Ψ)−Q_k-1(Ψ))

• • where δ_k is a decreasing sequence of positive numbers.

To fit a nonlinear model to data, initial parameter estimates are also needed. For this model, initial estimates of fixed effects were taken from previous studies, and estimates of variance were found by what is called a ‘Standard Two Stage Approach’. This involves estimating fixed effects for all data specific to each individual. Then, the variance of these estimates is calculated and used as an estimate for Ω.

The first stage of the modeling process was to find a suitable residual variance model such that the residuals are normally distributed around zero so that tests, such as the likelihood ratio test, are valid. Given the initial estimates, the first run uses a diagonal variance-covariance matrix with all four parameters having a normally distributed random component and the model having additive residuals. The residual plot is shown in FIG. 6.

This shows that the assumption about additive variance holds reasonably well and the residuals are centered around zero. Whilst the plot is not completely random at low frequencies, due for example to technical errors of measurements and inaccuracies associated with using the Cole equation to model bioelectrical impedance, this is considered acceptable.

This additive variance model can then be compared with the multiplicative and additive plus multiplicative error models. The residual plot was best when the additive error model was used.

The initial model can then be compared with different models having various fixed effects parameters. For example, the initial model can be compared with a model setting the parameter c to have no random component to it, that is having a fixed effect only. The likelihood ratio test for nested models can then be used to compare these models with the initial run. In this instance it was found that the initial run significantly increased the log-likelihood to justify having all parameters in the model.

The final step of the modeling process is to allow some or all of the random effects to have a log-normal distribution. This shows that found parameter estimates were not biologically plausible and can therefore be disregarded.

In the end, the initial model chosen to start the modeling process was the one chosen based on the above reasoning. It was also the model that provided the maximum value of the log-likelihood of (y; θ).

Using the model described above, the final estimates can be found in table 1, where the random effects are normally distributed around zero with variance SZ; a diagonal matrix with diagonal elements [ω₁², . . . , and ω₄²] is the residual variance.

	TABLE 1
	Parameter	Estimate (CV %)
	R0	861.00	(2.04)
	Rinf	320	(1.58)
	C	2.4e−006	(2.08)
	α	0.649	(1.07)
	ω12	1.92e+004	(18.39)
	ω22	1.66e+003	(19.04)
	ω32	1.64e−013	(19.27)
	ω42	0.00302	(16.39)
	s2	5.46	(0.93)

It will be appreciated that these values are derived from actual physical measurements and will therefore depend on collected data. The above values are therefore for the purpose of illustration only and are not intended to be limiting.

A coefficient of variation (CV %) is calculated by dividing the standard error of an estimate by its estimate and multiplying this by 100 to form a percentage. Typically, acceptable CV %'s for fixed parameters and random parameters are 20% and 50%, respectively. From the table 1, we can see that the CV %'s for our estimates are more than reasonable.

D-Optimal Design for Fixed Effects Model

Examining a design for one individual using the fixed effects model will now be described using the initial estimates found in the previous section, θ=[R₀, R_inf, C, α]′=[861, 320, 2.4e−006, 0.649]′. The design which maximises the determinant of the information matrix is:

${\xi }_{D_{\mathrm{fixed}}^{*}}=\left\{\begin{array}{c}0.0001\\ 3.7357\\ 31.8801\\ 1000.0000\end{array}\right\}$

However, as described above, this is for the values shown in Table 1 and if different values are used, a different matrix will be determined.

D-optimality generally selects extreme values of covariates, but this also occurs to an extent due to the nature of the Cole equation. If ω approaches zero, the Cole equation will approach Z=R₀. Alternatively, if ω approaches infinity Z=R_∞. Thus, extreme frequencies provide simpler expressions for R₀ and R_∞.

The calculation of the D-efficiency will be described below, providing an indication of the limitation of not accounting for between subject variability seen in the next section.

Mixed Effects

For the Cole equation, parameter estimates are known to vary between subjects, and accordingly a mixed effects model must be used.

For the purpose of this analysis it will be assumed that it is possible to administer electrical currents through the, body at frequencies [0, 1000] kHz, and obtain an accurate reading of bioelectrical impedance Thus, this can assume a model:y_i=f(θ,ξ)+sε_i  (15)

Using the initial estimates from the fixed effects model determined above, and shown in table 1, the following D-optimal design was found by maximizing the determinant of equation 11:

${\xi }_{D_{\mathrm{add}}^{*}}=\left\{\begin{array}{c}0.0001\\ 5.0374\\ 77.4409\\ 1000.0000\end{array}\right\}$

When forming the expected fisher information matrix for this design, C(E, V) is set to a block matrix of zeros. Then, under this assumption, the fixed and random effects of the model are independent.

The D-efficiency of the fixed effect design under the mixed effects model is 0.9692. Such a high efficiency of the fixed effects D-optimal design suggests that, in this case, accounting for between subject variance only marginally improves the ability to estimate the parameters of the Cole equation.

Dealing with Practical Limitations

With bioelectrical impedance studies, it is known that low and high frequencies produce more highly variable readings of bioimpedance compared with less extreme frequencies.

Two possible approaches to account for this variability will now be described. The first restricts the design space to exclude frequencies where readings of bioimpedance are inaccurate. The second allows the residual variance to depend on the frequency in some manner that mimics real life patterns.

D-optimal Design with Restricted Frequencies

The straight forward approach to dealing with such a variance structure is to avoid frequencies which produce highly variable bioimpedance measures. At present, such bioimpendance studies limit the frequencies to being between a variety of ranges due to this added variability. Examples of such ranges are [4, 200], [4, 500] and [4, 750]. These intervals, therefore are used to define restricted designs.

Allowing parameters to vary between individuals the following D-optimal designs were found for the above example:

TABLE 2
		D-eff under Mixed
Range	Frequency	Effects Design
[4, 200]	4.0000	0.4320
	15.1909
	60.4311
	200.0000
[4, 500]	4.0000	0.5977
	18.8903
	114.7364
	500.0000
[4, 750]	4.0000	0.06752
	21.5242
	143.0364
	750.0000

The D-efficiencies of these restricted designs under the mixed effects models are also shown in table 2. As the range becomes more restricted, the efficiencies decrease. This is highlighted by the D-efficiencies of these designs, which suggest that, under the assumption of constant additive variance, restricting the frequencies to certain intervals creates a significant loss in our ability to estimate parameters precisely. Consequently we will be forced to sample more often in the restricted case to obtain as precise estimates as in the unrestricted case.

D-optimal Design with Variance Function

An assessment of the preferred model can be made by adding a variance function into the model that mimics the variance of bioimpedance in real life studies.

In this case, once the true variance structure has been captured by the function, the D-optimal design will provide the best estimates of the parameters. Inherent in this design will be the ideal range to restrict our frequencies, based on the variance function used.

To achieve this, a model of the following form is used:y_i=f(θ_i,ξ_i)+ε_i(s₁+s₂h(ξ₁))  (16)where h(ξ_i), a function of frequency, specifies the variance function.

Relaxing the assumption about constant variance of the residuals by forming a variance function described above, allows accurate capture and mimicking of the residual structure, without violating any nonlinear regression assumptions.

Bioimpedance is made up of two parts, resistance and reactance, relating to the real and imaginary parts of the Cole equation, respectively. An example plot of resistance vs reactance is shown in FIG. 7. This plot is of real data and as such shows the real life deviations from the theoretical semi-circle that would be produced from the pure circuit seen in FIG. 4.

The peak of the locus of the semi-circle identifies the characteristic frequency f_c. This frequency yields the smallest variance due to the variance function and consequently, it is assumed that at this frequency the constant additive variance dominates. The variance function needs to be such that there is relatively small variance between the chosen interval, for example [4, 750], with rapidly increasing variance for frequencies outside these bounds.

Accordingly, a double exponential model, centred around f_c, can be used to adequately model the variance of such a study. Thus, for the range 0 kHz to 1000 kHz, frequencies between f_c and 1000 are rescaled to be between 0 and 1. Similarly, frequencies between 0 and f_c are rescaled so that they are between 1 and 0. The new vector of rescaled frequencies is then exponentiated twice, and scaled such that the vector has a minimum and a maximum at 0 and 1, respectively. Finally, slightly extra variability is added to frequencies greater than f_c by multiplying the resealed numbers which relate to these frequencies by 2. This whole vector is then multiplied by σ_extra².

Example residual plots of this variance function h(ξ) for various σ_extra² values are shown in FIG. 8. In order to illustrate the variation of impedance at small frequencies, the log of frequency is shown in all four plots.

From this it can be seen that the actual structure of the function does not change, but is merely at a different scale. This also shows how variable readings of bioimpedance can be at low and high frequencies, particularly when σ_extra² is large, and how there exists relatively small variance between 4 and 750.

This leads to another consideration for this variance structure, in particular, how much should bioimpedance vary at low and high frequencies, and therefore how large should σ_extra² be such that it will capture the real life variance structure of bioimpedance.

Table 3 shows D-optimal designs found for various values of σ_extra². The D-efficiencies shown refer to the efficiency of each design compared with the D-optimal design described above. For all designs, the model considered is from equation 16.

TABLE 3
σextra2	ξD*multi	D-eff
0	0.0001	1.0000
	5.0371
	77.4409
	1000.000
0.1	0.0001	0.6837
	15.3454
	109.5900
	699.9127
0.2	14.6420	0.6383
	59.5391
	186.4495
	680.5794
0.5	20.6291	0.5831
	68.7527
	183.2815
	595.1341
1	25.7367	0.5409
	74.3527
	175.7765
	530.8203

This shows that as the variance increases at the low and high frequencies, the criterion starts to choose frequencies at the less extreme values, thereby avoiding regions which relate to a highly variable reading of bioimpedance. Thus, it is up of the experimenter to decide how variable bioimpedance can be. Once this is decided, the D-optimal design will take this variability into account and find the design which will give the best estimates of parameters.

For example, at a first instance, a broad set of preferred frequency ranges can be defined as follows:

• • 0-50 kHz;
  • 2-200 kHz;
  • 50-500 kHz; and,
  • 200-1000 kHz.

However, using a more limited design space based on the above mentioned practical constraints can lead to more specific frequency ranges as follows:

• • 0-24 kHz;
  • 5-74 kHz;
  • 77-200 kHz;
  • 530-1000 kHz.

In a more extreme limitation, for example, assuming σ²=0.2 is a true representation of the variance, then the optimal design can be seen in the above table, indicating that measurements at frequencies lower than 14.64 kHz and higher than 680.60 kHz are too variable and should be avoided.

Uncertainty in Prior Estimates

Inaccurate prior estimates of parameters for nonlinear models can lead to designs that have a limited ability to estimate parameters that diverse considerably from what was expected. In order to find a design that offers efficient parameter estimates across a wide range of prior estimates, we can look at product design optimality discussed above.

A product design is generally more robust to changes in initial parameter estimates than local D-optimal designs for each parameter set. Thus, a product design is preferred if uncertainty exists in the initial parameter estimates.

Product Design

Given the initial parameters estimates above, 5th and 95th percentiles of the population can be found using the estimates of the between subject variability. For example, given the initial estimate of R₀=861 with ω₁=138, the 5th and 95th percentiles would be 584 and 1138, respectively.

Such percentiles can be determined for all four model parameters giving a set of eight different initial estimates which would be used as eight different impedance models in the product design. The eight parameter sets considered can be seen in table 4.

TABLE 4
Model	R0	Rinf	C	α
1	1138	370	2.40e−06	0.6490
2	584	370	2.40e−06	0.6490
3	861	401	2.40e−06	0.6490
4	861	239	2.40e−06	0.6490
5	861	320	3.21e−06	0.6490 ′
6	861	320	1.59e−06	0.6490
7	861	320	2.40e−06	0.7589
8	861	320	2.40e−06	0.5391

The product design is formed by optimizing over the product of the determinants of the respective variance-covariance matrices of the eight models as a single function. The idea behind forming product designs in this way is that optimizing across all eight models will provide a design which is efficient across a wide range of parameter estimates. The model considered is of the form of equation 16 (σ_extra²=0.2), the values for which are set out in Table 3 as:

${\xi }_{D_{\mathrm{multi}}^{*}}=\left\{\begin{array}{c}\mathrm{14.6426`}\\ 59.5391\\ 186.4495\\ 680.5794\end{array}\right\}$

The following product design is given by:

${\xi }_{{\left(D_1,\dots ,D_8\right)}_{\mathrm{multi}}^{*}}=\left\{\begin{array}{c}14.1844\\ 56.9201\\ 187.5397\\ 679.1141\end{array}\right\}$ Design Evaluation

To analyze how efficient the product design is at estimating parameters across all models, the efficiency of the design under each model is determined, as shown in table 5, in which the efficiency of the product design is compared to each D-optimal design for each parameter set.

TABLE 5
Model D-eff
1     0.9923
2     0.9932
3     0.9890
4     0.8129
5     0.9573
6     0.7298
7     0.9982
8     0.9997

The efficiency of the product design can be calculated by comparing its D-value under each model compared to the D-optimal value under the respective model. Table 5 shows the efficiency of the product design compared to each D-optimal design for each parameter set. Relatively high efficiencies across all models suggest that the product design is a good means for allowing for uncertainty in initial parameter estimates. Further, given the range of parameter values considered, it is believed that these designs will efficiently estimate model parameters for a variety of individuals. Hence, this approach and the frequencies found should benefit bioimpedance analysis.

A practical approach, applies real data to determine how well model parameters can be estimated. Given the design:

${\xi }_{{\left(D_1,\dots ,D_8\right)}_{\mathrm{multi}}^{*}}=\left\{\begin{array}{c}14.1844\\ 56.9201\\ 187.5397\\ 679.1141\end{array}\right\}$ and the data on 61 individuals, impedance at these frequencies were extracted from the data set of all individuals and a nonlinear mixed effects model was fitted to this extracted data set.

Using the same model and initial parameter estimates as discussed in the modelling section above, MONOLIX provides estimates for the parameter values as shown in table 6 and estimates of coefficients of variation found in table 7.

	TABLE 6
	Parameter	Full Data Set	Extracted Data Set
	R0	861	879
	Rinf	320	303
	C	2.4e−006	3.25e−006
	α	0.649	0.68
	ω12	1.92e+004	1.1e+04
	ω22	1.66e+003	1.52e+003
	ω32	1.64e−013	5.31e−012
	ω42	30.2e−003	1.37e−003
	s2	5.46	5.24

FIG. 9 shows the residual plot for this fit, indicating that the nonlinear regression assumptions discussed above hold.

The new estimates shown in table 6 are similar to those found using the full data set of observations. Given this and the residual plot, it is concluded that this D-optimal design performs well in practice and as such, with the use of only four frequencies, we have been able to form estimates of parameters of the Cole equation similar to those found by using the full data set.

It is also clear that the product design performs well in practice and requires of only four is frequencies per individual. That is, the optimal choice of these frequencies yields the ability to efficiently estimate model parameters. Further, it highlights the apparent lack of information gained by choosing over 100 unique frequencies per individual to model bioimpedance.

Table 7 shows that the coefficients of variation estimated from the full data set are similar to those estimated from the extracted data set. This suggests that only four measurements of impedance per individual are needed to efficiently estimate the parameters of the Cole equation.

	TABLE 7
	Full Data Set	Extracted Data Set	Expected
	R0	2.04	1.76	4.15
	Rinf	1.58	2.51	6.33
	C	2.08	11.11	14.24
	α	1.07	1.57	4.74
	w12	18.39	20.09	19.14
	w22	19.04	22.04	42.18
	w32	19.27	24.48	177.34
	w42	16.39	28.69	36.82
	s2	0.93	12.35	9.80

Table 7 also shows the expected coefficients of variation calculated by using the expected Fisher information matrix to form the expected standard errors of estimates.

Accordingly, this shows that experimental design for nonlinear fixed and mixed effects models can be applied BIA, allowing preferred frequency measuring models to be determined. Numerous four frequency designs were found relating to D-optimal designs based on various assumptions about how frequency relates to bioimpedance and how parameters vary between individuals. Product designs were also found. These designs should be robust to changes in initial estimates.

D-optimal designs can also be applied to real data where it was shown to perform well against the full data set. It is hoped that the D-optimal design approach used here and the designs found will aid BIA researchers in the design of optimised multiple frequency BIA instrumentation. This will mitigate the acknowledged inadequacy of some currently used instrumentation and provide for better clinical utility, for example, the accurate prediction of lean body mass for drug dosing in the obese.

Persons skilled in the art will appreciate that numerous variations and modifications will become apparent. All such variations and modifications which become apparent to persons skilled in the art, should be considered to fall within the spirit and scope that the invention broadly appearing before described.

Claims

1. A method of measuring the impedance of a subject, the method including, in a measuring device including a processing system: a) causing one or more electrical signals to be applied to the subject using a first set of electrodes, the one or more electrical signals having four frequencies;b) measuring electrical signals across a second set of electrodes applied to the subject in response to the applied one or more signals;c) determining from the applied signals and the measured signals at least one measured impedance value at each of the four frequencies, wherein the four frequencies are:i) 14 kHz;ii) 57 kHz;iii) 188 kHz; and,iv) 679 kHz.

2. A method of measuring the impedance of a subject, the method including, in a measuring device including a processing system: a) causing one or more electrical signals to be applied to the subject using a first set of electrodes, the one or more electrical signals having four frequencies;b) measuring electrical signals across a second set of electrodes applied to the subject in response to the applied one or more signals;c) determining from the applied signals and the measured signals at least one measured impedance value at each of the four frequencies, wherein the four frequencies are:i) 14.1844 kHz;ii) 56.9201 kHz;iii) 187.5397 kHz; and,iv) 679.1141 kHz.

3. A method according to claim 1, wherein the method includes determining one or more impedance parameter values based on the plurality of measured impedance values.

4. Apparatus for measuring the impedance of a subject, the apparatus comprising: a first set of electrodes;a second set of electrodes; anda processing unit;wherein the processing unit is configured to:a) cause one or more electrical signals to be applied to the subject using the first set of electrodes, the one or more electrical signals having four frequencies;b) measure electrical signals across the second set of electrodes applied to the subject in response to the applied one or more signals; andc) determine from the applied signals and the measured signals at least one measured impedance value at each of the four frequencies, wherein the four frequencies are:i. 14 kHz;ii. 57 kHz;iii. 188 kHz; andiv. 679 kHz.

5. Apparatus according to claim 4, the measuring device including: a) a current source for applying current signals to the subject;b) a voltage sensor for measuring voltages signals across the subject; and,c) the processing system coupled to the current source and the voltage sensor for: i) causing the current signals to be applied to the subject; and,ii) determining the impedance using the applied current signals and measured voltage signals.

6. Apparatus for measuring the impedance of a subject, the apparatus comprising: a first set of electrodes;a second set of electrodes; anda processing unit;wherein the processing unit is configured to:a) cause one or more electrical signals to be applied to the subject using the first set of electrodes, the one or more electrical signals having four frequencies;b) measure electrical signals across the second set of electrodes applied to the subject in response to the applied one or more signals; andc) determine from the applied signals and the measured signals at least one measured impedance value at each of the four frequencies, wherein the four frequencies are:i. 14.1844 kHz;ii. 56.9201 kHz;iii. 187.5397 kHz; andiv. 679.1141 kHz.