I E E E TRANSACTIONS O N BIOMEDICAL ENGINEERING. VOL 39. NO. I .

JANUARY 1992

R. E. Bolz and G. L. Tuve, "Liquids." in CRC Hut7dhook Applircl Engineering Science. Boca Raton, FL: CRC Press, 1973, p. 93. G . K. Weber and W. J . Teale, "Interaction of proteins with radiation," in The Proteins. Vol. 3 , 2nd ed., Neurath H. Ed. New York: Academic, 1965, pp. 445-521. R . E. Marc and H. G . Sperling, "Chromatic organization of primate cones," Science, vol. 196, pp. 454-456, 1977. Y. Le Grand. "The anatomy of the retina," in Light, Colour. a d Vision. New York: Chapman and Hall. 1968, pp. 357-374. R. Klein. "The epidemiology of diabetic retinopathy: Findings from the Wisconsin epidemiology study of diabetic retinopathy." /tit. Ophthul. Clinics, vol. 27, pp. 230-238, 1987. J . Cunha-Vaz. F. De Abreu, A. Campos, and G. Figo, "Early breakdown of the blood-retinal barrier in diabetes," Brit. J . Ophthul.. v o l . 59. pp. 649-656, 1975. S . K. Williams, J . J . Devenny, and M. W. Bitensky. "Micropinocytic ingestion of glycosylated albumin by isolated microvessels: Possible role in pathogenesis of diabetic microangiopathy." in Pro(.. Nut. Acud. Sri.. vol. 78, pp. 2393-2397, 1981. B . O'Brien. "Vision and resolution in the central retina," J . Opt. Soc. A m e r . , vol. 41. pp. 882-894. 1951. M. Marre and E. Marre, "The blue-mechanism in diseased eyes with eccentric fixation," Docum. Ophthul. Proc. Series, vol. 3 3 . pp. 133138, 1982. V. C. Smith, J . Porkorny. and K. R. Diddie, "Color matching and Stiles-Crawford effect in central serous choroidopathy," Moclcjrti Prob. Ophthul.. vol. 19, pp. 284-295, 1978.

I . I N T R O D ~ JI C ION
When biological signals are measured with any significant degree of accuracy and temporal resolution their subsequent analysis invariably involves the fitting of some kind of model. which is usually done by finding those values of the model parameters that minimize the sum of the squared residuals (SSR) between the data and the fit [1]-[3]. Once the best-fit values of the model parameters have been determined they are, of course, of little use unless accompanied by some measure of how precisely they are determined by the data. In the general case one can search the parameter space for those combinations of parameter values that cause the SSR t o be some specified amount above its minimum value 141. T h e limits of this space then define the respective parameter sensitivities. If the structure of the model is sufficient to enable it to mimic all the deterministic parts of the data then the residuals are merely random measurement noise. O n e can then assign a probabalistic interpretation to the SSR from which it may be possible to derive statistically motivated confidence intervals about the estimated parameter values [1]-[4]. In practice one often finds that the residuals between the data and the fit are not entirely random but instead exhibit systematic deviations either side of zero. indicating that the model structure is not flexible enough to properly account for the deterministic parts of the data. T h e standard recipe for dealing with this state of affairs is to increase the order of the model (i.e., increase the complcxity of the model structure and hence the number of parameters) until the SSR ceases to decreasc significantly with addition of new parameters a s dictated by. for example. the F-ratio test [ I ] . Hopefully, the final model will fit the data sufiicicntly well that the residuals will exhibit the ncccssary stochastic characteristics to enable confidence intervals t o be calculated on the parameters. T h e problem with this approach. however. is that one may end up with a model that is a good deal niorc complicated than that which originally motivated the exercise, and one is then left with having t o interpret the final model structure. This is frequently impossible in the biological sciences because there are invariably a number of different physiologically plausible realizations of a model of a given order. and one usually never knows which one is the most appropriate for describing the real system. T h u s , although statistically satisfactory confidence intervals on the parameters may have been obtained. o n e is no longer able to interpret the parameters themselves in any meaningful way. In this paper we attempt to deal with this apparent impasse by examining thc question of parameter value confidence when the order of the model is clearly less than that of the system being modeled. Although in dealing with this situation we need to develop a somewhat different concept of parameter value confidence to that usually invoked, we avoid the interpretational difficulties associated with more complicated models. We illustrate our approach applied to multiple linear regression with a specific problem arising in the modcling of rcspiratory system mechanics.

A Nonstatistical Approach to Estimating Confidence Intervals About Model Parameters: Application to Respiratory Mechanics
J . H . T. Bates and A . - M . Lauzon

Abstract-Estimates of parameters obtained by fitting models to physiologic data are of little use unless accompanied by confidence intervals. The standard methods for estimating confidence intervals are statistical, and make the assumption that the fitted model accounts for all the deterministic variation in the data while the residuals between the fitted model and the data reflect only stochastic noise. I n practice, this i s frequently not the case, as one often finds the residuals to he systematically distributed about zero. I n this paper, we develop an approach for assessing confidence in a parameter estimate when the order of the model i s clearly less than that of the system being modeled. Our approach does not require a parameter to have a single value located nithin a region of confidence. Instead, we let the parameter value vary over the data set i n such a way as to provide a good fit to the entire data set. We apply our approach to the estimation of the resistance of the respiratory system i n which a simple model i s fitted to measurements of tracheal pressure and flow by recursive multiple linear regression. The values of resistance required to achieve a good fit are represented as a modified histogram in which the contribution of a particular resistance value to the histogram i s weighted by the amount of information used in its determination. Our approach provides parameter frequency distribution functions that convey the degree of confidence one may have i n the parameter, while not being based on erroneous statistical assumptions.

11. THEOKY

Manuscript received August I , 1991; revised September 18. 1991. This work was supported by the Medical Research Council of Canada (MRC) and the EL1JTC Memorial Research Fund. J . H. T. Bates is a scholar o f the MRC. A.-M. Lauzon was supported by the Research Institute of the Royal Victoria Hospital, Montreal. The authors are with Meakins-Christie Laboratories, McGill University, Montreal. P.Q. H2X 2P2. Canada. IEEE Log Number 9104297.

Models

of Respirutory

S y s t e t ~ iMechrrtiicJ

The mechanical characteristics of thc respiratory system are important determinants of our ability to breathe and may become scverly compromised in disease. Thus, it is important to understand the nature of these characteristics and also t o be able to monitor them in patients. A convenient means of probing the respiratory

0018-9294192SO3.00

0 1992 IEEE

IEEE 'TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 39, NO. I , JANUARY 1992

95

0.4

a
Y

r J J 0.00

m
3 U
. . I

a
-0.05

al

-0.8

'

10

20

10

20

(SI

Fig. 2 . (a) Measured tracheal pressure (solid line) and predicted pressure (dashed line) provided by the single-compartment linear model [see ( I ) ] fitted by multiple linear regression to the data shown in Fig. I . (b) Residuals between measured tracheal pressure and that predicted using the recursively estimated parameters R . E , and K .

(SI

Fig. I . Respiratory data obtained from an anesthetized, paralyzed dog during mechanical ventilation: (a) tracheal pressure, (b) tracheal flow. and (c) integral of tracheal flow (i.e.. volume). system from a mechanical point of view is provided by the prcssure-flow relationships that can be easily and accurately measured at the mouth or airway opening of a patient or animal being mechanically ventilated. Fig. 1 shows typical records of tracheal pressure (Ptr), tracheal flow (b'), and integrated flow ( V ) measured in an anesthetized, paralyzed, tracheostomized dog. The data were low-pass filtered at 20 Hz with an 8-pole Bessel filter and sampled at 50 Hz. Ptr was measured with respect to atmosphere with a differential piezoresistive pressure transducer (Microswitch 1 6 3 P C O I D 3 6 , Honeywell, Scarborough, Ontario) and b' was measured with a # 2 Fleisch pneumotachograph connected to another differential piezoresistive pressure transducer. The simplest physiologically sensible model one could fit to the data in Fig. 1 is the classic single-compartment linear ( R C ) model whose equation is Ptr(r) = b'R

VE

+K

(1)

where R and E are parameters commonly refcrred to a s resistance and elastance of the respiratory system, respectively, and K is the pressure in the model when b'and V are zero. This model has been employed for many years to account for measured pressure-flow relationships, and yields in R and E quantities that have been enormously valuable in assessing, for example, the importance of mechanical changes produced by bronchoactive drugs IS]. Fig. 2(a) shows, however, that this simple modcl does not account perfectly for the relationships betwcen Ptr, b', and V . Indeed, the residuals between the measured Ptr and the fitted Ptr [Fig. 2(b)] are highly systematic. In recent years considerable progress has been made in understanding how this model should be improved to better account for respiratory data. It has been shown, for example. that the respiratory system of a normal dog is much more accurately modeled during normal ventilation as a single uniformly ventilated alveolar region surrounded by lincarly viscoelastic tissue [6]. Unfortunately, however, the next step in model sophistication is usually unclear. One could add more compartments, o r alternatively make the resistances and elastances of the existing model elements nonlinear (that is, they could depend on flow and volume). Both approaches increase the number of model parameters and thus the goodnessof-fit provided by the model. but the most appropriate direction to take depends on the particular perturbation applied to the system, such as the depth and rate of ventilation. When the lungs become diseased even the situation at the linear two-compartment level is not clear, since disease can induce significant regional differences in lung mechanical properties 171. Thus, although particular models of two or more compartments may be shown to be applicable to the respiratory system in particular circumstances. the only model that

96

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL 39. NO. I , JANUARY 1992

can be applied with uniform confidence to virtually all data from patients and experimental animals remains the single-compartment linear model described by ( I ) . Indeed, this model still forms the analytical basis for most bronchopharmacological studies. and its parameters R and E continue to serve as useful indexes of the dissipative and elastic properties of the respiratory system.

Parameter Vuriability f r o m Recursit~eMultiple Linear Regression

In order to apply our new approach to the data in Fig. I , the model [see ( l ) ] must be fit to sections of data that are sufficiently small that the fit to each section is satisfactory (that is. the residuals from each fit reflect only the noise in Ptr). A convcnient way to achieve this is by using the recursive lcast squares (RLS) algorithm with an exponential memory [SI. Indeed, such an algorithm as recently been applied to the estimation of respiratory mechanical parameters [SI. [ l o ] . This algorithm allows the parameters R , E , and K to be updated each time a new set of measurements of Ptr, V. and V arrives. In terms of the general linear model fonnalism developed above, if A, is the estimatcd parameter vector obtained from the first k measurements of the dependent variable y and the dependent variable vcctor X, then the estimated parameter vector I mcasurements is obtained from the first k

Confidence Intervals in Multiple Linear Regression

The problem we face with fitting the single-compartment model to respiratory data is that of obtaining confidence intervals on the parameters R and E . The parameter values themselves are readily obtained by multiple linear regression using Ptr(t) as the dependent variable and v(t), V(r), and 1 a s the independent variables, as follows. Let there be N measurements for the variables at equally spaced points in time. Denote the N values of Ptr(r) as y,, i = I , . . . , N, and express them collectively as a single N-element column vector Y . Denote the N values of b'(t), V ( t ) and I as x,,. xi?. and x , ~ ,respectively, and express them collectively as the N x 3 matrix X. Express the three model parameters collectively as a three-element column vector A. The relationship between Y and X in terms of ( I ) then becomes

A,+I
where

A,

+ P,+IX,+I(?.,+I- x : + , A , , / ( P + x:+,P,x,+,)

P,+,=

(l/P)

[PA

P,x,+,x:+,P,/(P +

x ~ * l ~ , x , + ,(6) )l.

Y=XA+C

(2)

where C is an N-element column vector containing all differences between the measured y , and those predicted by ( 1 ) . The wellknown least squares estimate A of A is [ 11

The algorithm is started with A(, = 0 and Po = 10hI where I is the identity matrix. p is a constant between 0 and 1 whose value determines the effective memory length of the algorithm. If p = l the algorithm remembers all past data with equal weight. As p tends to 0 the memory length also tends to zero. The time-constant 7 of the memory is
T =

[X7X]-'X7Y

(3)

-&/In

(p)

(7)

while the covariance matrix for A is defined as c o v {A} =

(T?

[X'XI-'

(4)

where o 2 is the variance of the noise in Y. Cov {A} has a meaningful interpretation when C consists of samples of random uncorrelated noise of equal variance at each point, in which case 0' can be approximated as SSR divided by N = 3. The elements of Cov {A} may be used to estimate confidence intervals on the elements of A. In particular, the leading diagonal of Cov {A} are estimates of the variances of the corresponding elements of A. The situation we face here, however, is that C is not random at all, but instead is almost entirely deterministic as evidenced by the systematic nature of the residuals between model and fit in Fig. 2(b). Now Cov {A} has no obvious meaning and, indeed, the very notion of a confidence interval in a statistical sense becomes inappropriate when there is virtually no stochastic element to the system being modeled because the usual meaning of a confidence interval is based in probability theory. That is, given a particular parameter value, how likely is it that the value might actually be something else? This is clearly a meaningless question when the deviations between data and fit arise because the model does not have sufficient degrees of freedom to account for all the deterministic variation in the data. W e propose to deal with this situation by discarding the conventional notion of parameter confidence and replacing it with that of parameter variability. That is, rather than insisting that each parameter have a single value for describing the entire data set, let the parameter values vary in such a way as to give as good a fit as possible over the data set. The usual concept of a parameter with a single value bounded by a region of confidence is then replaced by that of a parameter function whose range of values is used to make comparisons between estimations from different data sets.

where 6t is the data sampling interval. The value of T determines the effective length of the portion of data to which the model is being fitted at any instant. If 7 is large then the model will have difficulty describing all the systematic variation in each portion. similar to the situation cxcmplified in Fig. 2. If T is too small then the model will be able to describe the noise in each portion. What we seek is an intermediate 7 that allows the model to describe the deterministic variation in each portion, but not the noise. Fig. 3(a) shows the recursive estimate of R obtained with 7 = 1000 s, so that the RLS algorithm effectively remembers all past data. R shows some initial variation but quickly settles down to a stable value which has the same value at 20 s as the value of R produced by fitting the entire data set to the model by regular multiple linear regression. Fig. 3(b) shows the value of P(1, l ) (the first element in the leading diagonal of P,which corresponds to R ) . P ( 1 , 1 ) has an initial value of 10'. but rapidly and monotonically decreases towards a small value reflecting the large amount of information available in the entire data set for determining R. Fig. 3(c) shows the residuals between the measured Ptr and the values calculated using the recursively estimated parameters. The residuals show systematic variations that are very similar to those obtained by multiple linear regression (Fig. 2), which is not surprising because the recursively estimated parameters approached values close to those obtained by multiple linear regression in only a few seconds. Figs. 4-6 show plots corresponding to those in Fig. 3 , except that T has values of 1.0, 0.2, and 0.05 s, respectively. As 7 decreases the oscillations in R with each breath become more pronounced because each value of R is based on less data. The corresponding values of P( 1 , l ) also become larger with greater oscillations again reflecting the decreased information utilized to determine R at each point. Note that P(1.1) achieves its largest values (corresponding to the least confidence in R ) at the beginning

IEEE TRANSACTlONS ON BIOMEDICAL ENGINEERING. VOL. 39. NO

1

I . JANUARY 1992

rl

I
l-4

a
Y

(D

-1

1.0

N I N

l-4

rl 4

m -

0.5

a
0.0 0.05

I
10

20

n (0

d
(0
Y

0.051

m 0.00
I+

1$ 0.00
(D

(0

U
Sl-4

U
.rl

to

a , U
-0.05
10

m a ,

20

-0.05

10

20

(SI

(SI

Fig. 3. (a) Estimate of resistance parameter R provided by fitting ( I ) to the data shown in Fig. 1 by recursive multiple linear regression with a memory time-constant of 1000 s. (b) Recursive estimate of corresponding diagonal element of P matrix. (c) Residuals between measured tracheal pressure and that predicted using the recursively estimated parameters R. E . and K .

Fig. 1. (a) Estimate of resistance parameter R provided by fitting ( I ) to the data shown in Fig. I by recursive multiple linear regression with a memory time-constant of I s. (b) Recursive estimate of corresponding diagonal element of P matrix. (c) Residuals between measured tracheal pressure and that predicted using the recursively estimated parameters R , E , and K .

of inspiration. At this point the estimate of R is based predominately on the preceding end-expiratory data where Ptr, V and V change very slowly and so contain little information. The residuals also decrease as 7 decreases because the model is able to provide a better fit to short segments of data than to longer segments. When 7 is 1.0 s the residuals are smaller than Figs. 2 and 3, but still show clear systematic variations about zero. When 7 is 0.2 s the residuals consist virtually exclusively of cardiogenic oscillations, while when 7 is 0.05 s the residuals are even smaller than the cardiogenic oscillations. In the present respiratory mechanics application we consider the cardiogenic oscillations to be extraneous noise, so would consider an appropriate value for T as

being one that allowed the model to account for all the variation in the data except for the cardiogenic oscillations and any random measurement error. We therefore choose 0.2 s as being the desired value for 7. We should note at this point that the choice of 7 is an important consideration in any application of recursive least squares, as exemplified by two recent studies of its use in respiratory mechanics [9], [lo]. However, in conventional applications where the method is used to follow time-varying parameters, 7 is usually chosen on the basis of its ability to return accurate parameter values from simulated data. In the present study we determined T purely on the basis of the residuals it produces. When T has a value that allows the model to fit the data with the

Y8

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 39. NO

I . JANUARY 1992

e I e

cn
0

a
Y
Y

tu

a
I I ' l l II
I I

- 0

10

-1

10

20

c u 4

lo6

4 0 7

cu

cn

d d
Y

05

0.05

h h

a
Y

a
Y
tn
0.00

(0

3
m
3

0.00

tu
3

U -*

U
-rl

10 al

a
-0.05
10

10 al

-0.05

20

10

20

(SI

t (SI
Fig. 6 . (a) Estimate of resistance parameter R provided by fitting ( I ) to the data shown in Fig. I by recursive multiple linear regression with a memory time-constant of 0.05 s. (b) Recursive estimate of corresponding diagonal element of P matrix. (c) Residuals between measured tracheal pressure and that predicted using the recursively estimated parameters R, E , and K .

Fig. 5 . (a) Estimate of resistance parameter R provided by fitting ( I ) to the data shown in Fig. I by recursive multiple linear regression with a memory time-constant of 0.2 s. (b) Recursive estimate of corresponding diagonal element of P matrix. (c) Residuals between measured tracheal pressure and that predicted using the recursively estimated parameters R , E , and K . required degree of precision, the corresponding recursive estimate of R [Fig. 5(a)] shows how the value of this parameter must vary in order to achieve such a fit. A convenient way to represent the variation of this parameter is to display the values in Fig. 5(a) as a histogram. This has been done as the solid line in Fig. 7, which shows that R varies over a wide range including negative values that are not physically meaningful. Some of the values of R in Fig. 5(a), however, are based on very poor information as in'dicated by the magnitude of the corresponding values of P(1, 1) in Fig. 5(b). Indeed, inspection of Fig. 5 shows that this applies to the physically unreasonable negative values of R. This suggests that we should give greater consideration to those values of R whose P( 1.

1) values are small, while effectively disregarding those values of R whose P ( I , 1) values are very large. We propose, therefore, to represent R in the form of a modified histogram calculated as follows. When assigning each value of R to a particular bin in the histogram, rather than adding 1 to each bin we add 1 divided by the square root of the corresponding P( I , I ) value. (We choose to weight by the square root of P( 1 , 1) since this is proportional to the estimate of a parameter's standard deviation in conventional
multiple weighted to fit the modified linear regression.) The result is thus an informationhistogram of the values of R required to allow the model data satisfactorily. The dashed line in Fig. 7 shows the histogram which now no longer has physically unreason-

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL 39, NO

I . JANUARY 1992

99

C
0)

c r
+
L

a ,

a l

3
U

UI
I

-0.5

0.0

0.5

1 .o

R (kPa.s.1-I)

Fig. 7. Histogram o f R values shown in Fig. 5(a) (solid line) and the corresponding modified histogram calculated as described in the text. tically satisfactory fit is achieved frequently produces an impossible interpretational problem since the final model structure may have no clear analog in the system being modeled. In this paper we have proposed an alternative approach to the parameter confidence problem in which we no longer consider the parameters of a model to be random variables. Instead, they are considered to be deterministic variables whose values vary over the data set in such a way as to allow the model to fit the entire data set satisfactorily. Parameter variability is conveniently determined for a linear model by the recursive least-squares algorithm. The time-constant of the algorithms memory determines the mean magnitude of the residuals between data and fit, and so can be adjusted so that the residuals include only noise. In this way the algorithm effectively has to fit to only a segment of the complete data set at any one time. However, not all segments contain the same amount of information, as is reflected by the elements of the matrix P whose values are estimated recursively along with those of the parameters. Thus, we require a representation of a parameters range of values that is weighted in some way according to the amount of information determining each value. The representation we have chosen is that of the information-weighted histogram. Note that this is a rather different type of weighting to that invoked in conventional weighted least squares in which each data point is weighted according to the confidence one has in its accuracy [ I I ] . In constrast, our type of weighting assigns importance to sets of data points on the basis of how ditferent they are from each other (which is related to how much information they contain). This presentation of a parameter is not based on any statistical assumptions about the data. Indeed, it even avoids the concept of a true parameter value, since it is applicable when the order of the model is insufficient to describe the complete data set all at once. The model parameters thus cannot be conaidered to correspond precisely to physical quantities in the real system, and thus cannot have true values. This is not to say, however. that the model is reduced to the status of being purely empirical. The parameters may still reflect specific aspects of the system to a useful degrce. In the case of the respiratory system model above, the parameters R and E serve as gross indices of the dissipative and elastic processes within the system. The information-weighted histograms of these parameters serve as philosophically satisfactory measures with which to make comparisons between fits of the model to different data sets. Such comparisons may be made quantitatively through measures of dispersion. skewness. etc. of the histograms. These measures could then be compared between different histograms using standard nonparametric statistical tests. On the other hand,

a ,
wL

U
a ,
N
-rt
r(

/I

z
R (kPa . s . 1-l)
Fig. 8. Modified histogram as shown in Fig. 7 (solid line) together with gaussian centered about the value of R obtained by fitting ( I ) using conventional multiple linear regression (dashed line). The standard deviation of the Gaussian is equal to that estimated for R by conventional multiple linear regression. Note that the Gaussian extends considerably above the plot shown since the areas under the solid and dashed lines are equal. able negative values. Indeed, it has the intuitively appealing form of a rather tight distribution function centered about a value not too dissimilar to that obtained by standard multiple linear regression. Fig. 8 shows the modified histogram together with R obtained by regular multiple linear regression. The latter is represented as a Gaussian centered at the estimated value of R and with a standard deviation equal to that estimated for R from (4). T w o observations are noteworthy. First, the modified histogram is centered at a value o f R significantly lower than that returned by multiple linear regression (indeed, as r decreased from 1000 s to 0.05 s the modified histogram moved steadily toward lower R values and became progressively wider). Secondly. the width of the histogram is significantly greater than that of the Gaussian representing the multiple linear regression estimate of R.
111. DISCUSSION

L 0

II

Fitted model parameters are generally treated as random variables whose statistics are inferred from the statistics of the residuals between data and fit. This requires that the residuals be stochastic. The problem with this is that the residuals are often very clearly deterministic. such as occurs when the model does not have sufticient degrees of freedom to account for all the deterministic variation in the data. Increasing the order of the model until a statis-

100

IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING. VOL. 39. NO. I . JANU.4RI

1991

since there is n o stochastic element t o the derivation o f the histograms themselves, it might b e more appropriate t o compare them on the basis of a nonstatistical measure such as an arbitrarily chosen fractional degree of overlap.

[ 5 ] M . S. Ludwig, P. V. Romero, and J . H. T. Bates. A comparison of the dose-response behavior of canine airways and parenchyma. J . Appl. Physiol., vol. 67, pp. 1220-1225, 1989. 161 J. H . T. Bates, K. A . Brown, and T. Kochi, Respiratory mechanic5 in the normal dog determined by expiratoq flow interruption. J . Appl. Phxsiol.. vol. 67. pp. 2276-2285. 1989. [7] M . S . Ludwig. P. V. Romero. P. D. Sly. J . J. Fredberp. and REFERENCES J. H . T. Bates. Interpretation of interrupter resistance after histamine-induced constriction in the doe, J . Appl. PhJsiol., vol. 68. [I] N. R . Draper and H . Smith, Applied Regression h u l y s i s . New pp. 1651-1656. 1990. York: Wiley 1966, ch. 2. [8] T. C . Hsia, System Ideririficafior~. Lexington. MA: D. C. Heath. [2] G . Belforte, B. Bona, and M. Milanese, Advanced modeling and 1977. identification techniques for metabolic proccsses, CRC Crit. RC~I. [9] G . Avanzolini, P. Barbini. A . Cappello, and G . Cevenini. RealBiorned. Eng., vol. 10, pp. 275-316, 1984. time tracking of parameters of lung mechanics: Emphasis on alpo[ 3 ] K . R. Lutchen and A . C. Jackson, Statistical measurcs of parameter rithm tuning, J . Biojned. Eng.. vol. 12. pp. 489-495. 1990. estimates from models fit to respiratory impedance data: Emphasis on [IO] A.-M. Lauzon. and J. H. T. Bates. Estimation of time-var! inp rc5joint variabilities, IEEE Trans. B i o r ? d . 0 1 g . . vol. 33. pp. 1000piratory mechanical parameters by recursive least squares, J . Appl. 1009, 1986. Plzysiol., vol. 71, pp. 1159-1165. 1991. [4] T. Csendes, Nonlinear parameter estimation by global optimiza[ 1 I] K . R . Lutchen, Sensitivity analysis of respiratory parameter uncertion-efficiency and rcliability, Acfu Cyberriet.. vol. 8, pp. 361tainties: impact of criterion function form and constraints. J . Appl. 370, 1988. Physiol., vol. 69, pp. 766-775. 1990.

COMPUTERS,

October 11-14,1992 Durham, North Carolina

Call for Papers

The 19th edition of Computers in Cardiology will return to the United States in 1992 following its successful 1991 meeting in Venice, Italy. Durham, North Carolina, known as the City of Medicine, has been chosen as the 1992 location due to the high concentration of scientists in the Research Triangle area who are working in the fields of cardiology, biomedical engineering, and computer applications. The purpose of this annual international meeting, which is dedicated to the applications of computer technology in clinical cardiology and cardiovascular research, is to facilitate interaction and the exchange of ideas and experiences among clinicians and researchers in cardiology on the one hand and engineers and physical scientists on the other. The meeting will be co-hosted by the Division of Cardiology in the Department of Medicine and the Department of Biomedical Engineering at Duke University. Any contribution pertinent to the scope of the conference is welcome. On Sunday, October 11th, a series of tutorials will focus on cardiac imaging modalities and cardiac electrophysiology. Tours of Duke University laboratories will be available. The scientific sessions of the Conference will start on Monday morning, October 12th. During Monday afternoon and evening, the traditional social program will feature Southern hospitality and bluegrass musicians. Scientific sessions will be held all day on Tuesday and will be followed by a late afternoon poster session and reception. The conference will conclude by early afternoon on Wednesday, October 14th.

FOR MORE INFORMATION, CONTACT:

1992 Computers in Cardiology P.O. Box 2477 Durham, North Carolina 27715 U.S.A. phone: 919-660-5100 fax: 919-660-5405 e-mail: smb@sunbar.mc.duke.edu

Puke

Cosponsored by the IEEE Engineering in Medicine and Biology Society

w-

Local Committee: Susan M. Blanchard, William E. Hammond, Raymond E. Ideker, Mitchell W. Kruchoff, The0 C. Pilkington, William M. Smith, C. Frank Sfarmer

Abstract deadline May 1,1992

IEEE