DIABETES TECHNOLOGY & THERAPEUTICS

Volume 12, Number 5, 2010

ª Mary Ann Liebert, Inc.
DOI: 10.1089=dia.2009.0158

Enhanced Accuracy of Continuous Glucose Monitoring

by Online Extended Kalman Filtering

Andrea Facchinetti, Ph.D., Giovanni Sparacino, Ph.D., and Claudio Cobelli, Ph.D.

Abstract
Background: Most continuous glucose monitoring (CGM) devices measure a current, proportional to the in-
terstitial glucose (IG) concentration, which is converted into a glucose level by a standard device calibration step
that exploits some blood glucose (BG) references. However, data show that deterioration of sensor gain may
occur, which can affect CGM output by a systematic and possibly large (e.g., up to 15=20 mg=dL) error. En-
hanced calibration algorithms for improving the accuracy of CGM are thus of critical importance, especially in
real-time applications.
Methods: In this work we present an enhanced Bayesian calibration method that can be implemented online by
using the Extended Kalman Filter. The method takes into account the existence of BG-to-IG kinetics by incor-
porating a population convolution model and exploits only four BG reference samples per day.
Results: The new method is successfully applied on 10 simulated virtual patients. Its performance in improving
the accuracy of CGM profiles is significantly better than that of other current calibration procedures. Further-
more, the new method is shown to be robust to changes in its parameters. Improvement in the accuracy of CGM
is also shown on a representative subject.
Conclusions: Realistic simulations show that the new enhanced calibration method significantly improves the
accuracy of CGM signals, suggesting potential benefits by its inclusion in real-time applications of CGM devices.

Introduction lyzer [YSI, Inc., Yellow Springs, OH]), and a CGM pro-
file (line), collected using a subcutaneous sensor (FreeStyle

C ontinuous glucose monitoring (CGM) devices can

measure, in real-time, the glycemic level at a continuous
time for up to several days,1,2 allowing the improvement of
diabetes management.3–5 Most commercial CGM systems,
e.g., the Seven
(DexCom, Inc., San Diego, CA),6 the FreeStyle
Navigator
(Abbott Diabetes Care, Alameda, CA),7 and the
CGMS
(Medtronic Minimed, Inc., Northridge, CA),8 exploit
the glucose oxidase principle, which requires that the current
levels (measured, e.g., in mA) be transformed into glucose
levels (given, e.g., in mg=dL) by exploiting a transformation
rule with parameters determined using one or more glucose
reference, e.g., self-monitoring of blood glucose (SMBG),
samples. This step is commonly referred to as standard device
calibration.
In order to reduce invasiveness, CGM sensors are placed in
the subcutis and thus measure interstitial glucose (IG) rather
than blood glucose (BG) concentration. In dynamic condi-
tions, IG and BG differ because of the existence of BG-to-IG
kinetics. To show this, Figure 1 displays a comparison, per-
formed in a clinical study,9 between BG reference samples
(stars), collected every 15 min (measured with a YSI BG ana-
Navigator); amplitude and phase distortions are evident. The
BG-to-IG kinetics have been described by a two-compartment
model (Fig. 2),10 from which one can easily obtain:
: 1 g
IG(t) ¼
IG(t) þ BG(t) (1)
s s
In other terms, IG can be interpreted as the output of a first-
order linear system driven by BG, where g represents the
‘‘static gain’’ of the ‘‘BG-to-IG system,’’ which we can consider
equal to 1,11 and t is the time constant of the system, which
could vary between individuals. The system acts as a low-pass
filter and introduces a distortion (attenuation in amplitude
and distortion in phase) that is well observable, e.g., in time
window 0–5 h of Figure 1. The existence of BG-to-IG kinetics,
however, is not able to explain some discrepancies that are
evident along the y-axis, e.g., in the interval 12–22 h. This dif-
ference seems due to a change of behavior of the CGM sensor
performance after the initial standard device calibration.
The fact that CGM profiles may differ from BG pro-
files because of calibration problems can be critical in several

Department of Information Engineering, University of Padova, Padova, Italy.

353
354
400
BG references
350 CGM data
300
250
mg/dL
200
150
100
50
0 5 10
Time (hours)
FIG. 1. Representative real data taken from Kovatchev et al.9: BG reference samples (stars) versus CGM data (line) profiles.
applications, e.g., artificial pancreas methods relying on CGM
output.12,13 Improving the accuracy of CGM data is therefore
an important task to deal with in real-time. In this work we
present a new enhanced calibration method that can poten-
tially work in cascade to standard calibration of any CGM
device. The aim of this enhanced calibration is to process the
sensor outcome in order to improve the quality of the CGM
readings. The method is based on the Extended Kalman Filter
(EKF), which, by taking into account the BG-to-IG kinetics,
four SMBG samples per day, and a model of the time–
behavior of sensor accuracy, significantly enhances the qual-
ity of CGM data in real-time applications. The new method is
successfully validated on a simulated dataset reproducing
10 virtual patients. In particular, the method is shown supe-
rior in terms of both root mean square error (RMSE) and ro-
bustness with respect to other current calibration procedures
proposed in the literature. In addition, a representative subject
dataset is also used show the practical usability of the method.
CGM Calibration: State of the Art
Different calibration procedures have been presented in the
literature. A preliminary and important work that considered
the efficiency of calibration was performed by the DirecNet
FIG. 2. Two-compartment model describing the BG-to-IG
kinetics. In order to obtain the formulation of Eq. 1, the fol-
lowing reparameterization has been used: g ¼ (k21V1=V2)t
and t ¼ 1=(k02 þ k12).
FACCHINETTI ET AL.
15 20 25
Study Group,14 which retrospectively analyzed the im-
provement in CGMS sensor accuracy by modifying the
number and timing of the calibration points. Results showed
that the timing of the calibration points is even more impor-
tant than the number. In particular, performing calibrations
during periods of relative glucose stability, i.e., where the
point-to-point difference due to the BG-to-IG kinetics is min-
imized, significantly improves the accuracy.
The most adopted calibration strategy15,16 is the one pre-
sented by King et al.,17 which is based on the linear regression
model of Eq. 2:
y ¼ ax þ b (2)
where a and b are calibration parameters that are determined
by fitting them against a couple of BG (y) and CGM (x) pairs
collected at the same time.17,18 The procedure works retro-
spectively and exploits all available BG references. This
method suffers by several limitations. First, it can be applied
only retrospectively. Second, it does not take into account the
distortion introduced by the BG-to-IG kinetics. For instance, it
does not consider that, when glucose is changing rapidly, e.g.,
after a meal, BG and IG are significantly different (e.g., BG can
be 20 mg=dL higher than IG). Finally, it cannot deal with a
possible time-varying behavior of sensor performance (i.e.,
the parameters of Eq. 2 are estimated only once for the entire
monitoring).
Another approach is presented in the work of Kuure-
Kinsey et al.,19 where a dual-rate Kalman filter is presented to
improve the accuracy of CGM data. The procedure exploits
sparse SMBG measurements and estimates in real-time the
sensor gain. However, this algorithm also does not account
for BG-to-IG kinetics.
A comprehensive description of the CGM measurement
process is due to Knobbe and Buckingham.11 In their work,
the BG-to-IG kinetics model was explicitly taken into ac-
count in order to reconstruct BG levels at continuous time
from CGM measurements. In such an attempt, a state-space
Bayesian framework exploiting a priori knowledge of the un-
known variables was adopted (for details see The state space
ENHANCED ACCURACY OF CGM
model), and an EKF was implemented to perform state esti-
mation. However, in the article of Knobbe and Buckingham11
the EKF was tested on four subjects principally to estimate the
BG concentration. The calibration of CGM was only a by-
product of the method. Even if results were insightful, the
robustness of the method was not fully assessed because
validation on simulated data was missing.
Below we start from the work of Knobbe and Bucking-
ham11 and further develop the approach in order to tackle
calibration issues. In particular, our method aims to improve
the accuracy of CGM data by performing an enhanced cali-
bration, which works in cascade to the standard device cali-
bration. Also, we design and implement a simulation strategy
that is able to precisely define the domain of validity of the
method.
Simulated Database
The database used for the validation of the method consists
of 10 simulated type 1 diabetes virtual patients. Each subject
file contains a 7-day simulated CGM profile and four simu-
lated SMBG samples per day, created as described below.
First, BG profiles were created at 1-min time sampling by
using an in silico type 1 diabetes simulator. Model parameters
of the 10 virtual patients were randomly chosen from realistic
distributions.13,20,21 In this way, each 7-day BG time series has
its own characteristics and differs from all the others. More
specifically, to build the BG time-series, three meals per day
were considered. In order to simulate real-life episodes,
breakfast, lunch, and dinner times were allocated randomly
within the following time intervals: 6:00–8:00 h for breakfast,
12:00–14:00 h for lunch, and 19:00–21:00 h for dinner. Carbo-
hydrate intakes were differentiated from meal to meal and
from day to day. Their quantities were drawn randomly from
three different normal distributions with mean  SD given by
45  10 g (breakfast), 75  10 g (lunch), and 85  10 g (din-
ner).20 For instance, Figure 3 displays the BG profile (black
thick line) obtained for a representative subject (subject
220 BG
IG
200 CGM
SMBG
glucose (mg/dL)
180
160
140
120
100
0 6
Time (hours)
FIG. 3. Representative simulated virtual patient number 3, Day 6: actual BG (black thick line) and IG (gray thick line),
measured CGM (black line), affected by random multiplicative calibration error (Eq. 5) and by random additive error (Eq. 4),
and measured SMBG (gray circles), affected by additive noise (Eq. 3).
355
number 3) in a representative day (Day 6). SMBG samples
were simulated by taking four BG samples per day and
adding a measurement error with coefficient of variation (CV)
of 2% (gray circles in Fig. 3), as follows:
SMBG(k) ¼ u(k)BG(k) þ v2 (k) (3)
where j(k) is a flag that is equal to 1 only when the SMBG
sample is collected.
Then, by applying Eq. 1 with time-invariant parameters
(i.e., g ¼ 1 and t ¼ 12 min), IG concentrations are obtained
(gray thick line in Fig. 3). Finally, the measured CGM profile
(black thin line in Fig. 3) is obtained as:
CGM(k) ¼ (1 þ a(k))IG(k) þ v1 (k) (4)
where IG(k) represents the interstitial glucose level at sam-
ple k, a(k) is a random multiplicative value that simulates
the possibly time-varying deviation of the sensor gain from
unitary level, and v1(k) is the random additive measurement
error. In particular, v1(k) is modeled as a zero mean white
Gaussian process with variance s2 ¼ 4 mg2=dL2.19,22,23 Here,
the gain deviation a(k), which, from a theoretical point, re-
sembles the scale factor of Knobbe and Buckingham,11 is
created as the triple integration of a zero mean white noise
process w1(k):
a(k þ 1) ¼ 3a(k)
3a(k
1) þ a(k
2) þ w1 (k) (5)
In the literature no models for simulating=generating the
gain deviation a(k) have been proposed. Here, we decided to
propose it as in Eq. 5 by using three, rather than one or two,
integrators because this better simulates the experienced loss
of performance of CGM sensors during long-term monitoring.
An example of a(k) profile is reported in the bottom panel of
Figure 4 (gray dashed line), where calibration errors of the
order of 10% of the reference value are obtained. Of note is
that the model for a(k) is zero mean. As in the work of Knobbe
12 18 24
356 FACCHINETTI ET AL.

True IG
220 CGM (enhanced)

200

glucose (mg/dL)
180

160

140

120
Day 2 Day 3 Day 4 Day 5 Day 6 Day 7
6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24

True BG
220 Estimated BG

200
glucose (mg/dL)

180

160

140

120 Day 2 Day 3 Day 4 Day 5 Day 6 Day 7

6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24
0.2

0.15
Day 2 Day 3 Day 4 Day 5 Day 6 Day 7
0.1

0.05
α

−0.05
Simulated
−0.1
Estimated

6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24 6 12 18 24
Time (hours)

FIG. 4. Representative subject number 3. (Top) True IG (gray dashed) versus enhanced CGM (black) time series. (Middle)
True (gray dashed) versus estimated (black) BG concentrations. (Bottom) Simulated (gray dashed) versus estimated (black) a
profiles.

and Buckingham,11 this assumption can be taken without any
loss of generality because the introduction of a bias term
would not introduce any change in the performance of the
method.
Enhanced Bayesian Calibration Method (BCM)
The new online calibration method exploits a stochastic
framework and Bayesian estimator. For sake of simplicity, in
the following, the acronym BCM will be used. As detailed
below, the inputs of BCM are the noisy CGM time series
(black thin line in Fig. 3) and the four SMBG samples per day
(gray circles in Fig. 3). The outputs are the enhanced CGM
time series, i.e., an estimation of the IG profile, an estimation
of the BG concentration, and, as a by-product, an estimation of
the gain deviation a(k).
BCM exploits a state space model representation of all un-
known signals (see The state space model below), and it is
ENHANCED ACCURACY OF CGM 357

implemented by using EKF (see The EKF implementation time process governed by a nonlinear stochastic differential
below). equation:

The state space model x(k þ 1) ¼ f (x(k), w(k)) (10)

Assuming that measured CGM and BG are described as with a measurement vector
in Eqs. 3 and 4, the next step is to choose suitable a priori
models for the unknown signals BG, IG, and a on the con- y(k) ¼ h(x(k), v(k)) (11)
sidered 1-min evenly spaced grid. The model selected to
describe the unknown BG at 1-min time sampling is an inte- where f and h are nonlinear vectorial functions and w(k) and
grated random walk model: v(k) are the vectors of the process and measurement noises,
respectively. Usually, w(k) and v(k) are supposed to be un-
BG(k þ 1) ¼ 2BG(k)
BG(k
1) þ w2 (k) (6) correlated each other and to be zero mean white noise pro-
cesses with covariance matrices Qk and Rk, respectively. In this
where BG(k) represents the blood glucose value at the k-th specific application, f and h are designed as follows:
sample and w2(k) is a zero mean Gaussian noise with variance
equal to k22 . The multiple integration of a white noise process is 2 3
2x1 (k)
x2 (k) þ w2 (k)
commonly used in Bayesian signal estimation to represent a 6 7
6 x1 (k) 7
regular signal, e.g., a physiological process.24 In addition, this 6 (1
(1=s)x3 (k) þ (1=s)x1 (k) 7
particular model has also been recently used to represent the f (x(k), w(k)) ¼ 6 7
6 3x4 (k)
3x5 (k) þ x6 (k) þ w1 (k) 7 (12a)
regular pattern of the glycemic signal in a context of CGM 6 7
4 x4 (k) 5
data denoising.25 Here, the choice of two, rather than one or x5 (k)
three, integrators has been performed by a preliminary anal-
ysis on real, high-frequently sampled, BG datasets taken from
 
the literature.26 x3 (k)x4 (k) þ v1 (k)
The a priori model for IG simply reflects the continuous- h(x(k), v(k)) ¼ (12b)
u(k)x1 (k) þ v2 (k)
time differential equation of Eq. 1, which, after discretization,
turns into Eq. 6: The estimation of the state vector x is assessed by exploiting

 equations that are very similar to the ones used in the linear
1 g case.27,28 In particular, the time-update step equations are:
IG(k þ 1) ¼ 1
IG(k) þ BG(k) (7)
s s
x(k þ 1jk) ¼ f (^
^ x(kjk), 0) (13a)
Finally, the a priori model for a(k) is the triple integration of
P(k þ 1jk) ¼ Ak P(kjk)ATk þ Wk Qk WkT (13b)
a zero mean white noise process w1(k) with variance equal to
k21 (see Eq. 5).
In Eqs. 13a and 13b x^(kjk) and P(kjk) are, respectively, the
In order to obtain a state-space dynamic model, formulated
estimate of the state vector x and its covariance matrix, ob-
by letting x1(k) ¼ BG(k), x2(k) ¼ BG(k
1), x3(k) ¼ IG(k),
tained at the k-th sampling time given the measurements until
x4(k) ¼ a(k), x5(k) ¼ a(k
1), and x6(k) ¼ a(k
2), one has:
time k included. The measurement update equations are:
2 3 2 3 2 3
x1 (k þ 1) 2x1 (k)
x2 (k) w2 (k)
6 x2 (k þ 1) 7 6 7 6 0 7 Kk ¼ P(k þ 1jk)HkT (Hk P(k þ 1jk)HkT þ Vk Rk VkT )
1 (14a)
6 7 6 x1 (k) 7 6 7
6 x3 (k þ 1) 7 6 (1
(1=s)x3 (k) þ (1=s)x1 (k) 7 6 0 7 x ^(k þ 1jk) þ Kk (y(k þ 1)
h(^
^(k þ 1jk þ 1) ¼ x x(k þ 1jk), 0)) (14b)
6 7¼6 7þ6 7
6 x4 (k þ 1) 7 6 3x4 (k)
3x5 (k) þ x6 (k) 7 6 w1 (k) 7 (8) P(k þ 1jk þ 1) ¼ (I
Kk Hk )P(k þ 1jk) (14c)
6 7 6 7 6 7
4 x5 (k þ 1) 5 4 x4 (k) 5 4 0 5
x6 (k þ 1) x5 (k) 0 In Eqs. 14a–c Ak and Hk are the Jacobian matrices of the
partial derivatives of f and h with respect to x, whereas Wk and
The measurements model, obtained from Eqs. 3 and 4, Vk are the Jacobian matrices of the partial derivatives of f and h
reads: with respect to wk and vk. Finally, Kk is the Kalman gain matrix
      at k-th sample, and I is the identity matrix (with size corre-
y1 (k) x (k)x4 (k) v (k) spondent to that of x).
¼ 3 þ 1 (9)
y2 (k) u(k)x1 (k) v2 (k) In this specific application, Ak and Hk are designed as
follows:
where y1(k) and y2(k) are the measured CGM and BG signals.
In Eq. 9 a nonlinear relationship between the two states x3(k) 2 3
2
1 0 0 00
and x4(k) [corresponding to IG(k) and a(k)] is present. For this 6 1
6 0 0 0 0077
reason, in order to recover the estimates ^x(k) starting from 6 7
6 1=s 0 (1
1=s) 0 007
y1(k) and y2(k), the version of the Kalman filter for nonlinear Ak ¼ 6 7,
6 0 0 0 3
3 17
problems, i.e., the so-called EKF, has to be used. 6 7
6 7
4 0 0 0 1 0 05
The EKF implementation 0 0 0 0 1 0
 
In general, an EKF27,28 can be used to provide a minimum 0 0 x4 (k) x3 (k) 0 0
Hk ¼ (15)
variance estimate of the state vector x of a dynamic discrete- u(k) 0 0 0 0 0
358
while Wk and Vk are two identity matrices of size 66 and
22, respectively. It is important to evidence that Hk change
with k.
Note that the method of Eqs. 8 and 9 and the EKF im-
plementation described above represent an improvement of
the approach proposed by Knobbe and Buckingham.11 In fact,
the aim is now focused on improving the accuracy of CGM
data, rather than on estimation of BG and physiological
parameters (e.g., t as in Knobbe and Buckingham11), by per-
forming an enhanced calibration in cascade to the standard
device calibration. Also, the model is different and, for some
aspects, simpler. The original model of Knobbe and Buck-
ingham11 has five state variables (i.e., BG, rate of change of
BG, IG, t, and the scale factor), each one with its own variance
(which has to be known a priori). Even if the new model has
apparently six states, the true unknown variables are only
three, i.e., BG, IG, and a. A further simplification is that, now,
only two variances (i.e., k21 and k22 ) have to be set (compared
to the five variances that had to be trained in Knobbe and
Buckingham11). Finally, a different description of the vari-
ables (i.e., BG is an integrated random walk process, and a is
the triple-integration of a random white noise) has also been
used in order to improve the method performance.
Results
BCM on simulated data
BCM has been tested on the 10 simulated virtual patients
created in Simulated Database. Data of Day 1 are used as the
burn-in interval. Matrices Rk and Qk in Eqs. 13 and 14 are
considered available. Concerning Rk, the variances of v1(k)
and v2(k) in Eq. 9 have been set to r21 (k) ¼ 4 mg2 =dL2 and
r22 (k) ¼ (0:02 · y2 (k))2 mg2 =dL2 , respectively. Concerning Qk,
the variance of w1(k) has been estimated for each subject in the
burn-in interval by resorting to a stochastically based
smoothing criterion based on maximum likelihood,24,25 and
the variance of v1(k) was set at k21 ¼ 10
19 (the same value
used to generate a profiles in Simulated Database). For each
subject, k22 estimated values are reported in Table 1. The four
SMBG samples per day were ‘‘optimally’’ drawn at breakfast
and dinner and at the correspondent glycemic peaks, ac-
cording to the accepted notion that using reference values
collected at very different glycemic levels (e.g., one at val-
ley and one at peak) improves calibration.17 This particular
scheduling suits real life because type 1 diabetes subjects
likely measure their glycemic level before a meal to calculate
the dosage of the insulin bolus and 1 h after to check if an
additional bolus correction is necessary.
Figure 4 shows the results for a representative subject
(number 3). The top panel displays the true IG (gray dashed)
and the enhanced CGM (black) profiles. The middle panel
shows the true BG (gray dashed) and the reconstructed BG
(black) profiles. The bottom panel displays the true a (gray
dashed), used to simulate this dataset, and its estimate (black)
obtained by BCM. Even if the procedure is based only on four
SMBG samples, IG is correctly reconstructed. A good estimate
of BG is obtained thanks to the incorporation of the BG-to-IG
dynamics into the state space model. The trend of a is nicely
reconstructed thanks to the inclusion of the description of
sensor gain by Eq. 5. Notably, a reflects the systematic under-
or overestimation of the glucose level made by the CGM de-
vice, and therefore it is a sort of indicator of its accuracy.
FACCHINETTI ET AL.
Table 1. Estimated k22 Values for Each
of the 10 Simulated Subjects
Subject k22
1 3.7
2 2.7
3 2.3
4 3.7
5 3.8
6 4.7
7 2.9
8 3.1
9 5.9
10 2.7
Focusing on the middle panel, we can see that BCM allows us
to correctly identify an underestimation in Days 4 and 5 and
an overestimation in Day 7.
In order to better understand the quality of BCM outputs,
in Figure 5 a 1-day detail (Day 6) has been reported. The top
row shows the true IG (gray dashed line), the measured CGM
(black thin line), and its enhanced version (black thick line),
whereas the middle row displays the true BG (gray dashed)
and the reconstructed BG (black) profiles. Looking at the
top row, it is possible to see that the enhanced calibration
completely eliminates the amplitude distortion present in the
measured CGM signal. As a by-product of the procedure,
the CGM profile has also been significantly denoised. In the
bottom row, a zoom of 10–14 h is highlighted (IG, CGM, and
enhanced CGM on the left side, BG and estimated BG on the
right side).
Comparison with other methods
In order to better assess the potential usefulness of the
method, below we make a comparison with both the model of
Knobbe and Buckingham11 and a recently proposed calibra-
tion procedure based on the linear regression model of Eq. 2.17
Because the procedure of King et al.,17 in its original formu-
lation, works only retrospectively, we have modified it using
a two-point strategy for the update of the parameters of Eq.
218 (with this modification, the method performs better than
the other calibration procedures previously proposed in the
literature).15,17
Optimal SMBG scheduling. To compare the three cali-
bration methodologies, the RMSE between true IG and en-
hanced CGM profiles is performed. Columns 1, 2, and 3 of
Table 2 show the RMSE for BCM, the model of Knobbe and
Buckingham,11 and the two-point calibration procedure, re-
spectively, in the 10 simulated virtual patients, together with
mean and SD values. BCM performs better than the other
approaches. In particular, the RMSE mean value has been
significantly reduced (82%, P < 0.001 for the model of Knobbe
and Buckingham11; 33%, P ¼ 0.01 for two-point calibration,
Wilcoxon rank sum test). Furthermore, by looking at the SD
values, BCM performs similarly and satisfactory for all sub-
jects, whereas the algorithm of Knobbe and Buckingham11
performs unsatisfactorily in all subjects except for number 6,
ENHANCED ACCURACY OF CGM 359

True IG
220
CGM
CGM (enhanced)
200
glucose (mg/dL)
180

160

140

120

100
0 6 12 18 24

True BG
220 Estimated BG

200
glucose (mg/dL)

180

160

140

120

100
0 6 12 18 24
220 220
True IG True BG
CGM Estimated BG
200 CGM (enhanced) 200
glucose (mg/dL)

180 180

160 160

140 140

120 120
10 11 12 13 14 10 11 12 13 14
Time (hours) Time (hours)

FIG. 5. Subject number 3, detail of Day 6. (Top) True IG (gray dashed), noisy CGM (black thin) with lack of calibration, and
enhanced CGM (black thick) time series. (Middle) True (gray dashed) versus estimated (black) BG profiles. (Bottom) Zoom of
the time window 10–14 h, true IG (gray dashed), noisy CGM (black thin) with lack of calibration, and enhanced CGM (black
thick) time series (left panel) and true (gray dashed) versus estimated (black) BG profiles (right panel).
360 FACCHINETTI ET AL.

Table 2. RMSE (in mg=dL) for Optimal (i.e., Valleys and Peaks) and Suboptimal
(i.e., Close to Valleys and Peaks) SMBG Scheduling

Optimal scheduling Suboptimal scheduling (30-min window) Suboptimal scheduling (2-h window)
Two SMBGs,
Subject BCM Knobbe 2-point BCM 2-point BCM 2-point BCM

1 3.7 29.5 5.2 3.8 6.6 4.0 14.0 5.9

2 2.7 12.4 4.6 2.3 3.7 2.9 17.5 2.5
3 2.3 20.9 4.2 2.6 6.3 2.9 15.6 4.2
4 3.7 6.9 3.7 3.9 6.2 4.1 14.0 6.5
5 3.8 16.4 5.3 4.1 4.8 3.8 7.8 6.7
6 4.7 6.4 5.4 4.6 5.6 4.9 13.0 7.2
7 2.9 34.4 9.0 3.0 10.3 2.9 19.0 4.8
8 3.1 10.1 3.5 2.9 3.4 2.9 8.9 3.1
9 5.9 40.2 6.4 6.5 9.7 6.9 9.6 8.2
10 2.7 16.6 5.1 2.6 5.9 2.3 7.9 3.7
Mean 3.5 19.4 5.2 3.6 6.3 3.7 12.7 5.3
SD 1.1 11.7 1.6 1.3 2.3 1.4 4.0 1.9

and the two-point procedure presents unsatisfactory results
for some of them, in particular for subject number 7. Because
the two-point procedure performs better than the EKF of
Knobbe and Buckingham11 (73%, P < 0.001), hereafter only
the two-point calibration will be used as reference.
Figure 6 shows a graphical comparison on Day 6 of the
representative dataset. The true IG (gray dashed), the mea-
sured CGM (black thin), the two-point calibrated (black dot-
ted), and the BCM calibrated (black thick) signals are plotted.
As apparent even by graphical inspection, BCM is able to
perform a satisfactory CGM data enhancement, producing
better results than the two-point method. In addition, BCM
also returns a smoothed profile with a significant reduction in
the noise component.
Robustness to suboptimal SMBG scheduling. In this
subsection we investigate the situation in which a suboptimal
scheduling of the four SMBG measurements per day is pro-
vided. In particular, each of the four SMBG samples per day is
collected randomly in a time interval of 30 min centered at the
‘‘optimal’’ sampling times used in the previous simulations.
Columns 4 and 5 of Table 2 report the RMSE for this first
simulation. Also in this case, BCM performs much better than
the two-point method, with a significant reduction of the
RMSE of 43% (P < 0.01). Notably, BCM performance is almost
insensitive to the suboptimality of SMBG sampling (þ3%,
comparing average results of columns 1 and 4), whereas the
two-point procedure has a significant decrease in perfor-
mance (þ21%, comparing average results of columns 3 and 5).
In order to stress the performance of BCM, a second test
has been performed by increasing the time interval to 2 h.
RMSE values are reported in columns 6 and 7 for BCM and the
two-point procedure, respectively. The BCM mean RMSE
increases of only 5.4% if compared to results of optimal
scheduling (3.7 vs. 3.5, respectively) and, not surprisingly, is
significantly better than the two-point calibration in terms of
RMSE (71%, P < 0.001). In addition, by looking at SD values,
RMSE variability from subject to subject is similar to the ideal
situation only for BCM.
Robustness to reduction of SMBG measurements. It is
of interest to test the performance of BCM by reducing the
number of SMBG samples exploited. In particular, a new
SMBG scheduling has been drawn in which only two mea-
surements per day are provided. The two samples have been
collected one at breakfast and one at the postprandial dinner
peak, respectively, and both with suboptimal timing (ran-
domly taken in the 2-h window around each point). For each
subject RMSE values are reported in the last column of
Table 2. By looking at the results it is evident that even if
resorting only to two SMBG samples per day, BCM mean
RMSE increases by only 33% if compared to results of four
SMBG samples with optimal scheduling. This confirms that
BCM is able to perform satisfactorily also in suboptimal
conditions concerning both number and scheduling of SMBG
samples.
Finally, we also tested the performance of BCM by apply-
ing higher CV values to SMBG measurements and obtained
results comparable to the ones presented above (data not
shown).
Robustness to errors in t. We also investigated the in-
fluence of the value of t in Eq. 7. The value of t can signifi-
cantly vary from subject to subject,15 and its individual
identification is impossible in practice because it would re-
quire abundant BG references per day.
The first simulation is aimed to test BCM if, inside the
model, a wrong t value is used. In particular, we choose
t values from 6 to 18 min, i.e., we consider the worst case in
which the BG-to-IG model has 50% of error. Figure 7 shows
the BCM average RMSE values (black profile) obtained
comparing true IG and enhanced CGM time series. Even in
the worst case, i.e., by using t ¼ 6 min, BCM performance in
terms of RMSE is better than for the two-point method (3.9 vs.
5.2, respectively). If compared with the result of the optimal
SMBG scheduling of section (see BCM on simulated data)
(average RMSE ¼ 3.5), the average RMSE increases only by
12%. This means that, even if we use inside BCM a t value that
is lower=greater than 50% of the real one, BCM still performs
better than the calibration procedure of King et al.17
In the second simulation, we compared the performance of
individualized versus ‘‘population’’ BCM. The dataset created
in Simulated Database has been modified in order to take into
account the interindividual variability of t. Ten new simu-
lated subjects have been created, each one with a different t
drawn from a Gaussian distribution (18  4 min [see Facchi-
ENHANCED ACCURACY OF CGM 361

True IG
220 CGM
2−point
200 BCM
glucose (mg/dL)

180

160

140

120

100
0 6 12 18 24
Time (hours)

FIG. 6. Subject number 3, detail of Day 6: true IG profile (gray dashed), CGM data (black thin), two-point calibrated (black
dotted), and BCM enhanced (black thick) CGM profiles.

netti et al.15]). The mean and SD values of the RMSE were
3.6  1.1 and 4.1  1.4 mg=dL, by using the individualized
and the ‘‘population’’ BCM, respectively. This result evidences
that BCM, even with ‘‘population’’ parameters, performs
satisfactorily.
BCM on a representative subject
BCM was also tested on the same subject of Figure 1. As one
can see by comparing the two time series, the CGM profile is
characterized by an evident underestimation of the glycemic
level in the time interval 17–21 h. Two of the BG references
have been selected to simulate two SMBG measurements (see
the gray circles plotted in the top panel of Fig. 8). Then, BCM
has been applied as data were received in real time. By
looking at the results, a zoom of which is shown in Figure 8,
not only has the underestimation of the original CGM data
been correctly identified by estimating a negative value of a in
correspondence to the time window 15–21 h (bottom panel),
but also it has been satisfactorily recovered. In fact, the en-
hanced CGM profile (black series in the top panel) presents
increased CGM values exactly in that window if compared to
the original one. Quantitatively, the RMSE calculated in 18–
26 h, i.e., after the two SMBG samples have been exploited,
between BG references (we remember not used by BCM) and
CGM profile has been reduced from 32.0 to 25.5 mg=dL
thanks to the enhanced calibration.
Conclusions
In this work we have proposed a new online enhanced
calibration method that can be applied in cascade to the
standard CGM device calibration in order to improve
the accuracy of CGM output. This method, which is an

3.9

3.8
mg/dL

3.7

3.6

3.5
6 8 10 12 14 16 18
τ (min)
FIG. 7. The average BCM RMSE obtained comparing true IG versus enhanced CGM time series by using a wrong t value in
the BG-to-IG kinetics model (50% of the true value, i.e., 12 min).
362 FACCHINETTI ET AL.

400
CGM
CGM (enhanced)
350
SMBG

300
glucose (mg/dL)

250

200

150

100
12 14 16 18 20 22 24 26
0.1
Estimated α

0.05
α

−0.05

−0.1
12 14 16 18 20 22 24 26
Time (hours)

FIG. 8. Representative real subject. (Top) Measured (gray) versus enhanced (black) CGM time series as well as SMBG
samples (gray circles). (Bottom) Estimated a profile.

improvement of the algorithm of Knobbe and Buckingham,11
is embedded within a stochastic framework and implemented
by EKF. By taking into account BG-to-IG kinetics, few SMBG
samples per day, and a model describing the sensor gain
variability, the enhanced calibration method significantly
improves the accuracy of CGM data in real-time. Further-
more, the method intrinsically denoises the CGM signal and,
as a by-product, also provides a continuous-time estimate of
both BG and sensor gain.
Compared to state of art literature calibration strategy, the
new method performs significantly better, when both optimal
and suboptimal scheduling of the four SMBG measurements
is done and when the number of the SMBG per day is reduced.
Also, the method exhibits small sensitivity to variations of the
BG-to-IG time constant t. In addition, when applied to a real
dataset, BCM is shown to correctly identify under- and
overestimations of original CGM profile and to improve
sensor outcome.
Finally, we note that the application of the BCM to real data
requires the knowledge of the variances of both state and
measurement processes, which in real-life conditions are
usually unknown. Therefore, a real-time method for this es-
timation should be considered. Another aspect is related to
the need of a burn-in interval. Here, 1 day was exploited in
order to estimate the covariance matrix of the state vector in
the transient phase.
In conclusion, the enhanced calibration algorithm signifi-
cantly improves the accuracy of CGM signals, suggesting that
it could be profitably included in real-time applications of
CGM devices, e.g., algorithms for open=closed-loop diabetes
control. The model here used for the description of the sensor
gain deviation a is rather general and could be used in cascade
to make an enhanced calibration for any CGM device. In
addition, the method could be also used retrospectively to
efficiently improve the accuracy of CGM profiles before ap-
plying to those data any type of post-processing.
Acknowledgments
We thank Dr. Chiara Dalla Man (University of Padova,
Padova, Italy) for having provided the software13,20,21 and for
her helpful support in its use in the simulations reported in the
present work.
Author Disclosure Statement
The authors have no conflicts of interest.
ENHANCED ACCURACY OF CGM 363

References 15. Facchinetti A, Sparacino G, Cobelli C: Reconstruction of

glucose in plasma from interstitial fluid continuous glucose
1. Nichols JH, Klonoff DC: The need for performance stan-
monitoring data: role of sensor calibration. J Diabetes Sci
dards for continuous glucose monitors. J Diabetes Sci
Technol 2007;1:617–623.
Technol 2007;1:92–95.
16. Breton M, Kovatchev BP: Analysis, modeling, and simula-
2. De Block C, Vertommen J, Manuel-y-Keenoy B, Van Gaal L:
tion of the accuracy of continuous glucose sensors. J Dia-
Minimally-invasive and non-invasive continuous glucose
betes Sci Technol 2008;2:853–862.
monitoring systems: indications, advantages, limitations and
17. King C, Anderson SM, Breton M, Clarke WL, Kovatchev
clinical aspects. Curr Diabetes Rev 2008;4:159–168.
BP: Modeling of calibration effectiveness and blood-to-
3. Garg K, Zisser H, Schwartz S, Bailey T, Kaplan R, Ellis S,
interstitial glucose dynamics as potential confounders of
Jovanovic L: Improvement in glycemic excursions with a
the accuracy of continuous glucose sensors during hyper-
transcutaneous, real-time continuous glucose sensor. Dia-
insulinemic clamp. J Diabetes Sci Technol 2007;1:317–322.
betes Care 2006;29:44–50.
18. Lodwig V, Heinemann L; Glucose Monitoring Study Group:
4. Klonoff DC: Continuous glucose monitoring: roadmap for
Continuous glucose monitoring with glucose sensors: cali-
21st century diabetes therapy. Diabetes Care 2005;28:1231–
bration and assessment criteria. Diabetes Technol Ther
1239.
2003;5:572–586.
5. Battelino T, Bolinder J: Clinical use of real-time continuous
19. Kuure-Kinsey M, Palerm CC, Bequette BW: A dual-rate
glucose monitoring. Curr Diabetes Rev 2008;4:218–222.
Kalman filter for continuous glucose monitoring. Conf Proc
6. Simpson P, Brister M, Wightlin M, Pryor J, inventors; Dex-
IEEE Eng Med Biol Soc 2006;1:63–66.
Com, Inc., assignee: Dual electrode system for a continuous
20. Dalla Man C, Raimondo DM, Rizza RA, Cobelli C: GIM,
analyte sensor. Publication Number WO=2008=042918; April
simulation software of meal glucose-insulin model. J Dia-
10, 2008. International Application Number PCT=US2007=
betes Sci Technol 2007;1:323–330.
080228; October 4, 2007.
21. Dalla Man C, Rizza RA, Cobelli C: Meal simulation model of
7. Feldman BJ, McGarraugh GV, inventors; Abbott Diabetes
the glucose-insulin system. IEEE Trans Biomed Eng 2007;54:
Care, Inc., assignee: Method of calibrating an analyte-
1740–1749.
measurement device, and associated methods, devices and
22. Kovatchev BP, Clarke WL, Breton M, Brayman K, McCall A:
systems. U.S. Patent 7,299,082. November 20, 2007.
Quantifying temporal glucose variability in diabetes via
8. Mastrototaro JJ, Gross TM, Shin JJ, inventors; Medtronic
continuous glucose monitoring: mathematical methods and
Minimed, Inc., assignee: Glucose monitor calibration meth-
clinical application. Diabetes Technol Ther 2005;7:849–862.
ods. U.S. Patent 6,424,847. July 23, 2002.
23. Chase JG, Hann CE, Jackson M, Lin J, Lotz T, Wong XW,
9. Kovatchev BP, Gonder-Frederick LA, Cox DJ, Clarke WL:
Shaw GM: Integral-based filtering of continuous glucose
Evaluating the accuracy of continuous glucose-monitoring
sensor measurements for glycaemic control in critical care.
sensors: continuous glucose-error grid analysis illustrated by
Comput Methods Programs Biomed 2006;82:238–247.
TheraSense Freestyle Navigator data. Diabetes Care
24. De Nicolao G, Sparacino G, Cobelli C: Nonparametric input
2004;27:1922–1928.
estimation in physiological systems: problems, methods and
10. Rebrin K, Steil GM, van Antwerp WP: Subcutaneous glucose
case study. Automatica 1997;33:851–870.
predicts plasma glucose independent of insulin: implications
25. Facchinetti A, Sparacino G, Cobelli C: An on-line self-
for continuous monitoring. Am J Physiol 1999;277:E561–
tuneable method to denoise CGM sensor data. IEEE Trans
E571.
Biomed Eng 2010;57:634–641.
11. Knobbe EJ, Buckingham B: The extended Kalman filter for
26. http:==glucosecontrol.ucsd.edu (accessed April 1, 2009).
continuous glucose monitoring. Diabetes Technol Ther 2005;
27. Anderson BDO, Moore JB: Optimal Filtering. Dover Pub-
7:15–27.
lications, Mineola, NY, 2005.
12. Magni L, Raimondo DM, Bossi L, Dalla Man C, De Nicolao
28. Grewal MS, Andrews AP: Kalman Filtering: Theory and
G, Kovatchev B, Cobelli C: Model predictive control of type
Practice Using MATLAB. John Wiley and Sons, New York,
1 diabetes: an in silico trial. J Diabetes Sci Technol 2007;1:
2001.
804–812.
13. Kovatchev BP, Breton M, Dalla Man C, Cobelli C: In silico
preclinical trials: a proof of concept in closed-loop control of Address correspondence to:
type 1 diabetes. J Diabetes Sci Technol 2009;3:44–55. Prof. Claudio Cobelli, Ph.D.
14. Diabetes Research in Children Network (Direcnet) Study Department of Information Engineering
Group, Buckingham BA, Kollman C, Beck R, Kalajian A, University of Padova
Fiallo-Scharer R, Tansey MJ, Fox LA, Wilson DM, Weinzi- Via Ognissanti 72
mer SA, Ruedy KJ, Tamborlane WV: Evaluation of factors 35131 Padova, Italy
affecting CGMS calibration. Diabetes Technol Ther 2006;8:
318–325. E-mail: cobelli@dei.unipd.it