Bequette 2010 Continuous Glucose Monitoring Real Time Algorithms For Calibration Filtering and Alarms
Bequette 2010 Continuous Glucose Monitoring Real Time Algorithms For Calibration Filtering and Alarms
Abstract
Algorithms for real-time use in continuous glucose monitors are reviewed, including calibration, filtering of
noisy signals, glucose predictions for hypoglycemic and hyperglycemic alarms, compensation for capillary
blood glucose to sensor time lags, and fault detection for sensor degradation and dropouts. A tutorial on
Kalman filtering for real-time estimation, prediction, and lag compensation is presented and demonstrated via
simulation examples. A limited number of fault detection methods for signal degradation and dropout have
been published, making that an important area for future work.
Introduction
Author Affiliation: Department of Chemical and Biological Engineering, Rensselaer Polytechnic Institute, Troy, New York
Abbreviations: (ANN) artificial neural networks, (AR) autoregressive, (CGM) continuous glucose monitoring, (CGMS) continuous glucose
monitoring system, (FD) finite differences, (FDA) Food and Drug Administration, (FIR) finite impulse response, (FSN-CGM) FreeStyle Navigator
continuous glucose monitoring, (IIR) infinite impulse response, (SMBG) self-monitoring blood glucose
Keywords: alarms, algorithms, blood glucose, calibration, filtering, hypoglycemia, optimal estimation, prediction
Corresponding Author: B. Wayne Bequette, Ph.D., Department of Chemical and Biological Engineering, Rensselaer Polytechnic Institute, Troy,
NY 12180-3590; email address [email protected]
404
Continuous Glucose Monitoring: Real-Time Algorithms for Calibration, Filtering, and Alarms Bequette
for the U.S. market was the GlucoWatch G2 Biographer. The slope and intercept are estimated from
An article by the JDRF CGM Study Group7 summarized
the design and methods used to test the three continuous m=
( y2 − y1 ) b = y2 − mx2 . (4)
glucose monitoring systems (CGMS) currently approved ( x2 − x1 )
by the U.S. FDA for the U.S. market.
Further, when multiple data points are available, then
Skyler provided a comprehensive history of CGM
8 linear regression can be used to fit slope and intercept to
development as an editorial introduction to a special issue data. The assumption is that there is uncertainty in the
on CGM. A tutorial overview of current CGM technology output measurement, and the equation has the form15
was provided by Kerr and Fayers,9 while Oliver and
colleagues10 provided a more comprehensive review of yi = mxi + b + ei , (5)
noninvasive glucose sensor technology. Cox11 provided
a head-to-head comparison of the currently available where the subscript i represents the ith data point.
CGM systems. Standard linear regression techniques find the slope (m)
and intercept (b) that minimize the sum of the squares
of the errors (differences between measurements and
A CGM Standards report12 provided consensus guidelines
model predictions, where ŷi = mxi + b) over N data points
on how CGM data should be presented and compared
between CGM devices, and different glucose measurement N N
2
methods. Important topics included the evaluation of min ∑ ei2 = min ∑ ( yi − ŷi ) . (6)
m,b m,b
threshold alarms, calibration, and sensor and physiological i=1 i=1
and only one sensor signal (y) – blood glucose (x) pair is where the parameters m' and b' are found by minimizing
used. the objective function (where x̂i = m' yi + b')
ŷ = a0 yk + a1 yk −1 + + aM yk − M , (11)
The DirecNet Study Group25 evaluated factors affecting k
calibration of the Medtronic CGMS. Sensor accuracy was where y represents the measured glucose value and ŷ
improved slightly with more calibrations per day; also, the filtered value. A simple example is a moving average
accuracy was degraded when calibrating with glucose filter, with all data points in the data window weighted
rates of change ±1.5 mg/dl/min. Another finding was equally. For example, with a window of five samples,
that overnight sensor accuracy was improved if only the following fifth-order filter
overnight calibrations were used; it is suggested that a
separate calibration algorithm be used for overnight ŷk = 0.2yk + 0.2yk −1 + 0.2yk − 2 + 0.2yk − 3 + 0.2yk − 4 (12)
reading. Finally, it should be noted that the calibrations
were retrospective, but the authors suggested that similar would create an effective delay of the order of 2.5 sample
improvements would be expected from prospective times. Panteleon and colleagues19 cited use of a seventh-
calibration. order FIR filter for signals sampled at a 1-minute interval.
Medtronic patents27,28 show raw signals being obtained alarm, whereas the Medtronic Guardian RT and
at 10-second intervals. At the end of each 1-minute Abbott Navigator systems have both a threshold and a
interval, the lowest and highest values are removed, and projected alarm.36
the remaining four are averaged to obtain the 1-minute
average. Similarly, every 5-minute interval removes Linear Regression and Linear-in-Time Projections
the lowest and highest 1-minute averages and averages
Perhaps the most intuitive (and most common) method of
the remaining three data points for the 5-minute average.
predicting glucose values is to perform linear regression
Removing the highest and lowest values has a similar
using a “window” of recent glucose measurements
effect to median filtering. “Clipping limits” can be applied
(filtered, raw, or estimates of glucose) and assuming
to limit the rate of change of either the raw signal
that the slope remains constant into the future. As an
or calibrated glucose values. Keenan and associates29
example, consider the simulated set of glucose sensor
noted that the Medtronic CGMS Gold and Guardian RT
measurements shown in Figure 1, where t = 0 is the current
have different filtering algorithms.
time. Standard linear regression analysis yields not
only an estimate of the slope and intercept parameters
An infinite impulse response (IIR) filter has the form but confidence intervals as well. These confidence intervals
can also be used in the future prediction, yielding a
ŷ = −a1 ŷk −1 − − aN ŷk − N + b0 yk + b1 yk −1 + + bM yk − M , (13)
k region of possible glucose values, as shown 30 minutes
into the future in Figure 1.
where the current filtered value of the output is a function
of N previously filtered values, as well as current and M
Past and future glucose
previous measurements. A patent assigned to Dexcom16
suggests the use of IIR filters for raw signal filtering,
with an example of N = 3 and M = 3. It should be
noted that filters can also be applied to the derivative
of the glucose (or its change between two successive
measurements); this approach is cited in the Medtronic
patent of Steil and Rebrin30 and is particularly useful
G, mg/dl
their paper, they showed that achieving false negative of multiple window lengths of past data for multiple
and false positive ratios of <5 and <10%, respectively, linear regressions. In addition, they considered several
would require a CGM to have an mean absolute relative different prediction horizons. A similar approach is
deviation of ≤7.5%. Brauker38 suggested a “zone” of future proposed in the patent by Dunn and colleagues,45
glucose possibilities, which is the same idea as that shown who suggested a “mixture of experts” based on the
in Figure 1. superposition of multiple linear regressions.
The Abbott FreeStyle Navigator continuous glucose The previous methods require that filtered glucose
monitoring (FSN-CGM) system has both threshold and concentrations be used to develop the models for future
projected alarms. McGarrough and Bergenstal39 presented projections. The theory of optimal estimation and prediction,
in-clinic and in-home results for the FSN-CGM. however, combines signal filtering with model-based
For projected alarms, a window of 15 minutes of sensor estimation and is discussed in the following subsection.
data is used to estimate the rate of change of glucose
and its uncertainty. If the uncertainty is below a certain Optimal Estimation and Prediction Theory
value and the projected glucose is below a threshold of Palerm and colleagues46 developed a Kalman filter-based
85 mg/dl within 10, 20, or 30 minutes, for low, medium, approach to estimate the future values of blood glucose.
and high sensitivity alarms, respectively (as selected by They explored the effect of sample time, hypoglycemic
the patient), an alarm is activated. threshold, and prediction horizon on the sensitivity and
specificity of the predictions and discussed how the
Another linear projection method (used in the voting optimal estimators can be tuned to trade-off the false
method presented by Cameron and colleagues40 and alarm rate with the rate of missed hypoglycemic
Dassau and colleagues41 and the pump shutoff studies episodes; this suggests that an individual can tune the
of Buckingham and associates42) is to consider the end alarm depending on their personal risk for hypoglycemia.
points of a data window Palerm and Bequette47 validated this approach to data
yk − yk − N obtained from hypoglycemic clamp studies. Buckingham36
rate k = (15)
NΔt also recommended that the hypoglycemic alarm threshold
be tunable to meet an individual’s particular needs.
and find the mean squared error of the rate based on
single sample times over the same interval
The underlying model is
. (16) xk +1 = Φxk + Γ w wk
, (19)
yk = Cxk + vk
The standard deviation of the rate estimate is
where x is a vector of states and y is measured output,
MSErate k wk is process noise (covariance Q), and vk is measurement
σ ( rate k ) = . (17) noise (covariance R). If it is assumed that process noise
N −1
drives the first derivative of glucose with time, then the
Further, this can be scaled by the absolute value of the following relationships result
rate of change
g k + 1 = g k + dk
σ ( rate k )
unc k = (18) dk + 1 = dk + w k , (20)
rate k
yk = gk + vk
and if the scaled uncertainty is greater than some
threshold value, the prediction is not accepted. where g and d represent the glucose and the change in
glucose from step to step, respectively. The state space
Choleau and associates43 used a window of 10 minutes, model corresponding to Equation (20) is
with a sensor sample time of 0.5 minute, and linear
regression to estimate the slope (glucose rate of change)
and to project the glucose concentration 20 minutes ⎡ g ⎤ ⎡ 1 1 ⎤⎡ g ⎤ ⎡ 0 ⎤
⎢ ⎥ = ⎢ ⎥⎢ ⎥ + ⎢ ⎥w
into the future. If the projected glucose is 70 mg/dl or less, ⎢⎣ d ⎥⎦ k +1 0 1 ⎦ ⎢⎣ d ⎥⎦
⎣ ⎣ 1 ⎦ k
the alarm is triggered. Experimental data are presented k
xk +1 Φ xk Γw
based on rat studies.
⎡ g ⎤
Cameron and colleagues44 developed a similar approach
yk = ⎡⎣ 1 0 ⎤⎦ ⎢ ⎥ + vk . (21)
⎢⎣ d ⎥⎦
in their development of a statistical-based hypoglycemic C
k
alarm. Their technique, however, involves the consideration xk
Process and measurement noise are considered stochastic Pk = ΦPk −1 ΦT + Γ wQΓ wT . (28)
processes, and the process noise covariance (Q) is known
only approximately and is often used as a tuning parameter. For this problem, the confidence interval grows with
The states are estimated using predictor–corrector equations each step not followed by a measurement update as
of the form: there are two “integrators” in the glucose model; this
behavior is demonstrated in the simulations that follow.
x̂k |k −1 = Φx̂k −1|k −1 predictor (time update) (22)
and the glucose sensor measurement is used to update Consider the simulated noise sequence presented in
the state estimate Figure 1 where Kalman filtering is used to obtain
glucose and rate of change of glucose, in real time,
(
x̂k |k = x̂k |k −1 + Lk yk − Cx̂k |k −1 ) despite substantial measurement noise. The state vector
estimate is initialized with the measured glucose at
corrector (measurement update) (23) t = –15 minutes, with an assumed rate of change of
0 mg/dl/min. Figure 2 displays real-time estimates of
where x̂ represents an estimate of the states and the
glucose (top) and its rate of change (bottom) along with
subscript k|k – 1 means the estimate at step k is based
the uncertainty bounds of each. Although the actual
on measurements up (and including) step k – 1. Note that
rate of change was –2 mg/dl/min, the Kalman filter
a model is used to propagate the state estimate from the
converges to this value within 15 minutes. Figure 3
previous time step (k – 1) to the current time step (k).
shows predicted glucose values based on measurements
The measurement at the current time step is then used
available until t = 0 minute. Because of the uncertainty
to update the state estimate based on Kalman gain (Lk).
about future rates of change, confidence intervals of the
glucose predictions increase each step into the future;
The state estimate covariance is determined by solving simulation details are provided in the Appendix.
The increase in uncertainty bounds in Figure 3
( )
−1
Pk = ΦPk −1 ΦT + Γ wQΓ wT − ΦPk −1CT CPk −1CT + R CPk −1 ΦT , (24) (based on Kalman filtering), compared to Figure 1
(based on linear regression analysis), is an artifact of
where the first two terms on the right-hand side of the
assumptions used in the simulation. The smaller linear
equals sign represent the propagation of state covariance
regression confidence bounds are due to the fact that the
and process noise, while the third term represents a
real underlying signal has a constant slope, while the
correction due to the measurement update (including the
Kalman filter “recognizes” that the slope may change in
effect of measurement noise)
the future.
( )
−1
Lk = Pk CT CPk CT + R . (25)
⎣ ⎦
yk = gk + vk , (29) yk = ∑ ai yk − i + wk , (32)
i=1
which has the state space form where y is the glucose value, w is a white noise sequence,
and there are n coefficients, ai (i.e., the model is nth order).
⎡ x ⎤ ⎡ 2 −1 ⎤ ⎡ x1 ⎤ An autoregressive moving average model has a more
⎥ + ⎢ 1 ⎥ wk
⎡ ⎤
⎢ 1 ⎥ = ⎢ ⎥⎢
x
⎢⎣ 2 ⎥⎦ 1 0 x 0 general form with n ai coefficients and m ci coefficients
⎣ ⎦ ⎢ 2 ⎥⎦ k ⎣⎦
⎣
k+1 n m−1
xk +1
Φ
xk Γw
yk = ∑ ai yk − i + ∑ ci wk − i . (33)
⎡ x ⎤ i=1 i=0
1
yk = ⎡⎣ 1 0 ⎤⎦ ⎢ ⎥ + vk , (30)
⎢ x2 ⎥ Reifman and colleagues51 studied a sensor with a sample
C ⎣ ⎦k
time of 1 minute and used 2000 data points (2000 minutes)
xk
to “train” a 10th-order AR model (n = 10); that is, they fit
where the first state is the glucose and the second the 10 model parameters to data. The performance of this
state is the value of glucose at the previous time step. model is based on the ability to match 4000 additional
Although the performance of the resulting Kalman filter minutes of data. They found that a prediction horizon of
is equivalent to results shown in Figures 2 and 3, we 30 minutes yields a root mean square error of 22 mg/dl,
feel that our formulation is more natural, as our second with a prediction delay. In order to reduce the prediction
state is the rate of change of glucose, which is applied delay, Gani and colleagues52 used a smoothing procedure
directly to future predictions. Facchinetti and associates49 and regularization to minimize changes in the glucose
developed a nice procedure for tuning the Kalman filter first derivatives. This approach, with a 30th-order model,
based on data from individual subjects. reduces the prediction lag compared to the results of
Reifman and associates.51
The Kalman filter estimates shown in Figures 2 and 3
are based on a second-order (two-state) model. However, Sparacino and colleagues53 used an adaptive first-order
if it is assumed that process noise drives the second model, where the parameter is updated at each time
derivative of glucose with time, the following third-order step to estimate future glucose values and compared
(three-state) model can be used: this with a first-order “polynomial-in-time” approach.
Zanderigo and associates assessed the accuracy using prediction alarms, however. A number of different
a continuous glucose error-grid analysis. Sparacino and structures and training procedures can be used for
colleagues provided a tutorial overview of algorithms for artificial neural networks. Kuure-Kinsey and associates60
continuous glucose monitoring and glucose prediction presented a process application of a feed-forward ANN,
for use in hypo/hyperglycemic alarms. They noted that while Kuure-Kinsey and Bequette61 showed that a
a limitation to the approach of Reifman and colleagues51 recurrent ANN yields better future predictions. A major
is the substantial amount of training data required for disadvantage to nonlinear techniques, such as ANN,
estimation of a large number of parameters in a fixed- is that much training data are needed (for model
parameter formulation. Gani and colleagues addressed parameter estimation).
this criticism by showing that a model could be produced
once for a particular individual and then used on other Discussion
individuals.
Gani and colleagues52 stated that a disadvantage to
Kalman filtering is that it requires a high-fidelity, first-
Eren-Oruklu and associates57,58 developed a more general principles model for meals and physical activity, but it
approach, including real-time adaptation of parameters, should be clear from Equations (20), (21), and (31) that
and found that the best model is second order, with the this is not the case. The two- and three-state models
form of Equations (21) and (31) are nothing more than basic
principles of differential calculus applied to glucose; that
yk = a1 yk −1 + a2 yk −1 + wk + c1 wk −1 , (34)
is, the two-state model related “distance” (glucose) and
“velocity” (rate of change) of glucose, whereas the three-
while Sparacino and colleagues53 used a first-order model
state model relates distance, velocity and “acceleration”
yk = a1 yk −1 + wk . (35) (second derivative of glucose). These models are exact,
with the only uncertainties being the distribution of
In these studies, model coefficients are estimated recursively, process noise (perturbations to either the first or second
using weighted least squares, where data further in derivatives of glucose) and sensor noise. A major advantage
the past have less of an impact than more recent to the Kalman filtering approach is that it, unlike
data. That is, the parameters are found to solve the empirical modeling, does not require much data for
optimization problem parameter estimation. Depending on what is known,
N −1
or assumed, about the process and sensor noise variances,
2 a single tuning parameter, the Q/R ratio, can be used to
min ∑ m i ( yk − i − ŷk − i ) , (36)
change Kalman filter estimator performance.
i=0
Artificial Neural Networks (ANN) Keenan and colleagues29 provided an overview of delays
Artificial neural networks and other nonlinear models can in continuous glucose sensor devices. For common
be used for glucose prediction. Pappada and colleagues59 transcutaneous sensors on the market, the physiologic
used an ANN, based on CGM and additional patient diary lag is 3–12 minutes, the electrochemical sensor lag is
information (meals and insulin infusion), to predict 1–2 minutes, and there can be additional lags due to filtering
glucose values; they do not propose hypoglycemic algorithms. In an in vitro study of the Medtronic CGMS,
Keenan and associates29 found that a substantial time Glucose concentration, blood vs. s.c. measured
delay can be induced by the sensor noise filtering
algorithm under conditions of glucose rate of change
that are not physiologic.
G, mg/dl
fluid as a first-order differential equation with the form
dy 1 k
= ay + bg = − y + g, (37)
dt t t
Consider now the simple simulation presented by compensation in their products; it is used here merely to
Rebrin and associates65 for a capillary blood glucose show the major advantages of optimal estimation-based
decrease from 200 to 100 mg/dl, where a time constant techniques for this type of problem. The Medtronic patent
(lag) of 12 minutes and a gain of 1 are assumed; the by Steil and Rebrin30 cites the use of a Weiner filter for
sensor sample time is 1 minute. The corresponding lag compensation, but no technical details are provided.
interstitial fluid glucose, with sensor noise (standard The Abbott patent by Feldman and McGarraugh17 describes
deviation = 1 mg/dl), is shown in Figure 4. While the a lag compensation procedure identical to Equation (39),
measurement noise in Figure 4 does not appear too when y is a calibrated sensor value (and thus a/b = –1).
significant, it does have a strong effect on the ability to
estimate blood glucose from subcutaneous measurements, Bequette68 developed a Kalman filter-based algorithm
as demonstrated in the simulations that follow. to compensate for the blood glucose to interstitial fluid
transport lag, even when there is substantial sensor noise.
Rebrin and colleagues65 presented an intuitive estimation A formulation that assumes that the stochastic noise
approach by rearranging Equation (37) to solve for blood term perturbs the glucose rate of change is shown in the
glucose (g) from the subcutaneous sensor signal (y): following equation
dy ⎡ s ⎤ ⎡ ⎤⎡ s ⎤ ⎡
⎥ ⎢ Φ Γ 0 ⎥⎢ k ⎥ ⎢ 0 ⎥
⎤
− ay ⎢ k +1
ĝ = dt , (38) ⎢ gk +1⎥ = ⎢ 0 1 1 ⎥ ⎢ gk ⎥ + ⎢ 0 ⎥ wk
b ⎢ ⎥ ⎢ 0 0 1 ⎥⎢ ⎥ ⎢ 1 ⎥
⎢⎣ dk +1
⎥⎦ ⎣ ⎦ ⎢⎣ dk ⎥⎦ ⎣ ⎦
where ^ is used to indicate an estimated value. A finite-
⎡ s ⎤
differences (FD) approximation for the derivative yields ⎢ ⎥
yk = ⎣ 1 0 0 ⎦ ⎢ gk ⎥ + vk
⎡ ⎤ , (40)
⎢ d ⎥
⎣ k ⎦
where s is the subcutaneous glucose concentration, g is
. (39) the capillary blood glucose, and d is the rate of change
of blood glucose. The time lag of 12 minutes and the
gain of 1 result in values of Φ = exp ( −Δt t ) = 0.92 and
Rebrin and colleagues65 noted that this numerical Γ = 1 − Φ = 0.08. The state estimates are obtained solving
derivative-based approach is very sensitive to measurement Equations (22) and (23) as before. For these simulation
noise and also applied a three-point moving average studies, the steady-state Kalman gain is obtained by
filter to the derivative term. It should be made clear iterating Equations (24) and (25) until they converge to
that Medtronic does not use this FD approach for lag constant values.
A comparison of FD with Kalman filter estimates is blood glucose and the subcutaneous sensor signal, which
shown in Figure 5, where the Kalman filter is designed may actually be time varying. Another limitation is the
based on Q = 0.005 and R = 1 and on the steady-state static calibration curve relating the subcutaneous signal to
Kalman gain (Lss). Clearly the Kalman filter estimates the capillary blood glucose; this is strongly dependent
are insensitive to sensor noise, with the downside of on the quality of the calibration, which is dependent on the
a slight lag in the estimates when glucose begins to glucose meter variability as well as the “human element.”
decrease. This lag can be reduced by using a larger
Q value in the design, with a slight increase in noise
sensitivity.
Signal Dropouts and Related Artifacts
A simple way to detect signal dropouts is to set a bound
on allowable measurement changes. That is, a measurement
is invalid if
yk − yk −1 > d
G, mg/dl
, (42)
yk < ε
⎢⎣ d̂ ⎥⎦
k|k
1. Bequette BW. A critical assessment of algorithms and challenges in 21. Choleau C, Klein JC, Reach G, Aussedat B, Demaria-Pesce V,
the development of an artificial pancreas. Diabetes Technol Ther. Wilson GS, Gifford R, Ward WK. Calibration of a subcutaneous
2005;7(1):28-47. amperometric glucose sensor implanted for 7 days in diabetic
patients. Part 2. Superiority of the one-point calibration method.
2. Doyle FJ 3rd, Jovanovic L, Seborg DE. A tutorial on biomedical Biosens Bioelectron. 2002;17(8):647-54.
process control: glucose control strategies for treating type 1
diabetes mellitus. J Process Control. 2007;17(7):572-6. 22. Clark WL, Cox D, Gonder-Frederick LA, Carter W, Pohl SL.
Evaluating clinical accuracy of systems for self-monitoring of
3. Kumareswaran K, Evans ML, Hovorka R. Artificial pancreas: an blood glucose. Diabetes Care. 1987;10(5):622-8.
emerging approach to treat type 1 diabetes. Expert Rev Med Dev.
2009;6(4):401-10. 23. Aussedat B, Thome´-Duret V, Reach G, Lemmonier F, Klein JC,
Hu Y, Wilson GS. A user-friendly method for calibrating a
4. Cobelli C, Dalla Man C, Sparacino G, Magni L, De Nicolao G, subcutaneous glucose sensor-based hypoglycaemic alarm. Biosens
Kovatchev B. Diabetes: models, signals and control. IEEE Rev Bioelectron. 1997;12(11):1061-71.
Biomed Eng. 2010;3. In press.
24. King C, Anderson SM, Breton M, Clarke WJ, Kovatchev B.
5. Klonoff DC. Continuous glucose monitoring: roadmap for 21st Modeling of calibration effectiveness and blood-to-interstitial
century diabetes therapy. Diabetes Care. 2005;28(5):1231-9. glucose dynamics as potential confounders of the accuracy of
6. Klonoff DC. A review of continuous glucose monitoring technology. continuous glucose sensors during hyperinsulinemic clamp.
Diabetes Technol Ther. 2005;7(5):770-775 J Diabetes Sci Technol. 2007;1(3):317-22.
7. JDRF CGM Study Group. JDRF randomized clinical trial to assess 25. DirectNet Study Group. Evaluation of factors affecting CGMS
the efficacy of real-time continuous glucose monitoring in the calibration. Diabetes Technol Ther. 2006;8(1):318-25.
management of type 1 diabetes: research design and methods. 26. Poitout V, Moatti-Sirat D, Reach G, Zhang Y, Wilson GS,
Diabetes Technol Ther. 2008;10(4): 310-21. Lemonnier F, Klein JC. A glucose monitoring system for on
8. Skyler JS. Continuous glucose monitoring: an overview of its line estimation in man of blood glucose concentration using a
development. Diabetes Technol Ther. 2009;11 Suppl 1:S5-10. miniaturized glucose sensor implanted in the subcutaneous tissue
and a wearable control unit. Diabetologia. 1993;36(7):658-63.
9. Kerr D, Fayers K. Continuous real-time glucose monitoring systems:
time for a closer look. Pract Diab Int. 2008;25(1):37-41. 27. Mastrototaro JJ, Gross TM, Shin JJ. Glucose monitor calibration
methods. United States patent US 6,424,847. 2002 Jul 23.
10. Oliver NS, Toumazou C, Cass AE, Johnston DG. Glucose sensors:
a review of current and emerging technology. Diab Med. 28. Shin JJ, Holtzclaw KR, Dangui ND, Kanderian S Jr,
2009;26(3):197-210. Mastrototaro JJ, Hong PI. Real time self-adjusting calibration
algorithm. United States patent US 6,895,263 B2. 2005 May 17.
11. Cox M. An overview of continuous glucose monitoring systems.
J Pediatr Health Care. 2009;23(5):344-7. 29. Keenan DB, Mastrototaro JJ, Voskanyan G, Steil GM. Delays
in minimally invasive continuous glucose monitoring devices:
12. Klonoff D, Bernhardt P, Ginsberg BH, Joseph J, Mastrototaro J,
a review of current technology. J Diabetes Sci Technol.
Parker DR, Vesper H, Vigersky R. Performance metrics for
2009;3(5):1207-14
continuous interstitial glucose monitoring; approved guideline.
CLSI document POCT05-A (ISBN 1-56238-685-9). Clinical and 30. Steil G, Rebrin K. Closed-loop system for controlling insulin
Laboratory Standards Institute; 2008. infusion. United States patent US 7,354,420 B2. 2008 Apr 8.
31. Knobbe EJ, Buckingham B. The extended Kalman filter for 51. Reifman J, Rajaraman S, Gribok A, Ward WK. Predictive
continuous glucose monitoring. Diabetes Technol Ther. 2005;7(1):15-27. monitoring for improved management of glucose levels. J Diabetes
Sci Technol. 2007;1(4):478-86.
32. Knobbe EJ, Lim WL, Buckingham BA. Method and apparatus for
real-time estimation of physiological parameters. United States 52. Gani A, Gribok AV, Rajaraman S, Ward WK, Reifman J. Predicting
patent US 6,575,905 B2. 2003 Jun 10. subcutaneous glucose concentration in humans: data-driven glucose
modeling. IEEE Trans Biomed Eng. 2009;56(2):246-54.
33. Bequette BW, Palerm CC, Willis JP, Desemone J. Incorporation of
glucose meter uncertainty in the calibration of continuous glucose 53. Sparacino G, Zanderigo F, Corazza S, Maran A, Facchinetti A,
sensors. Diabetes Technol Ther. 2005;7(2):366-7. Cobelli C. Glucose concentration can be predicted ahead in time from
continuous glucose monitoring sensor time series. IEEE Trans
34. Kuure-Kinsey M, Palerm CC, Bequette BW. A dual-rate Kalman filter
Biomed Eng. 2007;54(5):931-7.
for continuous glucose monitoring. Proc. IEEE EMB Conference;
2006. p. 63-6. 54. Zanderigo F, Sparacino G, Kovatchev BP, Cobelli C. Glucose
prediction algorithms from continuous monitoring data: assessment
35. Heise T, Koschinsky T, Heinemann, L, Lodwig V. Hypoglycemia
of accuracy via continuous glucose-error grid analysis. J Diabetes
warning signal and glucose sensors: requirements and concepts.
Sci Technol. 2007;1(5):645-51.
Diabetes Technol Ther. 2003;5:563-71.
55. Sparacino G, Facchinetti A, Maran A, Cobelli. Continuous glucose
36. Buckingham B. Clinical overview of continuous glucose monitoring.
monitoring time series and hypo/hyperglycemia prevention:
J Diabetes Sci Technol. 2008;2(2):300-6.
requirements, methods, open problems. Curr Diabetes Rev.
37. Noujaim SE, Horwitz D, Sharma M, Marhoul J. Accuracy 2008;4(3):181-92.
requirements for a hypoglycemia detector: an analytical model
56. Gani A, Gribok AV, Lu Y, Ward WK, Vigersky RA, Reifman J.
to evaluate the effects of bias, precision, and the rate of glucose
Universal glucose models for predicting subcutaneous glucose
change. J Diabetes Sci Technol. 2007;1(5):652-68.
concentration in humans. IEEE Trans Inf Tech Biomed. 2010;14(1):157-65.
38. Brauker J. Continuous glucose sensing: future technology
57. Eren-Oruklu M, Cinar A, Quinn L, Smith D. Estimation of future
developments. Diabetes Technol Ther. 2009;11 Suppl 1:S25-36.
glucose concentrations with subject-specific recursive linear
39. McGarraugh G, Bergenstal R. Detection of hypoglycemia with models. Diabetes Technol Ther. 2009;11(4):243-53.
continuous interstitial and traditional blood glucose monitoring
58. Eren-Oruklu M, Cinar A, Quinn L. Hypoglycemia prediction
using the Freestyle Navigator continuous glucose monitoring
with subject-specific recursive time-series models. J Diabetes Sci
system. Diabetes Technol Ther. 2009;11(3):145-50.
Technol. 2010;4(1):25-33.
40. Cameron FM, Niemeyer G, Palerm CC, Dassau E, Doyle FJ 3rd,
59. Pappada SM, Camero BD, Rosman PM. Development of a neural
Lee H, Bequette BW, Chase HP, Buckingham BA. Early detection of
network for prediction of glucose concentration in type 1 diabetes
hypoglycemia combining multiple predictive methods on retrospective
patients. J Diabetes Sci Technol. 2008;2(5):792-801.
clinical continuous glucose monitoring data. J Diabetes Sci Technol.
2008;2(2):A19. 60. Kuure-Kinsey M, Cutright R, Bequette BW. Computationally
efficient neural predictive control based on a feedforward
41. Dassau E, Cameron FM, Lee H, Bequette BW, Doyle FJ 3rd,
architecture. Ind Eng Chem Res. 2006;45(25):8575-82.
Niemeyer G, Chase P, Buckingham BA. Real-time hypoglycemia
prediction using continuous glucose monitoring (CGM), a safety 61. Kuure-Kinsey M, Bequette BW. Improved nonlinear predictive
net to the artificial pancreas. Diabetes. 2008;57:A13, Suppl. control performance using recurrent neural networks. Proceedings
of the 2008 American Control Conference; Seattle, Washington.
42. Buckingham B, Cobry E, Clinton P, Gage V, Caswell K,
p. 4197-202.
Kunselman E, Cameron F, Chase HP. Preventing hypoglycemia
using predictive alarm algorithms and insulin pump suspension. 62. Weinstein RL, Schwartz S, Brazg RL, Bugler JR, Peyser TA,
Diabetes Technol Ther. 2009;11(2):93-7. McGarraugh GV. Accuracy of the 5-day FreeStyle Navigator
continuous glucose monitoring system. Diabetes Care.
43. Choleau C, Dokladal P, Klein JC, Ward WK, Wilson GS, Reach G.
2007;30(3):1125-30.
Prevention of hypoglycemia using risk assessment with a continuous
glucose monitoring system. Diabetes Care. 2002;51(11):3263-73. 63. Voskanyan G, Keenan DB, Mastrototaro JJ, Steil GM. Putative
delays in interstitial fluid (ISF) glucose kinetics can be attributed
44. Cameron F, Niemayer G, Gundy-Burlet K, Buckingham B. Statistical
to the glucose sensing systems used to measure them rather
hypoglycemia prediction. J Diabetes Sci Technol. 2008;2(4):612-21.
than the delay in ISF glucose itself. J Diabetes Sci Technol.
45. Dunn TC, Jayalakshmi Y, Kurnik RT, Lesho MJ, Oliver JJ, 2007;1(5):639-44.
Potts RO, Tamada JA, Waterhouse SR, Wei CW. Method and device
64. Schmidtke DW, Freeland AC, Heller A, Bonnecaze R. Measurement
for predicting physiological values. United States patent US 6,653,091
and modeling of the transient difference between blood and
B1. 2003 Nov 25.
subcutaneous glucose concentrations in the rat after injection of
46. Palerm CC, Willis JP, Desemone J, Bequette BW. Hypoglycemia insulin. Proc Natl Acad Sci U S A. 1998;95(1):294-9.
prediction and detection using optimal estimation. Diabetes
Technol Ther. 2005;7(1):3-14. 65. Rebrin K, Steil GM, Van Antwerp WP, Mastrototaro JJ.
Subcutaneous glucose predicts plasma glucose independent of
47. Palerm CC, Bequette BW. Hypoglycemia detection and prediction insulin: implications for continuous monitoring. Am J Physiol.
using continuous glucose monitoring--a study on hypoglycemic 1999;277(3 Pt 1):E561-71.
clamp data. J Diabetes Sci Technol. 2007;1(5):624-9.
66. Steil GM, Rebrin K, Mastrototaro J, Bernaba B, Saad MF.
48. Stengel RF. Optimal control and estimation. New York: Dover; 1994. Determination of plasma glucose during rapid glucose excursions
49. Facchinetti A, Sparacino G, Cobelli C. An on-line self-tuneable with a subcutaneous glucose sensor. Diabetes Technol Ther.
method to denoise CGM sensor data. IEEE Trans Biomed Eng. 2003;5(1):27-31.
2010;57. In press. 67. Kulcu E, Tamada JA, Reach G, Potts RO, Lesho MJ. Physiological
50. Bremer T, Gough DA. Is blood glucose predictable from previous differences between interstitial glucose and blood glucose
values? A solicitation for data. Diabetes. 1999;48(3):445-51. measured in human subjects. Diabetes Care. 2003;26(8):2405-9.
Q = 0.12 mg2/dl2/min2
R = 22 mg2/dl2 .
The following initial conditions are used for the state vector and the state covariance
⎡ 22 0 ⎤
P0 = ⎢ ⎥
⎢⎣ 0 2 2 ⎥⎦
Thus there is significant initial uncertainty in the states, as the actual initial condition is
⎡ g ⎤ ⎡
⎥ = ⎢ 150 ⎥ .
0 ⎤
x0 = ⎢
⎢⎣ do ⎥⎦ ⎣ −2 ⎦
⎡ 0.2716 ⎤
Lss = ⎢ ⎥.
⎣ 0.0427 ⎦
⎡ 4.6415 0.3770 ⎤
Pk =15 = ⎢ ⎥
⎣ 0.3770 0.1025 ⎦
⎡ 1.4914 0.2343 ⎤
Pss = ⎢ ⎥.
⎣ 0.2343 0.0736 ⎦
That is, eventually (with continued measurement updates), variances of the glucose and rate-of-changes estimates
are 1.49 and 0.0736 mg2/dl2/min2, respectively, yielding long-term confidence intervals of ±1.22 mg/dl and ±0.27 mg/dl/min
for glucose and its rate of change.
A3. Estimation of Capillary Blood Glucose from Noisy Subcutaneous Sensor Example
Simulation details for Figures 4 and 5 are presented here. The process covariance, Q = 0.005, and the sensor noise,
R = 1 (consistent with the measurement noise variance of 1 mg2/dl2). The steady-state Kalman gain is
⎡ 0.2522 ⎤
Lss = ⎢ 0.7221 ⎥
⎢ ⎥
⎢⎣ 0.0611 ⎥⎦
That is, the steady-state variance of the estimate of subcutaneous glucose (measured) is 0.3373 mg2/dl2, capillary blood
glucose (unmeasured) is 3.4424 mg2/dl2, and the rate of change of blood glucose (unmeasured) is 0.064 mg2/dl2/min2.