> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
1
STAR Development and Protocol Comparison
Liam M. Fisk, Aaron J. Le Compte, Geoffrey M. Shaw, Sophie Penning, Thomas Desaive, J. Geoffrey
Chase

Abstract—Accurate glycemic control (AGC) is difficult due to
excessive hypoglycemia risk. Stochastic TARgeted (STAR)
glycemic control forecasts changes in insulin sensitivity to
calculate a range of glycemic outcomes for an insulin
intervention, creating a risk framework to improve safety and
performance. An improved, simplified STAR framework was
developed to reduce light hypoglycemia and clinical effort, while
improving nutrition rates and performance. Blood glucose (BG)
levels are targeted to 80 – 145mg/dL, using insulin and nutrition
control for 1-3 hour interventions. Insulin changes are limited to
+3U/hour and nutrition to ±30% of goal rate (minimum 30%).
All targets and rate change limits are clinically specified and
generalizable. Clinically validated virtual trials were run on
using clinical data from 371 patients (39,841hours) from the
SPRINT cohort. Cohort and per-patient results are compared to
clinical SPRINT data, and virtual trials of three published
protocols. Performance was measured as time within glycemic
bands, and safety by patients with severe (BG<40mg/dL) and
mild (%BG<72mg/dL) hypoglycemia. Pilot trial results from the
first 10 patients (1,458 hours) are included to support the in-silico
findings.
In both virtual and clinical trials, mild hypoglycemia was
below 1% versus 4% for SPRINT. Severe hypoglycemia was
reduced from 14 (SPRINT) to 6 (STAR), and 0 in the pilot trial.
AGC was tighter than both SPRINT clinical data and in silico
comparison protocols, with 91% BG within the specified target
(80–145mg/dL) in virtual trials and 93.4% in pilot trials. Clinical
effort (measurements) was reduced from 16.2/day to 12.0/day
(13.9/day in pilot trials). This STAR framework provides safe,
accurate glycemic control with significant reductions in
hypoglycemia and clinical effort due to stochastic forecasting of
patient variation – a unique risk-based approach. Initial pilot
trials validate the in silico design methods and resulting protocol,
all of which can be generalized to suit any given clinical
environment.
Index
Terms—Biomedical
computing,
forecasting, stochastic approximation
clinical
trial,
I. INTRODUCTION
S
TRESS-INDUCED hyperglycemia is a significant issue
in critical care, affecting up to 30-50% of patients and
Manuscript received January 10, 2012. This work was supported in part by
the New Zealand Tertiary Education Commission and the NZ Foundation for
Research Science and Technology.
L.M. Fisk, A. J. Le Compte, and J. G. Chase are with the Department of
Mechanical Engineering, University of Canterbury, Christchurch, New
Zealand (phone: +64-3-364-2596, fax: +64-3-364-2078, e-mail:
geoff.chase@canterbury.ac.nz)
S. Penning and T. Desaive are with the University of Liege, Belgium.
G. M. Shaw is with the Department of Intensive Care, Christchurch
Hospital, Christchurch, New Zealand.
increasing morbidity and mortality [1, 2]. Controlling
glycemia has proved difficult due to the associated risk of
hypoglycemia when highly dynamic patients are treated with
exogenous insulin [3]. Both extremes, as well as glycemic
variability, have been independently linked to increased
morbidity and mortality [4-6], creating a difficult clinical
problem safely and effectively regulating glycemia to a
physiologically and clinically safe range.
Accurate glycemic control (AGC) can mitigate these
outcomes [7-9], but has proven difficult to achieve safely and
consistently [10-12]. Only one study [8] reduced both
mortality and hypoglycemia. However, the higher nursing
workloads due to high density glucose readings are
impractical in many units [13, 14]. Hand-held glucometers are
easier for measurement, but their larger errors can add
additional difficulty for some AGC protocols. Finally, clinical
compliance determines much of the efficacy of any AGC
method, with quality of glycemic control thus also limited by
the confidence and compliance of nursing staff [14, 15]. All of
these issues interact with the inherent inter- and intra- patient
metabolic variability [16] to exacerbate the difficulty of
achieving good control. Hence, glycemic control targets have
been raised to mitigate these factors and avoid hypoglycemia
[17], as a best outcome compromise, despite the physiological
and clinical evidence on the negative impact of even moderate
hyperglycemia [1, 2, 18].
Directly quantifying and managing hyperglycemic and
hypoglycemic risk as a function of inter- and intra- patient
metabolic variability can leverage AGC benefits and minimize
risk of unintended harm. STAR (Stochastic TARgeted) is a
model-based AGC framework that can be implemented across
a range of clinical scenarios and approaches, and uses
dynamic and stochastic models to regulate BG levels,
workload and patient safety within a pre-defined risk
management approach. STAR uses stochastic forecasting of a
patient's potential metabolic variability [19, 20] in conjunction
with a clinically validated mathematical model [21-23] to
determine optimal insulin and nutrition treatment
combinations with specified risks of moderate hyper- and
hypo- glycemia. In essence, it is a patient-specific approach to
manage inter- and intra- patient variability that overlaps the
glycemic outcome range for a given intervention with a
clinically specified desired glycemic range. It thus provides
both control and a clinically specified risk of moderate
hypoglycemia. Hence, STAR provides a framework of models
and methods to manage intra- and inter- patient variability to
mitigate the significant difficulty and risk seen in current
glycemic control approaches [3, 16].
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
2
This paper presents an enhanced and simplified STAR
protocol, and its development and optimization using virtual
trials. Specific focus is placed on reducing mild hypoglycemia
(BG < 72mg/dl) and its associated risk [4], while maintaining
glycemia in bands with the best evidence for improved
outcome. Protocol simplicity and transparency are optimized
to increase compliance.
II. METHODS
A. Model
The clinically validated Intensive Care Insulin-NutritionGlucose (ICING) metabolic model was used to simulate the
fundamental metabolic dynamics [22]. Table I lists the
population constants of the model defined in Equations 1-6,
and a graphical overview is presented in Figure 1.
Ġ = -pG G(t)- SI G(t)
İ = -
nL I(t)
1+ αI I(t)
Q(t)
1+ αG Q(t)
+
P(t)+ EGPb -CNS+PN(t)
-nK I(t)- (I(t)-Q(t)) nI +
Q̇ = (I(t)-Q(t)) nI - nC
Q(t)
Vg
uex (t)
VI
+(1-xL )
(1)
uen (G)
VI
1+ αG Q(t)
(2)
(3)
Ṗ 1 = -d1 P1 +P(t)
(4)
Ṗ 2 = - min(d2 P2 , Pmax ) +d1 P1
(5)
uen = max (16.67,
14G(t)
1+0.0147G(t)
-41)
(6)
where G(t) [mmol/L] is the total plasma glucose, I(t) [mU/L]
is the plasma insulin and interstitial insulin is represented by
Q(t) [mU/L]. Exogenous insulin input is represented by uex(t)
[mU/min] and endogenous insulin production is estimated
with uen [mU/min], modeled as a function of plasma glucose
concentration determined from critical care patients with a
minimum pancreatic output of 1U/hr. First-pass hepatic
insulin clearance is represented by xL, and nI [1/min] accounts
for the rate of transport between plasma and interstitial insulin
compartments. Patient endogenous glucose clearance and
TABLE I
CONSTANTS USED IN SYSTEM MODEL OF EQUATIONS (1)-(6)
Model variable
pG
SI
αG
d1
d2
Pmax
EGPb
CNS
VG
nI, nC
αI
VI
xL
nK
nL
a
Numerical value (or typical range)
0.006 min-1
[1x10-7 – 1x10-2] L/(mU.min)a
1/65 L/mU
-ln(0.5)/20
-ln(0.5)/100
6.11 mmol/min
1.16 mmol/min typically
0.3 mmol/min
13.3 L
0.0075 min-1
1.7x10-3 L/mU
4.0 L
0.67
0.0542 min-1
0.1578 min-1
Insulin sensitivity (SI) is identified from clinical data in the range shown.
Fig.1. Schematic illustration of the ICING model physiological overview
insulin sensitivity are pG [1/min] and SI [L/(mU.min)],
respectively. The parameter VI [L] is the insulin distribution
volume, nK [1/min] and nL [1/min] represent the clearance of
insulin from plasma via renal and hepatic routes respectively.
Basal endogenous glucose production (unsuppressed by
glucose and insulin concentration) is denoted by EGPb
[mmol/min] and VG [L] represents the glucose distribution
volume. CNS [mmol/min] represents non-insulin mediated
glucose uptake by the central nervous system. MichaelisMenten kinetics are used to model saturation, with α I [L/mU]
used for the saturation of plasma insulin clearance by the liver,
and αG [L/mU] for the saturation of insulin-dependent glucose
clearance and receptor-bound insulin clearance from
interstitium. P1 [mmol] represents the glucose in the stomach
and P2 [mmol] represents glucose in the gut. The rate of
transfer between the stomach and gut is represented by d 1
[1/min], and the rate of transfer from the gut to the
bloodstream is d2 [1/min]. Enteral glucose input is denoted
P(t) [mmol/min], and P max represents the maximum disposal
rate from the gut.
B. Virtual Patients
Clinically validated virtual trials [23] were carried out using
the SPRINT AGC cohort clinical data [8] (371 patients,
39,841 hours, 26,646 measurements) to create virtual patients.
Virtual patients are created using clinical data to identify an
hourly treatment-independent insulin sensitivity profile SI(t)
[24], allowing virtual trials to realistically simulate patient
response to a given (modified) treatment. This approach has
been clinically validated on independent matched cohort data
[23] and in several AGC trials [21, 25-27]. Patient
demographics are given in Table II.
Patients were considered to require AGC once BG >
7.0mmol/L, and this value was used to determine the
beginning of a virtual trial. Interruptions in nutrition are
common for some patients in clinical practice, and are
incorporated by setting P(t) = 0 mmol/min over the same
periods they occurred in the clinical data. Equally, clinically
specified parenteral nutrition (PN) was included in the
simulations just as it was given clinically.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
PATIENT DEMOGRAPHICS
SPRINT cohort
Total patients
Age (years)
% Male
APACHE II score
APACHE II risk of death
Diabetic history
APACHE III Diagnosis:
Operative
Cardiovascular
Respiratory
Gastrointestinal
Neurological
Trauma
Other (Renal,
metabolic, orthopaedic)
Non-operative
Cardiovascular
Respiratory
Gastrointestinal
Neurological
Trauma
Other (Sepsis, renal,
metabolic, orthopaedic)
371
65 [49 – 74]
63.6%
18 [15 – 24]
25.7% [13.1% - 49.4%]
62 (16.7%)
Num. patients
76
9
60
7
14
4
%
20%
2%
16%
2%
4%
1%
Num. patients
39
66
10
20
32
17
%
11%
18%
3%
5%
9%
5%
C. Stochastic Control Method
STAR provides patient-specific treatment in real time.
Stochastic forecasting provides a framework for control of
future outcomes, particularly the mitigation of mild, moderate,
and severe hypoglycemia. STAR was also developed with the
intent to use the simplest, most transparent control logic [15].
This latter aspect had the goal of ensuring its choices were as
understandable as possible, and thus directly translated into
safe, effective and clinically acceptable treatment
recommendations to maximize compliance.
When a BG measurement is entered, the model is used to
evaluate current patient insulin sensitivity [24] and its likely
variation [20] over the next 1 to 3 hours. Insulin is
administered in bolus form for safety from unintended delays
[28]. However, infusions may also be used. Robustness to
glucometer measurement error limits increases in insulin rate
to +2U/hour, with upper limits on the total bolus dose (6U/hr)
and any added infusion rate for highly resistant patients
(3U/hr). Thus, total insulin is limited to 9U/hr. Insulin can be
reduced to 0U/hr from any rate if required. Enteral nutrition is
controlled between 30-100% of ACCP goal [29], and changes
are limited to ±30% per intervention cycle. However, nutrition
administration can be set to a fixed constant rate or zero if
clinically specified.
Measurement or treatment interval is specifically limited
when: 1) current measured BG is outside the specified target
range of 80-145mg/dL (1-hourly limits);; or 2) the patient is
unable to be fed (2-hourly maximum interval). Otherwise,
when the current measured BG is within the clinically
specified target range, STAR calculates intervention options
(insulin and nutrition) for 1, 2 and 3 hourly measurement
intervals.
For each allowed insulin/nutrition combination and
treatment interval based on the limits specified, stochastic
forecasts are generated for the predicted 5th percentile of BG
outcomes (Figure 2, points A, B and C) and the predicted 95th
percentile of BG outcomes (Figure 2, points D, E and F). For
all treatment intervals, glycemic level is controlled by
targeting the 5th percentile of BG outcomes to the lower limit
of the desired range including tolerance (80-85mg/dL).
Hypoglycemia is thus directly managed as insulin rates cannot
be recommended if the predicted 5th percentile is below this
limit. The tolerance on the lower limit ensures consistent
interventions between measurement intervals offered (1, 2 and
3-hourly). The 5th percentile target is prioritized for control
due to the skewed nature of the BG outcome distribution, as
depicted in Figure 2, which ensures BG outcomes best overlap
the lower (80 - 125mg/dL) desired portion of the 80 145mg/dL range. This 80 - 125mg/dL range is associated with
better outcomes [5, 9] and is also associated with reduced rate
and severity of organ failure [30].
Tightness of the AGC provided by STAR is thus
determined by the treatment of the 95th percentile forecast
outcomes. This upper limit is used to restrict treatment interval
only if a desired 2- or 3- hourly treatment allows 95th
percentile BG above the target range. Monitoring the
likelihood of predicted BG outcome, as shown in Figure 2,
allows for more explicit direct control over intra-patient
variability and therefore directly limits the risk and occurrence
of mild hyperglycemia. Hourly measurements are always
offered, regardless of the 95th percentile forecast.
Figure 3 details the logic used to determine whether an
insulin/nutrition combination is accepted for a treatment
interval longer than 1 hour. If the 5th percentile BG is forecast
to be within tolerance of the target band lower limit the
treatment is considered acceptable (Figure 3, panel A). For
cases where the forecast BG range does not meet this criteria
there are two possibilities. Figure 3, panel B is considered
acceptable as the entire expected BG outcome is in the
specified desired band. The risk of hyperglycemia shown in
Figure 3, panel C is higher, and the outcome range includes
less of the lower part of the band when the 80 - 85mg/dL level
160
Upper Limit of Target Band
BG (mg/dL)
TABLE II
3
E
140
D
120
Forecasted BG
100
F
Measured BG
A
80
60
Tolerance on Lower Limit
B
0
2
C
4
6
8
Time (hrs)
Fig.2. Controller forecast schematic for BG a target range of 80 –
145mg/dL. A BG measurement has been taken at 10hrs, and forecasts of
BG have been generated (points A-F). The depicted distribution
indicates the skewed nature of BG forecasts within the 5th-95th
percentiles.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
Is the predicted 5th percentile of BG outcomes within
80-85mg/dL?
No
180
A
Acceptable
160
140
120
100
80
0
2
4
Time since measurement (hrs)
180
B
Acceptable
160
140
120
100
80
0
2
4
Time since measurement (hrs)
BG (mg/dL)
BG (mg/dL)
BG (mg/dL)
Yes
Is 95th percentile of
BG outcomes below
145 mg/dL?
Yes
No
180
C
Not acceptable
160
140
120
100
80
0
2
4
Time since measurement (hrs)
Fig.3. Selection logic for the possible forecasts possible for treatments at a
3-hourly treatment interval. The lower blue line depicts the 5th percentile
of BG outcomes, and the upper blue line depicts the 95th percentile of BG
outcomes. Target glycemic range (including tolerance on the lower bound)
is shown by the horizontal black lines.
is not reached. Thus, this insulin/nutrition combination is
considered unacceptable.
STAR maximizes performance (time in glycemic bands)
with a minimum, clinically specified risk of mild
hypoglycemia (5% for BG < 80mg/dL, ≈1% for BG
<72mg/dL). Within this goal, the control determines the
allowable insulin/nutrition combinations. Importance is placed
on maximizing nutrition rates [31], particularly for longer stay
patients, but not at the risk of exacerbating hyperglycemia.
Hence, STAR ranks allowable treatments by nutrition rate,
ensuring the treatment with the highest enteral nutrition rate is
selected.
The longest measurement intervals are calculated first. If an
acceptable treatment is found (Figure 3), the selected nutrition
rate is used as a lower limit for shorter treatment intervals.
This approach ensures treatment consistency across all
intervention and measurement intervals to maximize
transparency and clinical acceptance, and thus compliance
[15]. Specifically, it ensures an intuitive combination of
treatment options, where longer measurement intervals also
generally yield wider stochastic forecasting bounds and thus
more conservative (lower insulin) treatment choices.
D. Virtual Trials
Virtual trials were carried out to verify performance before
clinical testing. STAR is simulated in two forms: a) maximum
measurement interval available is chosen (“STAR - Max”);
and b) select at maximum 2-hourly intervals when available
4
(“STAR 2-hourly”) to best compare with SPRINT, which had
a 2-hourly maximum interval. Results were compared to
clinical SPRINT data to demonstrate improvements in
performance and safety over the currently utilized SPRINT
protocol that successfully reduced mortality [8] and organ
failure [30]. Three other published protocols: UNC [32], Yale
[33] and GluControl A [10] were simulated in virtual trials for
comparison and context.
E. Analyses/Performance Metrics
Performance was defined as percentage of BG within
selected glycemic bands. Clinical effort is evaluated by BG
measurement frequency as a surrogate [13, 14]. Safety was
defined as the incidence of severe (number patients with BG <
40mg/dL) and mild (%BG < 72mg/dL) hypoglycemia. All BG
data was resampled hourly to provide a consistent time-basis
for comparison across protocols with different measurement
and intervention intervals.
Finally, the virtual trial approach is further validated in
comparison to results for the first 10 patients (1458 hours,
demographic shown in Table III) in initial STAR pilot clinical
trials. All patients were treated for length of stay after
informed consent was obtained. Approval for this study and
use of the data was given by the Upper South A Ethics
Committee (URA/10/09/069).
III. RESULTS
Table IV shows both versions of STAR reduce the number
of cases of severe hypoglycemia and more than halve the
measures of mild hypoglycemia compared to SPRINT. There
is a 79% reduction in severe hypoglycemia between STAR 2hourly and SPRINT showing the importance of the STAR
approach independent of measurement interval, as both
protocols have a 2-hour maximum interval and similar
Age
TABLE III
PILOT TRIAL PATIENT DEMOGRAPHICS
Risk of
APACHE
APACHE III
Gender
death
II
diagnosis
(%)
Mortality
85
F
38
Cardiovascular
95.4
N
70
M
14
Respiratory
13.6
Y
61
M
26
Respiratory
56.9
N
65
M
30
Neurological
52.6
Y
20
M
31
Trauma
60.5
N
69
F
27
Respiratory
52.2
N
26.0
N
48
M
17
Gastrointestinal
(Operative)
40
M
18
Trauma
19.7
N
80
M
23
Respiratory
46.0
Y
70
F
30
Respiratory
70.0
N
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
measurements per day. These safety gains are introduced with
reduced clinical effort of 11.8 and 14.9 measures/day
compared to 16.1 for SPRINT, and with equivalent or higher
time in desired glycemic bands. Importantly, median nutrition
rates are raised by 32% (absolute) of ACCP goal rate. Thus,
all indications show STAR will provide global improvements
over SPRINT in performance, safety and effort when applied
clinically.
Figures 4 and 5 display major behavioral differences
between STAR and SPRINT. Each figure categorizes the
insulin/nutrition combination selected to display relative
frequency. Figure 4 indicates the preferred behavior of STAR
is to modulate insulin at a higher nutrition rate, while Figure 5
indicates SPRINT typically modulates nutrition at moderate,
SPRINT Clinical Data
% of Total Insulin Interventions
TABLE IV
STAR VS. SPRINT FULL COHORT SIMULATION RESULTS
STAR STAR 2SPRINT
Whole cohort statistics
Max
hourly
Data
# patients:
371
371
371
40101
40006
39841
Total hours:
hours
hours
hours
# BG measurements:
19634
24218
26646
BG measures per day
11.8
14.9
16.1
BG median
109
109
101
[IQR] (mg/dL):
[100 - 120]
[100 - 119]
[90.0 - 115]
Desired glycemic bands:
% BG in 80 - 125mg/dL
81.0
82.5
78.5
% BG in 80 - 145mg/dL
91.6
92.1
86.0
Hyperglycemic bands:
% BG in 145 - 180mg/dL
4.69
4.59
4.45
% BG > 180mg/dL
1.65
1.64
2.00
Safety glycemic bands:
% BG < 80mg/dL
2.06
1.69
7.83
% BG < 72mg/dL
0.83
0.70
2.89
# patients < 40mg/dL
4
3
14
Interventions:
Median insulin rate [IQR]
2.5
2.5
3.0
(U/hr):
[1.5 - 4.0]
[1.5 - 4.0]
[2.0 - 4.0]
Median glucose rate [IQR]
5.0
5.1
4.1
(g/hr):
[2.2 - 6.4]
[2.3 - 6.5]
[1.9 - 5.6]
Median glucose rate [IQR]
90.0
90.0
68.1
(% goal):
[30 - 100]
[30 - 100]
[30 - 85]
Treatment Summary:
3-hourly interventions:
44%
0%
2%
2-hourly interventions:
17%
64%
47%
1-hourly interventions:
39%
36%
51%
5
12
10
8
6
4
2
0
0
1
2
Insulin (U)
3
4
5
6
30
40
50
60
70
80
90
100
Nutrition (% ACCP Goal)
Fig.5. SPRINT Nutrition/Insulin Combination Frequency
relatively constant insulin administration levels. This
difference implies a clinical advantage when high nutrition
input is preferred. It equally highlights that insulin dosage
behavior is relative to carbohydrate input, showing importance
of accounting for nutrition in AGC, which many published
protocols do not [34]. Finally, and equally importantly, the
most common outcome in Figure 4 is 100% goal feed and 0
U/hr (11% of total interventions = 1 in 9), a safe and effective
outcome enabled by the STAR framework’s stochastic modelbased approach that was not possible with SPRINT.
Global and per-patient statistics from virtual trials of the
UNC, Yale and GluControl A protocols are presented with
STAR-Max and clinical SPRINT data in Table V. STAR
compares favorably against all cases, providing tight control
(lowest BG IQR) and safety (lowest incidence of severe
hypoglycemia) with the least clinical effort (11.8
measures/day). The per-patient analysis in Table IV indicates
the performance and safety benefits of STAR are realized over
a wide range of heterogeneous patients. Safety against
hypoglycemia, quantified by the median and IQR of %BG <
72mg/dL, shows a significant reduction against both the
clinical data and other AGC protocols. Median and IQR of the
Accurate Glycaemic Control with STAR
100
STAR
STAR Pilot Trial Results
STAR Virtual Trial Results
SPRINT Clinical Results
90
Cumulative Density (%)
% of Total Insulin Interventions
80
12
10
8
6
4
2
70
60
50
40
30
20
0
0
90
1
2
Insulin (U)
3
4
5
6
30
40
100
80
70
60
50
Nutrition (% ACCP Goal)
Fig.4. STAR Nutrition/Insulin Combination Frequency
10
0
0
20
40
60
80
100 120
Blood Glucose (mg/dL)
140
160
180
Fig.6. Cumulative density plot of current pilot trial results against
simulation results and SPRINT clinical data after 10 patients.
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
6
TABLE V
COMPARISON OF ALL FULL COHORT AND PER-PATIENT VIRTUAL TRIAL RESULTS AND CLINICAL SPRINT DATA
Whole cohort statistics
# patients:
Target Range (mg/dL)
# BG measurements:
BG measures per day
BG median [IQR] (mg/dL):
Desired glycemic bands:
% BG in 80 – 125mg/dL
% BG in 80 – 145mg/dL
Hyperglycemic bands:
% BG in 145 - 180mg/dL
% BG > 180mg/dL
Safety glycemic bands:
% BG < 80mg/dL
% BG < 72mg/dL
# patients < 40mg/dL
Interventions:
Median insulin rate [IQR] (U/hr):
Median glucose rate [IQR] (g/hr):
Median glucose rate [IQR] (% goal):
Per-patient Analysis
Median BG measurements [IQR]:
BG median [IQR] (mg/dL)
Median % BG in 80 - 125mg/dL [IQR]:
Median % BG in 80 - 145mg/dL [IQR]:
Median % BG < 72mg/dL [IQR]:
STAR-Max
371
[145 – 180]
19634
11.8
109 [100 - 120]
SPRINT Data
371
[72 – 117]
26646
16.1
101 [90.0 - 115]
UNC
371
110
24563
14.8
108 [98.5 - 122]
GluControl A
371
[80 – 110]
25049
15.1
105 [90.9 - 126]
Yale
371
[90 – 119]
36233
21.9
108 [96.0 - 122]
81.0
91.6
78.5
86.0
75.9
87.3
65.8
78.6
74.9
86.1
4.69
1.65
4.45
2.00
6.74
2.31
9.12
2.85
6.72
2.58
2.06
0.83
4
7.83
2.89
14
3.60
1.89
13
9.47
4.75
37
4.58
1.65
3
2.5[1.5 - 4.0]
5.0 [2.2 - 6.4]
90.0 [30 - 100]
3.0 [2.0 - 4.0]
4.1 [1.9 - 5.6]
68.1 [30 - 85]
2.6 [1.1 - 5.0]
5.2 [4.2 – 5.9]
80.0 [80 - 80]
4.0 [1.5 – 8.0]
5.2 [4.2 – 5.9]
80.0 [80 - 80]
3.0 [1.8 - 5.5]
5.2 [4.2 – 5.9]
80.0 [80 - 80]
29.0 [13.0 - 66.0]
110 [106- 119]
76.9 [61.1 - 85.4]
89.4 [78.9 - 95.1]
0.0 [0.0 - 1.0]
36.0 [16.0 - 92.5]
104 [95.0 - 113]
73.3 [58.6 - 83.8]
83.3 [69.6 - 91.8]
1.1 [0.0 - 4.6]
number of BG measures also indicate that the reduced clinical
effort of STAR will be realized across patients.
The simulation results are supported by the initial pilot trial
results presented in Figure 6 for the first 10 patients with the
target band of 80 (1486 hours, 836 measurements). The
clinical BG results in Table VI are very similar to simulation
results in Table V with 89.4% of BG within 80-145mg/dL and
78.2% BG within 80-125mg/dL. Safety has been maintained
with 1.54% of BG < 72mg/dL and no severe hypoglycemia
events (BG < 40mg/dL). Median BG was 109mg/dL [IQR:
99.5-122 mg/dL], which matches very closely with the
location and spread of BG in virtual trials.
TABLE V
STAR CLINICAL RESULTS
Clinical data:
# patients:
Total number of episodes
Total hours:
# BG measurements:
BG measures per day
80-125 mg/dL
80-145 mg/dL
145-180 mg/dL
% Measures:
> 180 mg/dL
< 80 mg/dL
<72 mg/dL
# patients below 40 mg/dL
Median insulin rate [IQR] (U/hr):
Median glucose rate [IQR] (g/hr):
Median glucose rate [IQR] (% goal):
Resampled hourly for comparison:
BG median [IQR] (mg/dL):
80-125 mg/dL
80-145 mg/dL
145-180 mg/dL
% BG:
> 180 mg/dL
< 80 mg/dL
<72 mg/dL
10
19
1486
836
13.5
69.1
83.9
8.97
4.19
3.83
2.03
0
3.0 [1.5 - 4.5]
4.87 [0.00 – 6.09]
80.0 [0.00 - 109]
109 [99.5 - 122]
78.2
89.4
5.63
2.48
2.48
1.54
34.0 [13.0 - 82.5]
110 [105 - 121]
72.0 [54.1 - 82.8]
85.3 [73.5 - 92.9]
0.0 [0.0 - 2.3]
34.5 [14.0 - 88.5]
106 [100 - 119]
61.0 [46.0 - 72.2]
75.0 [62.6 - 84.0]
4.3 [0.0 - 8.5]
53.5 [18.0 - 129.0]
110 [105 - 120]
69.2 [53.4 - 80.4]
83.8 [72.4 - 91.8]
0.7 [0.0 - 2.4]
IV. DISCUSSION
STAR virtual trials demonstrate safe, effective AGC in a
clinically applicable fashion. Significant safety improvements
would be likely compared to the current generation of BG
control in Christchurch Hospital (SPRINT), which was the
safest published protocol targeting BG ≤ 110mg/dL. Most
significantly, only 4 patients out of 371 (1.1% by patient and a
reduction of 71% from SPRINT) showed severe
hypoglycemia, and, in all cases, this hypoglycemia was
quickly resolved. Bulk cohort statistics were supported by the
per-patient analysis as shown in Table IV, with STAR
showing a more significant benefit than SPRINT and UNC,
the next best protocol, for individual patient safety from
hypoglycemia. Clinical applicability is strongly supported by
the initial pilot results, with both the performance and safety
benefits seen in virtual trials being realized in practice.
Characteristics of AGC varied between protocols. UNC
showed improved performance over SPRINT, while providing
a relatively light nursing workload. Safety was also improved
with regards to mild hypoglycemia, although the incidence of
severe hypoglycemia was increased. Although Yale showed
improved safety, the number of required measurements was
high. STAR provided the tightest and safest control, while
requiring the lowest nursing workload.
Importantly, all protocols showed relatively similar
glycemic control in desired bands, and results not overly
different from other clinical trials with ICU cohorts [3]. The
main differentials are in mild (%BG < 72mg/dL) and severe
hypoglycemia (number of patients < 40mg/dL). Equally, perpatient control, or consistency across inter-patient variability,
is better for SPRINT and STAR. These results match
observations that those trials that have not succeeded in
reducing mortality also had significantly increased
hypoglycemia [3], which is associated with worsened outcome
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
[4]. Hence, the success of SPRINT [8, 30] and the potential
success of STAR rests on the (unique) ability to tightly control
glycemia over a highly variable cohort and specific patients
with increased risk of hypoglycemia. Thus, the key
differentiator, seen here in these virtual trial results, may lie
with the ability to provide safe and tight glycemic control that
is robust to significant inter- and intra- patient variability,
which is what STAR was explicitly designed to achieve.
The relatively higher enteral nutrition rate for STAR
compared to SPRINT is likely to increase clinical acceptance.
Higher feed is generally preferred in many cases [31], despite
some recent contradictory evidence [10, 35]. However, STAR
can be easily adjusted by setting nutrition administration goals
to match any emerging evidence, and insulin rates will
automatically modify to maintain glycemic balance.
The comparison between STAR-Max and STAR 2-hourly
illustrates a known trade-off between measurement rate
(nursing workload) and patient safety. However, both provide
quality AGC. Thus, the main impact of measurement interval
in the STAR framework is on safety from intra-patient
variability.
Equally importantly, the ability of nursing staff to choose
between measurement interval options means that an informed
decision can be made at each BG reading. Thus, nurses selfmanage this workload. This choice or feature is expected to
also have a positive impact on compliance and acceptance.
A notable enhancement of the model-based approach of
STAR compared to the fixed approach in SPRINT is the
ability to change the desired target range and other factors or
limits. This ability has implications for the balance between
workload, safety and nutrition. For example, raising the target
BG range could provide increased nutrition intake at the
expense of higher BG. Providing a wider target band may
allow reduced workload with fewer BG measurements under
the target-to-range scheme presented, but may result in
increased glycemic variability within that band for which the
clinical outcomes are not fully known. These decisions can be
made by the attending clinical team based on their goals for a
particular patient and assessment of current evidence. Hence,
the framework can be directly and easily adopted and
generalized to any clinical culture and practice, unlike
previously published protocols.
More importantly perhaps, STAR can thus provide the
power to automatically balance the clinical treatment, based
on clinical goals for each patient. It thus avoids forcing a
clinic-wide target or approach onto patients who have
requirements outside the norm, or subjecting them to ad-hoc
decisions to handle their particular cases. Thus, for example,
different BG target ranges could be readily specified for
different clinical conditions based on diagnosis and this target
range could be updated as treatment progresses.
glycemia. Both overall cohort and per-patient findings support
the implementation of STAR in pilot clinical trials whose
initial results have further validated the in-silico approach in
this research. The model-based nature of STAR allows easy
adjustment of BG and nutrition to match emerging clinical
evidence, and permits individualization of treatment goals to
particular patient outcomes. Finally, the overall framework
presented is unique in its stochastic, risk-based approach, as
well as completely generalizable.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
V. CONCLUSIONS
Achieving AGC requires accounting for patient metabolic
variability while balancing safety and workload requirements.
In-silico results indicate the developed STAR algorithm
provides a safe and effective method for management of
7
[14]
[15]
J. S. Krinsley, "Association between hyperglycemia and increased
hospital mortality in a heterogeneous population of critically ill
patients," Mayo Clin Proc, vol. 78, pp. 1471-1478, Dec 2003.
K. C. McCowen, A. Malhotra, and B. R. Bistrian, "Stress-induced
hyperglycemia," Crit Care Clin, vol. 17, pp. 107-124, Jan 2001.
D. E. Griesdale, R. J. de Souza, R. M. van Dam, D. K. Heyland, D. J.
Cook, A. Malhotra, R. Dhaliwal, W. R. Henderson, D. R. Chittock, S.
Finfer, and D. Talmor, "Intensive insulin therapy and mortality among
critically ill patients: a meta-analysis including NICE-SUGAR study
data,"
CMAJ,
vol.
180,
pp.
821-7,
Apr
14
2009.
doi:10.1503/cmaj.090206.
S. Bagshaw, R. Bellomo, M. Jacka, M. Egi, G. Hart, C. George, and t.
A. C. M. Committee, "The impact of early hypoglycemia and blood
glucose variability on outcome in critical illness," Critical Care, vol. 13,
p. R91, 2009.
M. Egi, R. Bellomo, E. Stachowski, C. J. French, and G. Hart,
"Variability of blood glucose concentration and short-term mortality in
critically ill patients," Anesthesiology, vol. 105, pp. 244-52, Aug 2006.
J. S. Krinsley, "Glycemic variability: a strong independent predictor of
mortality in critically ill patients," Crit Care Med, vol. 36, pp. 3008-13,
Nov 2008.
J. S. Krinsley, "Effect of an intensive glucose management protocol on
the mortality of critically ill adult patients," Mayo Clin Proc, vol. 79, pp.
992-1000, Aug 2004.
J. G. Chase, G. Shaw, A. Le Compte, T. Lonergan, M. Willacy, X.-W.
Wong, J. Lin, T. Lotz, D. Lee, and C. Hann, "Implementation and
evaluation of the SPRINT protocol for tight glycaemic control in
critically ill patients: a clinical practice change," Critical Care, vol. 12,
p. R49, 2008.
G. Van den Berghe, P. Wouters, F. Weekers, C. Verwaest, F.
Bruyninckx, M. Schetz, D. Vlasselaers, P. Ferdinande, P. Lauwers, and
R. Bouillon, "Intensive insulin therapy in the critically ill patients," N
Engl J Med, vol. 345, pp. 1359-1367, Nov 8 2001.
M. P. Casaer, D. Mesotten, G. Hermans, P. J. Wouters, M. Schetz, G.
Meyfroidt, S. Van Cromphaut, C. Ingels, P. Meersseman, J. Muller, D.
Vlasselaers, Y. Debaveye, L. Desmet, J. Dubois, A. Van Assche, S.
Vanderheyden, A. Wilmer, and G. Van den Berghe, "Early versus late
parenteral nutrition in critically ill adults," N Engl J Med, vol. 365, pp.
506-17, Aug 11 2011. doi:10.1056/NEJMoa1102662.
F. M. Brunkhorst, C. Engel, F. Bloos, A. Meier-Hellmann, M. Ragaller,
N. Weiler, O. Moerer, M. Gruendling, M. Oppert, S. Grond, D. Olthoff,
U. Jaschinski, S. John, R. Rossaint, T. Welte, M. Schaefer, P. Kern, E.
Kuhnt, M. Kiehntopf, C. Hartog, C. Natanson, M. Loeffler, and K.
Reinhart, "Intensive insulin therapy and pentastarch resuscitation in
severe sepsis," N Engl J Med, vol. 358, pp. 125-39, Jan 10 2008.
doi:358/2/125 [pii]10.1056/NEJMoa070716.
S. Finfer, D. R. Chittock, S. Y. Su, D. Blair, D. Foster, V. Dhingra, R.
Bellomo, D. Cook, P. Dodek, W. R. Henderson, P. C. Hebert, S.
Heritier, D. K. Heyland, C. McArthur, E. McDonald, I. Mitchell, J. A.
Myburgh, R. Norton, J. Potter, B. G. Robinson, and J. J. Ronco,
"Intensive versus conventional glucose control in critically ill patients,"
N Engl J Med, vol. 360, pp. 1283-97, Mar 26 2009.
I. Mackenzie, S. Ingle, S. Zaidi, and S. Buczaski, "Tight glycaemic
control: a survey of intensive care practice in large English hospitals,"
Intensive Care Med, vol. 31, p. 1136, 2005.
D. Aragon, "Evaluation of nursing work effort and perceptions about
blood glucose testing in tight glycemic control," Am J Crit Care, vol. 15,
pp. 370-7, Jul 2006.
J. G. Chase, S. Andreassen, K. Jensen, and G. M. Shaw, "Impact of
Human Factors on Clinical Protocol Performance: A Proposed
> REPLACE THIS LINE WITH YOUR PAPER IDENTIFICATION NUMBER (DOUBLE-CLICK HERE TO EDIT) <
[16]
[17]
[18]
[19]
[20]
[21]
[22]
[23]
[24]
[25]
[26]
[27]
[28]
[29]
[30]
[31]
Assessment Framework and Case Examples," Journal of Diabetes
Science and Technology, vol. 2, pp. 409-416, May 2008 2008.
J. G. Chase, A. J. Le Compte, F. Suhaimi, G. M. Shaw, A. Lynn, J. Lin,
C. G. Pretty, N. Razak, J. D. Parente, C. E. Hann, J.-C. Preiser, and T.
Desaive, "Tight glycemic control in critical care - The leading role of
insulin sensitivity and patient variability: A review and model-based
analysis," Computer Methods and Programs in Biomedicine, vol. 102,
pp. 156-171, 2011. doi:10.1016/j.cmpb.2010.11.006.
E. S. Moghissi, M. T. Korytkowski, M. DiNardo, D. Einhorn, R.
Hellman, I. B. Hirsch, S. E. Inzucchi, F. Ismail-Beigi, M. S. Kirkman,
and G. E. Umpierrez, "American Association of Clinical
Endocrinologists and American Diabetes Association consensus
statement on inpatient glycemic control," Diabetes Care, vol. 32, pp.
1119-31, Jun 2009. doi:10.2337/dc09-9029.
F. Weekers, A. P. Giulietti, M. Michalaki, W. Coopmans, E. Van Herck,
C. Mathieu, and G. Van den Berghe, "Metabolic, endocrine, and
immune effects of stress hyperglycemia in a rabbit model of prolonged
critical illness," Endocrinology, vol. 144, pp. 5329-38, Dec 2003.
J. Lin, D. Lee, J. Chase, C. Hann, T. Lotz, and X. Wong, "Stochastic
Modelling of Insulin Sensitivity Variability in Critical Care,"
Biomedical Signal Processing & Control, vol. 1, pp. 229-242, 2006.
J. Lin, D. Lee, J. G. Chase, G. M. Shaw, A. Le Compte, T. Lotz, J.
Wong, T. Lonergan, and C. E. Hann, "Stochastic modelling of insulin
sensitivity and adaptive glycemic control for critical care," Computer
Methods and Programs in Biomedicine, vol. 89, pp. 141-152, Feb 2008.
X. W. Wong, I. Singh-Levett, L. J. Hollingsworth, G. M. Shaw, C. E.
Hann, T. Lotz, J. Lin, O. S. Wong, and J. G. Chase, "A novel, modelbased insulin and nutrition delivery controller for glycemic regulation in
critically ill patients," Diabetes Technol Ther, vol. 8, pp. 174-90, Apr
2006.
J. Lin, N. N. Razak, C. G. Pretty, A. Le Compte, P. Docherty, J. D.
Parente, G. M. Shaw, C. E. Hann, and J. Geoffrey Chase, "A
physiological Intensive Control Insulin-Nutrition-Glucose (ICING)
model validated in critically ill patients," Comput Methods Programs
Biomed,
vol.
102,
pp.
192-205,
May
2011.
doi:10.1016/j.cmpb.2010.12.008.
J. G. Chase, F. Suhaimi, S. Penning, J. C. Preiser, A. J. Le Compte, J.
Lin, C. G. Pretty, G. M. Shaw, K. T. Moorhead, and T. Desaive,
"Validation of a model-based virtual trials method for tight glycemic
control in intensive care," Biomed Eng Online, vol. 9, p. 84, 2010.
doi:10.1186/1475-925X-9-84.
C. E. Hann, J. G. Chase, J. Lin, T. Lotz, C. V. Doran, and G. M. Shaw,
"Integral-based parameter identification for long-term dynamic
verification of a glucose-insulin system model," Comput Methods
Programs Biomed, vol. 77, pp. 259-270, Mar 2005.
S. Penning, A. J. Le Compte, K. T. Moorhead, T. Desaive, P. Massion,
J. C. Preiser, G. M. Shaw, and J. G. Chase, "First pilot trial of the
STAR-Liege protocol for tight glycemic control in critically ill patients,"
Comput Methods Programs Biomed, Aug 30 2011. doi:S01692607(11)00191-X [pii]10.1016/j.cmpb.2011.07.003.
A. Evans, G. M. Shaw, A. Le Compte, C. S. Tan, L. Ward, J. Steel, C.
G. Pretty, L. Pfeifer, S. Penning, F. Suhaimi, M. Signal, T. Desaive, and
J. G. Chase, "Pilot proof of concept clinical trials of Stochastic Targeted
(STAR) glycemic control," Ann Intensive Care, vol. 1, p. 38, Sep 19
2011. doi:2110-5820-1-38 [pii]10.1186/2110-5820-1-38.
J. G. Chase, G. M. Shaw, J. Lin, C. V. Doran, C. Hann, T. Lotz, G. C.
Wake, and B. Broughton, "Targeted glycemic reduction in critical care
using closed-loop control," Diabetes Technol Ther, vol. 7, pp. 274-82,
Apr 2005.
T. Lonergan, A. LeCompte, M. Willacy, J. G. Chase, G. M. Shaw, X. W.
Wong, T. Lotz, J. Lin, and C. E. Hann, "A simple insulin-nutrition
protocol for tight glycemic control in critical illness: development and
protocol comparison," Diabetes Technol Ther, vol. 8, pp. 191-206, 2006.
F. B. Cerra, M. R. Benitez, G. L. Blackburn, R. S. Irwin, K. Jeejeebhoy,
D. P. Katz, S. K. Pingleton, J. Pomposelli, J. L. Rombeau, E. Shronts, R.
R. Wolfe, and G. P. Zaloga, "Applied nutrition in ICU patients. A
consensus statement of the American College of Chest Physicians,"
Chest, vol. 111, pp. 769-78, Mar 1997.
J. G. Chase, C. G. Pretty, L. Pfeifer, G. M. Shaw, J. C. Preiser, A. J. Le
Compte, J. Lin, D. Hewett, K. T. Moorhead, and T. Desaive, "Organ
failure and tight glycemic control in the SPRINT study," Crit Care, vol.
14, p. R154, 2010. doi:10.1186/cc9224.
C. Alberda, L. Gramlich, N. Jones, K. Jeejeebhoy, A. G. Day, R.
Dhaliwal, and D. K. Heyland, "The relationship between nutritional
intake and clinical outcomes in critically ill patients: results of an
[32]
[33]
[34]
[35]
8
international multicenter observational study," Intensive Care Med, vol.
35, pp. 1728-37, Oct 2009. doi:10.1007/s00134-009-1567-4.
S. S. Braithwaite, R. Edkins, K. L. Macgregor, E. S. Sredzienski, M.
Houston, B. Zarzaur, P. B. Rich, B. Benedetto, and E. J. Rutherford,
"Performance of a dose-defining insulin infusion protocol among trauma
service intensive care unit admissions," Diabetes Technol Ther, vol. 8,
pp. 476-88, Aug 2006.
P. A. Goldberg, M. D. Siegel, R. S. Sherwin, J. I. Halickman, M. Lee, V.
A. Bailey, S. L. Lee, J. D. Dziura, and S. E. Inzucchi, "Implementation
of a safe and effective insulin infusion protocol in a medical intensive
care unit," Diabetes Care, vol. 27, pp. 461-7, Feb 2004.
F. Suhaimi, A. Le Compte, J. C. Preiser, G. M. Shaw, P. Massion, R.
Radermecker, C. G. Pretty, J. Lin, T. Desaive, and J. G. Chase, "What
makes tight glycemic control tight? The impact of variability and
nutrition in two clinical studies," J Diabetes Sci Technol, vol. 4, pp. 28498, Mar 2010.
J. A. Krishnan, P. B. Parce, A. Martinez, G. B. Diette, and R. G. Brower,
"Caloric intake in medical ICU patients: consistency of care with
guidelines and relationship to clinical outcomes," Chest, vol. 124, pp.
297-305, Jul 2003.