Intensive Care Med (2009) 35:1505–1517

DOI 10.1007/s00134-009-1542-0 REVIEW

Saeid Eslami
Ameen Abu-Hanna
Tight glycemic control and computerized
Evert de Jonge decision-support systems: a systematic review
Nicolette F. de Keizer

Received: 17 November 2008 Abstract Objective: To identify All controlled studies reported on at
Accepted: 19 April 2009 and summarize characteristics of least one quality indicator of the
Published online: 27 June 2009 computerized decision-support sys- blood glucose regulatory process that
Ó The Author(s) 2009. This article is tems (CDSS) for tight glycemic was improved by introducing the
published with open access at CDSS. Nine out of ten controlled
Springerlink.com control (TGC) and to review their
effects on the quality of the TGC studies either did not report on the
process in critically ill patients. number of hypoglycemia events (one
Methods: We searched Medline study), or reported on no change (six
(1950–2008) and included studies on studies) or even a reduction in this
critically ill adult patients that repor- number (two studies).
ted original data from a clinical trial Conclusions: While most studies
or observational study with a main evaluating the effect of CDSS on the
objective of evaluating a given TGC quality of the TGC process found
protocol with a CDSS. improvement when evaluated on the
Results: Seventeen articles met the basis of the quality indicators used, it
inclusion criteria. Eleven out of sev- is impossible to define the exact suc-
S. Eslami ())
A. Abu-Hanna
enteen studies evaluated the effect of cess factors, because of simultaneous
N. F. de Keizer implementation of the CDSS with a
Department of Medical Informatics, a new TGC protocol that was intro-
Academic Medical Center, University duced simultaneously with a CDSS new or modified TGC protocol and
of Amsterdam, Meibergdreef 15, J1b-124, implementation. Most of the reported the hybrid solutions used to integrate
1105 AZ Amsterdam, The Netherlands CDSSs were stand-alone, were not the CDSS into the clinical workflow.
e-mail: [email protected] integrated in any other clinical infor-
Fax: ?31-20-6919840 mation systems and used the Keywords Systematic review

‘‘passive’’ mode requiring the clini- Tight glycemic control
Insulin

E. de Jonge Computerized decision-support
Department of Intensive Care, Leiden cian to ask for advice. Different
University Medical Center, implementation sites, target users, system
Critically ill patients

Albinusdreef 2, 2333 ZC Leiden, and time of advice were used, Critical care
The Netherlands depending on local circumstances.

Introduction however, requires additional efforts, especially from the

nursing staff who are expected to adhere to the TGC
Blood glucose control aiming at normoglycemia (i.e., protocol. Such protocols are often complex, requiring
blood glucose levels 80–110 mg/dl), frequently referred specific timing of blood glucose level (BGL) measure-
to as ‘‘tight glycemic control’’, has been shown to reduce ments, and the availability of patient-specific data. One
mortality and morbidity of critically ill patients [1, 2] way to support nurses in adhering to protocols is by
although its implementation bears the risk of hypoglyce- applying information technology, in particular clinical
mia [1, 2]. Applying tight glycemic control (TGC), computerized decision-support systems (CDSSs).
1506

In general, a CDSS is a computer program that is

intended to help healthcare workers in making decisions
[3]. A CDSS can be characterized by the level of sup-
port, the consultation mode, and the communication
style. The level of support ranges from general to
patient-specific: e.g., from merely displaying the proto-
col chart to suggesting the specific amount of insulin to
be administered. Some systems are passive, providing
advice only on demand, whereas others are active,
providing feedback to the healthcare worker without
being asked for it. Finally, regardless of the level of
support and the consultation mode, a CDSS may operate
in two communication styles: in the critiquing mode the
system provides advice which is dependent on the
adherence of clinical practice to a protocol (e.g., noti-
fying the nurse that a BGL measurement has been
expected but not performed), in the non-critiquing mode
it provides advice regardless of whether a protocol is
followed or not [4]. Notwithstanding the potential ben-
efits of CDSSs, issues pertaining to their design,
implementation, evaluation, and to critical success fac-
tors, are largely still open. Glucose regulation is
specifically interesting as it forms an application in
Fig. 1 The search strategies applied and the search flowchart. The
which a CDSS is applied in a highly controlled clinical bold terms are MeSH terms
practice. Therefore lessons learned from glucose regu-
lation CDSSs can be used for other highly controlled
clinical practices, for example mechanical ventilation
resolved by consensus involving a third reviewer. Arti-
and blood-pressure control.
cles were selected if they reported original data from a
The objective of this systematic review is to identify
clinical trial or observational study on critically ill adult
and summarize characteristics of CDSSs and to review
patients and only if one of their main objectives con-
their effects on the quality of the TGC process in pub-
cerned the evaluation of a given TGC protocol with a
lished studies on the use of CDSS in TGC for critically ill
CDSS. A study was included if the TGC protocol
patients.
implied an upper normoglycemia limit of, at most,
150 mg/dl. Opinion papers, surveys, and letters were
excluded.
Methods From the selected papers, the same two reviewers
extracted data on the following items:
We searched for relevant English language articles based 1 method and study design aspects;
on keywords in title, abstract and MeSH terms, using 2 CDSS characteristics;
Ovid Medline and Ovid Medline In-Process (1950 to 3 TGC quality indicators in terms of their definition and
December 31, 2008). The final literature search was applicability; and
performed on January 11, 2008. 4 data on the effect of the CDSS on the quality of the
Figure 1 shows the applied search strategies and the TGC process or mortality or morbidity.
corresponding search flowchart. In the first stage (A),
we searched for terms related to glucose and insulin. In A quality indicator was defined as a measurable
the second stage (B) we limited the search using the quantity of the TGC process that may, alone or in com-
terms ‘‘critical illness’’ or ‘‘critical care’’ or ‘‘intensive bination with other quantities, indicate some aspect of its
care’’. In the third stage (C), keywords and MeSH quality. Discrepancies between the two reviewers were
terms referring to a decision-support system were again resolved by consensus after involving the same
searched. The results of these three stages were com- third reviewer.
bined using the Boolean operator ‘‘AND’’. Searching To obtain insight into the heterogeneous nature of
was supplemented by scanning bibliographies from these evaluation studies and according to the hierarchy of
identified articles. study designs developed by the University of California
Two reviewers independently examined all titles and San Francisco Stanford Evidence-Based Practice Center
abstracts. Discrepancies among the two reviewers were we classified the studies into:
 1507

– randomized controlled trial (level I), Support for insulin pump speed was ‘‘active’’ in one study
– non-randomized controlled trial (level II), only; the others used the ‘‘passive’’ mode for pump
– observational study with controls (level III), and adjustments requiring the clinician to ask for advice.
– observational study without controls (level IV) [5, 6]. In five studies the protocols were based on ‘‘if–then’’
statements on a sliding scale. They contained a list of
simple rules, with a condition (the ‘‘if’’ part) and a con-
clusion part (the ‘‘then’’ part). The condition was based on
Results the value of the current BGL measurement. The conclusion
specified the corresponding insulin amount (or insulin
Searching the online databases resulted in 70 articles. Ini- pump speed) and the time interval until the next BGL
tial screening of titles and abstracts rendered 21 articles measurement. In the other 12 studies the protocols were
eligible for further full-text review. One additional article formula-based. Formula-based protocols rely on an famil-
was identified by reviewing bibliographies, yielding a total iar and simple equation: insulin dose/hour = [BGL - 60]
of 22 articles. Based on the full text review, five studies 9 multiplier (insulin sensitivity) [7], which was sometimes
were excluded because they turned out not to address a adapted. In this group of studies, a default number for the
computerized system, leaving 17 articles for detailed multiplier was considered as the starting point for TGC.
analysis. Based on the latest BGL values and a statistical model, the
Table 1 lists the materials (included patients and study multiplier was recalculated and these values were used to
locations), TGC target range, design, and results of these calculate the required next pump speed.
17 studies. Table 2 summarizes the CDSSs’ characteris- In most studies (14 out of 17), users manually entered
tics. Table 3 reports all the used quality indicators of the the BGL values and pump speeds into a separate CDSS
TGC process. database, because these data were not electronically
available or they were not connected to the CDSS. In
three other studies these data were electronically retrieved
Methods and study design aspects from a laboratory or a hospital information system.

There were three randomized clinical trials, seven con-

trolled before/after trials, and seven observational studies. Quality indicators
In two studies only information was given on how often the
CDSS was used, in one of these studies this information Twenty-four different indicators of glycemic control were
was only given for the three months after the study period. extracted (Table 3). Hypoglycemia-related indicators were
In only one study the same TGC protocol was used used in 14/17 studies as a proxy for safety. Six different
before and after the implementation of CDSS. In five thresholds varying between 40 and 70 mg/dl were used to
other studies a modified version of the TGC protocol was define a hypoglycemia event. The most often used indi-
used after the implementation of the CDSS. Two of these cators were BGL summaries such as mean or median BGL
five studies evaluated the same CDSS and the protocol (15/17 studies), the number of measurements in a prede-
used, but in another setting. Eleven out of seventeen fined target range (10/17), the frequency of BGL
studies evaluated the effect of a new TGC protocol measurements (9/17), the time needed to reach the defined
introduced at the same time with the implementation of BGL target (8/17), the time spent in the predefined BGL
the CDSS. Three of these eleven studies also evaluated range (7/17), or compliance to protocol (6/17). Hypergly-
the same CDSS and the protocol in different settings. cemia-related indicators (number of hyperglycemia events
and hyperglycemia index) (6/17) were other frequently
used indicators. Because of different target ranges, com-
CDSS characteristics paring these indicators is difficult. Lower limits of target
ranges varied from 80 to 100 mg/dl. Similarly, the upper
Most of the reported CDSSs (14 out of 17) were stand-alone limits of target ranges varied from 110 to 150 mg/dl.
and were not integrated in other clinical information sys-
tems, for example computerized physician order entry
systems, patient data management systems, or intelligent Effect of the CDSS on the quality of the TGC process
pumps. CDSSs in all studies were ‘‘patient specific’’ and
operated in the ‘‘critiquing’’ mode. This means that deci- All controlled studies reported on at least one quality
sion support was given if clinical practice for a patient was indicator that was improved by introducing the CDSS
not according to the protocol. Reminders on the time of the (with or without the new protocol). Among controlled
next BGL measurement were ‘‘active’’ in nine out of sev- studies, one study reported that the number of hypogly-
enteen studies. ‘‘Active’’ means that the users automatically cemia events increased, but without mentioning whether
received the reminder without manually asking for it. this increase was statistically significant. Among other
Table 1 Description of study design, TGC target range, and results of the 17 included papers on DSS for TGC implementation 1508

Level Ref. Materials Target range Design Result

and monitoring
interval

I [23] 50 patients, nine-bed medical TR: 80–110 Prospective, five months, Median BGL and HGI were significantly lower in the CDSS group [BGL 106; HGI 7.2] than
ICU, 72 h of ICU stay M: NM RCT in control patients [BGL 133, p \ 0.001; HGI 29, p \ 0.001]. One hypoglycemic episode
was detected in the CDSS group but none in the control group. Sampling interval was
significantly shorter in the CDSS group [117 min vs. 174 min; p \ 0.001]. Thirty out of
thirty-four nurses answered the question of whether the algorithm could be applied in daily
routine in the affirmative.
[24] 40 patients with diabetes TR: 90–150 Prospective, RCT The mean BGL (147 vs. 126, p \ 0.01) and the mean time to capture range (171 min vs.
mellitus and scheduled M: 15 min–4 h 40 min, p \ 0.001) decreased during the ICU stay. Patients in the CDSS group spent more
cardiac surgery, time in the desired range during both the intraoperative phase (49 vs. 27%, p \ 0.001) and
Cardiothoracic ICU. the ICU phase (84 vs. 60%, p \ 0.0001). There were no statistical differences between
Operation time and first groups in the number of hypoglycemia episodes.
9 h in ICU
[8] 484 patients, 18-bed TR: 72–126 Prospective, 10–6–10– In the paper-based period, 29% of the samples occurred with optimal timing; during the
medical-surgical IC, M: 15 min–3 h 4 weeks, before protocol– second period, this increased to 35.5% for paper-based and to 40.2% for computerized
C24 h in ICU paper-based–randomized protocols. In the third study period timeliness scores reverted to the first period rates. The
between paper-based or same occurred for late sampling and insulin dose compliance. For the second study period,
CDSS–paper-based the time that a patient’s BGL fell within the target range improved for both the control
protocol (52.9%) and computerized (54.2%) groups compared with the first study period (44.3%),
the third period (42.3%), and before TGC (22%).
II [9] 552 trauma admissions, TR: 80–110 Retrospective, Median number of BGLs was 18 in control vs. 12 (4–54) in intervention group (p = 0.27).
31-bed ICU M: 2 h in manual and 4 ? 3 months, Mean BGL was lower (116 vs. 12; p \ 0.001), and median BGL measurement interval
1–2 h in computerized before/after decreased (2.79 vs. 2.14; p \ 0.001). Percentage of range BGL increased (41.8 vs. 34.0%;
protocol p \ 0.001), hyperglycemia and hypoglycemia frequency decreased (12.8 vs. 15.1% and 0.2
vs. 0.5%; both p \ 0.001), and within 12 and 24 h more patients reached normalization
(69.7 vs. 62.1%; p = 0.47 and 79.8 vs. 77.7%; p = 0.88). For post-initiation, computerized
protocol entries compliance to protocol’s recommendations was 98%.
[22] 1,080 patients, four ICUs TR: 80–110 Prospective, 22 months, Compliance with eProtocol-insulin advice was 91–98% among the four ICUs. Compared with
M: 15 min–2 h in paper- controlled the simple guideline, eProtocol-insulin BGLs in target increased from 21 to 39%
based protocol and (p \ 0.001) and mean BGL decreased from 142 to 115 (p = 0.001). Number of
1–4 h in computerized measurements and patients with at least one hypoglycemic event did not change
significantly.
[10] 2,398 patients, progressive TR: 80–110 Retrospective, 17 months, 61% of BGL were in target range, mean BGL was 106, and hypoglycemia frequency was
care unit and mixed M: 15 min–2 h Observational, 3 ? 3 0.4%. After a hypoglycemia event the mean interval until next measurement was 26 min,
medical-surgical ICU. before/after and the mean next BGL measurement was 106. Achieving the target range (with mean
BGL of 98) required 6.9 h. Percentage of measurements \110 increased from 32 to 52%
due to intervention (p \ 0.001). The frequency of hypoglycemia decreased from 0.5 to
0.4%.
[25] 97 coronary artery bypass TR: 80–120 Retrospective, There was a decrease in mean BGL (in first 48 h) from 154 (only type 2 diabetic patients
graft patients, six-bed M: 15 min–2 h 5 ? 5 months, before/ included) to 118 in type 2 diabetic patients and 116 in non-diabetic patients (p \ 0.0001),
Cardiothoracic ICU, first after in percentage not in target range within 48 h (26 vs. 4.6% and 3.0%, p \ 0.0001), in mean
48 h after surgery time to capture the target range (22.1 vs. 8.7 h and 5.9 h, p \ 0.0001) and in
hyperglycemia index (41.34 vs. 12.97 and 8.46, p \ 0.0001). The hypoglycemia rate was
not significantly increased after the intervention (1.14 vs. 1.42% and 1.94%, p = 0.26).
[26] 129 ? 128 trauma patients, TR: 80–130 Retrospective, 6 ? 6 months Significant reduction in mean BGL and total infection in all LOS categories were
C72 h in ICU M: not mentioned before/after demonstrated after intervention. However, case mix adjusted mortality was significantly
higher after the intervention. The percentage of hypoglycemic patients did not change (31
vs. 32%).
[15] 351 patients, 21-bed surgical TR: 80–110 Retrospective, 32 ? 49 days The computer-based protocol reduced time from first BGL measurement to initiation of
ICU, first five days in IC M: 1 h in manual protocol before/after insulin protocol, improved the percentage of in-range BGL (29.3 vs. 37.7%; p = 0.006),
and 1–2 h in and patients on IIT for C24 h were on average 116 min more in target range (p = 0.029).
computerized protocol The overall mean BGL for the first 5 days of IC admission were non-significantly lower in
post-intervention group (129 vs. 134). Hypoglycemia was rare in both groups (0.2% of
measurements).
[11] 891 patients, 16-bed ICU, TR: 97–128 Retrospective, 9 ? 15 ? 5 The mean BGL decreased from 131 before intervention to 119 thereafter and then to 112 after
at least 24 h in IC M: 30 min–4 h months before/after the first revision. The proportion of values \144 increased from 69 to 81% and then to
89%. One episode of hypoglycemia was observed before the intervention, 13 after
intervention and then six after revision. The hit counter added to the calculator after the end
of study showed the monthly use was 1,175 hits on average.
Table 1 continued
Level Ref. Materials Target range Design Result
and monitoring
interval

IV [27] 661 patients, 16-bed ICU TR: 97–128 Prospective, 12 months, Mean BGL was 121. There were 34 episodes of treated hypoglycemia 11 of which were on an
M: 30 min–4 h observational IIT. There were two troughs in the time of data entry that corresponded with staff handover.
There was no evidence of diurnal variation in BGL.
[17] 2,800 patients, three ICUs TR: 72–135 Prospective, 30 months, Patients were on GRIP-ordered pump rates 97% of time. Median measurement time was
(surgical, neurosurgical M: 30 min–12 h observational 5 min late (IQR 20 min early to 34 min late). Hypoglycemia was uncommon (7% of
and cardiothoracic). patients for \63; 0.086% for \40). Median time to capture target range was 5.6 h (0.2–
11.8) and maintained for 89% (70–100) of the remaining ICU stay. The glucose variability
was 22 (14.4–31.5). The frequency of measurements was 5.9 (4.8–7.3) per patient-day or
once every 245 min.
[28] 179, patients, ten-bed mixed TR: 81–135 Prospective, 3 months, Mean BGL decreased from 166 without protocol to 138 with the final protocol implemented
medical-surgical ICU M: 30 min–24 h observational study by a CDSS. BGLs were measured a total of 1,854 times in 179 ICU admissions during 553
ICU treatment days. The median BGL was 126, and 53.1% of BGLs were within the target
range. One episode of hypoglycemia occurred (0.5% patients or 0.05% measurements).
[29] 50 patients, 22-bed ICU, TR: 80–110 Prospective, 6 month, Median time percentage in target range was 23%. BGL was 50% of time in the range 112–
C24 h on mechanical M: 15 min–4 h observational study 144. Median proportion of time spent in hyperglycemic range 180–200 and[201 was 2 and
ventilation 1.4%. There were 28 hyperglycemic episodes including those of 15 (30%) patients treated
with IIT protocol. Median time to capture the target range was 10.5 h. Median of 47% of
sampling were not taken within the time frame stated in the protocol. Hypoglycemia
occurred 14 times and in five (10%) patients.
[7] 5,080 IIT run over TR: 100–140 NM, observational The mean BGL reached \150 in 3 h. The prevalence of hypoglycemia was 2.6% among all
120,683 h IIT. M: 20 min–2 h runs. All episodes of hypoglycemia were recognized within 20 min. The mean of all BGL
\60 was 49 and the follow up value in an average of 33 min was 83 (p \ 0.001). No
clinical symptoms due to hypoglycemia were reported.
[16] 179 patients, 12-bed TR: 72–135 Prospective, 4 months, Severe and mild hypoglycemia rates were 0.6%, and 11.2%. In patients staying[24 h, time to
surgical ICU. M: 30 min–12 h observational capture target was 5.7 h and hyperglycemia Index was 17.3 mg/dl. Median time percentage
for BGL in target range was 78% with median BGL of 88. Mean BGL change in the first
24 h was -21. After 24 h, mean BGL was 121. Nurses rated program as easy to work with
and as an improvement over the paper protocol.
[21] NM, surgical ICU TR: 100–150 NM, observational There was a significant improvement in glycemic control. The benefits to the participating
M: 20 min–2 h institution were increased awareness of the importance of improving BG control, decreased
calls to physicians, reduced need for sliding-scale insulin injections, increased nursing
satisfaction because of fewer calls to physicians for insulin adjustment, and increased
physician satisfaction because of improved BG control and fewer interruptive phone calls.

TR, target range; M, monitoring; NM, not mentioned

Unit of all BGL thresholds is mg/dl
1509
Table 2 Characteristics of CDSS
1510

Ref. Type of Type of Consultation Users Evaluation of Evaluation of Short description of CDSS
protocol system mode for: before/after usage, usability
pump/next new system and satisfaction
measurement

[23] Model-based Stand-alone Passive/ Nurse/Nurse CDSS ? modified Usability The system (called eMPC) runs on a bedside laptop
formula active protocol computer.
The model adapts itself to the input–output relationship
observed during TGC, i.e., an incoming glucose
measurement is used by the model to update insulin
resistance taking into account previously administered
insulin, parenteral and enteral glucose.
At start-up, the algorithm requires the patient’s ID and
body weight. The input of the BGL measurement is the
first step of a ‘‘wizard’’, subsequently querying the
current status of enteral and parenteral glucose
administration. Finally, any advice regarding the insulin
infusion rate is given.
Profiles of glucose level, insulin rate and the amount of
carbohydrates infused by enteral and parenteral
administration are displayed on the screen.
A countdown timer signals the time until the next glucose
measurement.
When the enteral/parenteral carbohydrate infusion rate is
changed then eMPC suggests a new insulin infusion rate
without requiring a new BG measurement.
[24] Model-based Stand-alone NM NM CDSS ? modified No This software (called EndoTool) upregulates and
formula protocol downregulates a quadratic insulin-dosing relationship
based on the entered BG readings from a point-of-care
device. It uses engineering-control mathematics that
consider the previous four dose responses to regulate
the dosing relationship.
System recommends the insulin dose, glucose
determination frequency, and a 50% dextrose dose
(when appropriate) for hypoglycemia.
How the necessary data should be entered (manually or
electronically) was not mentioned.
[8] ‘‘If–then’’ Implemented Active/ Physicians/ CDSS No The system extracted the pump rate and BGLs from
statements in PDMS active physicians PDMS, and suggested the pump rate and the next BGL
measurement time.
It reminded when the next BGL measurement should be
done according to the protocol. If the staff took no
action, the message would pop up again within a few
minutes.
When the patient record was activated (bedside computer
or any other workstation) the pop up windows would be
shown.
Table 2 continued
Ref. Type of Type of Consultation Users Evaluation of Evaluation of Short description of CDSS
protocol system mode for: before/after usage, usability
pump/next new system and satisfaction
measurement

[9] Model-based Implemented Passive/ Nurses/initiation by CDSS ? modified No Physician could initiate the CPOE-based insulin protocol,
formula in CPOE passive physicians and protocol which the patient’s nurse would carry out. Physicians
continuation by should enter current BGL and target high and low
nurses limits.
At initiation, a highlighted prompt reminds physicians to
provide a dextrose source to prevent hypoglycemia.
After verifying protocol the CPOE system generates
corresponding orders for physician verification, and
instructs the nurse to perform subsequent BGL testing.
Nurses enter new BGL into the system, and adjust insulin
drip rates based on the protocol provided. The nurse can
override system pump suggestions and enter it manually
based on his/her own decision. It included optional
instructions for nurses to notify physicians about out-of-
range values.
BGLs below target threshold generate an order for
intravenous dextrose dose; simultaneously, intravenous
insulin infusion is withheld for 1 h.
[22] Model-based Stand-alone Passive/ NM/mainly CDSS ? modified Kind of usability and Installed on stand-alone laptops. The laptops were
formula active nurses but also protocol satisfaction were distributed to clinical areas when patient care with
physicians mentioned (not e-Protocol insulin was started.
formal) Only in one ICU BGLs were automatically retrieved from
the hospital electronic medical records. In other three
ICUs, nurses manually entered the recent BGL and
patient weight. Initial infusion rate was weight-
dependent. Thereafter, the current infusion rate, the
difference between the most recent entered BGL, the
target, and the rate of change of BGL subsequently
determined the insulin infusion rate.
The e-Protocol-insulin bedside computer displayed the
time remaining to the next BGL measurement.
Clicking the electronic protocol screen signified an ‘‘intent
to’’ accept the recommendation. The nurse still needed
to verify drug administration through their usual
medical record documentation. When a
recommendation was declined, reason was recorded.
1511
Table 2 continued
1512

Ref. Type of Type of Consultation Users Evaluation of Evaluation of Short description of CDSS
protocol system mode for: before/after usage, usability
pump/next new system and satisfaction
measurement

[10] Model-based Stand-alone Passive/ NM/nurses CDSS ? new No Target range can be selected. It is menu-driven, and
formula active protocol includes options for starting a new insulin drip,
stopping/holding a drip, resuming a prior drip, entering
a BG value, specifying the initial insulin sensitivity
factor (ISF).
First BG should be entered manually. Then the system
calculates the initial pump rate and the amount of time
until next BG measurement. Time and ISF are adjusted
whenever a new BG is entered. Based on current and
previous BGs, the new rate and next BG measurement
time is announced by the program.
At the scheduled time for the next measurement, the
program sounds an alarm to remind nurse.
[25] Model-based Stand-alone, Passive/ Nurses/nurses CDSS ? new No A preprinted order was set on the web site that provided a
formula web-based passive protocol direct link to an on-line calculator.
Current, last BGL, and first multiplier should be entered
manually. Therefore the computerized system
calculates a new multiplier and pump rate.
A list of patient non-specific recommendations related to
the protocol was given by a computer interface without
mentioning which one is more important.
[26] Model-based Stand-alone Passive/ Physicians/ CDSS ? new No The system (called Glucommander) was used in
formula active physicians protocol combination with, but independently of, an insulin
infusion pump and a glucometer.
Nurse should enter the first multiplier and the BGLs
manually. Except for the first multiplier, other
multiplier and the next time for new BGL test was
calculated automatically.
Time for new BGL test was calculated for use by system
reminders.
[15] Model-based Implemented Passive/ Nurses/initiation by CDSS ? modified No usage but not Same system as mentioned in Ref. [9].
formula in CPOE passive physicians and protocol formal usability
continuation by and satisfaction
nurses evaluation
[11] ‘‘If–then’’ Stand-alone, Passive/ Medical staff, CDSS ? new Usage for three A nurse inputs current and last BGL and current pump rate
statements web-based passive pharmacists, and protocol months after end into the bedside computer. Based on entered BGL and
nurses/nurses of study was last BGL new pump rate was calculated and an interval
mentioned for the next test was also suggested.
A list of recommendations related to the protocol was
given by the computer interface.
[27] ‘‘If–then’’ Stand-alone, Passive/ Nurses CDSS ? new Usage was A slightly modified version of the protocol and web-based
statements web-based passive protocol mentioned calculator mentioned in Ref. [11] were used in this
study. The system was modified to record entered
insulin, BGL, and patient’s bed number, and the date
and time of calculation.
Table 2 continued
Ref. Type of Type of Consultation Users Evaluation of Evaluation of Short description of CDSS
protocol system mode for: before/after usage, usability
pump/next new system and satisfaction
measurement

[17] Model-based Stand-alone Passive/ NM/nurses CDSS ? new Only usability and Slightly modified version of mentioned protocol and
formula active protocol satisfaction were CDSS in Ref. [16] was used in this study. More safety
evaluated features such as label printing and other features
rendering the entire procedure paperless were added.
[28] ‘‘If–then’’ Stand-alone Passive/ Physicians/nurses CDSS ? new No Current, last BGL, current insulin rate and hourly rate of
statements passive protocol feeding should be entered manually. Therefore the
computerized system calculates needed bolus insulin,
pump rate, and next measurement interval.
A list of patient-non-specific recommendations related to
the protocol was given by a computer interface without
mentioning which is most important.
[29] ‘‘If–then’’ Stand-alone Passive/ Nurses CDSS ? new No A nurse inputs BGL and current pump rate into the system
statements passive protocol in the bedside computer. Based on entered BGL and last
BGL a new pump rate was suggested by the system.
50% tolerance for delay in BGL test was accepted because
the nurses manually entered the BGLs into the system.
[7] Model-based Stand-alone Passive/ NM/physicians and CDSS ? new No Same system as mentioned in Ref. [26] but now system is
formula active nurses protocol connected to the pump and can adjust it directly.
[16] Model-based Stand-alone Passive/ NM/nurses CDSS ? new Only usability and The system (called GRIP) extracts new BGLs from the
formula active protocol satisfaction were hospital information system. Nurses should validate
evaluated these BGL values and also could change them. Nurses
should enter current insulin pump rates, intravenous
glucose dose, related drug doses, etc.
Nurses can see the new rate suggestion and can accept or
change it.
Time for next BGL test is used in system reminders.
[21] Model-based Stand-alone Passive/ NM physicians and CDSS ? new Not formal Same system as mentioned in Refs. [7] and [26].
formula active nurses protocol satisfaction
evaluation
NM not mentioned
1513
1514

Table 3 List of quality indicators used

Indicator Measurement description Refs.

Blood glucose levels Represented as mean and/or median BGL values. Each BGL Fifteen articles [9–11,
over a time period measurement or each patient was considered as the unit of 15–17, 21–29]
observation.
Mean or median BGL was also calculated in the morning (6:00–12:00,
morning BGL) [11], after the target range was achieved [10], after
24 h [16, 17], at different time of a day [27], or at starting TGC
[29].
Measurements in Represented as the number or percentage of measurements in a Ten articles [9–11, 15,
predefined predefined BGL range during the study, after the target is achieved 17, 22, 24, 25, 27, 28]
blood glucose ranges [10] or among early, on time, or late measurements (based on
protocol) [17].
Each BGL measurement was considered as the unit of observation.
Frequency of Represented as: mean or median per patient [9, 22, 25, 28] or per Nine articles [9, 15–17,
measurements during patient treatment day [15–17, 28]; mean sampling interval (time) [9, 22–25, 28, 29]
the study 17, 23, 29]; frequency overall [24] per day [15]; and mean number
of measurements before first in-range BGL [22]
Time to capture defined Represented as mean and median of time or by Kaplan–Meier curve Eight articles [7, 9, 10,
blood glucose target [10, 25] 16, 17, 22, 24, 25]
Time in predefined range Represented as mean of percentage of time per patient [15, 24], Seven articles [8, 15–17,
median of percentage of time per patient [16, 17, 29], or 24, 25, 29]
cumulatively for all patients [8, 17, 25]
Protocol compliance A comparison of measurement times suggested by protocol to actual Six articles [8, 9, 17, 22,
times of measurements and/or pump speed suggested by protocol to 23, 29]
actual pump speed during TGC or at the time of hypoglycemia
events.
Hyperglycemic index Represented as median area between glucose–time curve and upper Four articles [16, 17, 23,
normal range divided by time per patient during the trial. Upper 25]
ranges were 110 [23], 117 [16], 120 [25], and 135 [17].
Hyperglycemia events Represented as: percentage of measurements[180a [25] or percentage Three articles [9, 25, 29]
of time [150 [9, 25, 29] and percentage of measurements and
patients with at least one BGL [180 for more than 2 h [29]
Number of patients who Represented as number and percentage of patients who achieved the Two articles [9, 22]
achieved predefined target range overall [22], within 12 and 24 h after starting IIT [9].
range
BGL change over time Represented as BGL change in first 24 h. One article [16]
BGL variability Represented as standard deviation of all measurements per patient. One article [17]
Number of patients who Represented as number and percentage. One article [25]
did not achieve
predefined range
Odds ratio of achieving Per additional IC day and per used drugs. One article [11]
certain BGL
Time until first BGL Represented as mean of time [16]. One article [15, 16]
Time until starting IIT Proportion of patients per time [15]. One article [15, 16]
Hypoglycemia-related indicator (proxy for safety)
Hypoglycemia events \40 [7, 9, 15, 22, 23, 25, 27], B40 [9, 22, 25, 28], \50 [7, 10, 11, 26], Twelve articles [7, 9–11,
\60 [7, 24], and \70 [25] were used as thresholds for defining a 15, 22–28]
BGL as hypoglycemia event.
Severe or marked \40 [16, 17] and B40 [15, 16, 29] were used to define a BGL as Five articles [7, 15–17,
hypoglycemia events severe hypoglycemia event. In one study clinical findings defined 29]
severe hypoglycemia [7]
Mild or moderate \63 was used define a BGL as a mild hypoglycemia event. Two article [16, 17]
hypoglycemia events
Next BGL after Represented as mean BGL. Two articles [7, 10]
hypoglycemia
Time until next in Represent as mean time. Two articles [7, 10]
predefined range after
hypoglycemia
Need for glucose injection \40 have to inject and \72 should be considered. One article [27]
Time until hypoglycemia Represented as maximum time until hypoglycemia recognition. One article [7]
recognition
Time until next BGL after Represented as mean time. One article [10]
hypoglycemia
a
Unit of all BGL thresholds is mg/dl
 1515

controlled studies one study did not report [8], six studies Therefore, it is difficult to conclude that the CDSS itself
reported no change, and two studies [9, 10] even reported was the causative reason for observed improvements in
a reduction in the number of hypoglycemia events. Seven glucose regulation. The fact that in some studies the users
observational studies also reported that the number of differed before and after the introduction of the CDSS
hypoglycemia events was in an acceptable range. No (based on a new protocol) further hampers identification
clinical symptoms attributable to hypoglycemia were of success factors for improved glucose regulation. Future
reported in any of the studies. The observational studies research should carefully choose a study design in order
also used efficiency-related indicators, for example mean to clearly separate the effect of the TGC protocol and the
BGL or time to capture TGC in the acceptable range by contribution of the CDSS.
introducing the CDSS (with or without a new protocol).

CDSS characteristics

Discussion and recommendations To achieve the optimum effect of TGC, the protocol should
be integrated into clinical workflows [14]. It should put the
In this review we have summarized the design, charac- knowledge for clinical decision-making at the point of care.
teristics, indicators, results, and limitations of 17 The question is, however, what is the best place and time for
published studies on glucose regulation CDSSs in inten- presenting the knowledge to the user. Boord et al. [15] used
sive care. Although most studies reported a positive effect the computerized physician order entry (CPOE) as the
on at least one quality indicator the diversity of studies in starting point and integrated the CDSS into this. They
term of case-mix, insulin therapy, associated therapies, believe this provided a ‘‘one stop shop’’ by capturing BGL
and indicators used (varying in their definition and the and other patient data from the clinician at the bedside,
ways of calculating them) severely hamper comparison of generating new orders, and logging the data into the elec-
the studies. Therefore, although meta-analysis is theoret- tronic medical record. Clinicians should first initiate an
ically possible, the results will not be reliable. In addition order in the CPOE before the CDSS could generate advice.
only three papers reported on an RCT. Our search Both pump adjustment and next measurement time
included all synonyms known to us of TGC, for example reminders were passive. In contrast, Vogelzang et al. [16,
‘‘intensive insulin therapy’’ (IIT). A limitation of our 17] believe that because most nurses spend some time near
search is that we only addressed studies whose main the glucose analyzer until the result of the analysis is
objective concerned evaluation of a given TGC protocol known, a computer situated next to the analyzer is the ideal
with a CDSS; we might have missed some studies with a spot for the system to interact with the nurses. This system
limited evaluation and TGC quality measurement focus. generated an active reminder but it could be shown only
The implementation of paper-based TGC protocols when nurses were near the analyzer, missing other clini-
with decision tables or charts for adjusting the insulin rate cians. As a final example, Rood et al. [8] evaluated a system
is cumbersome and time-consuming [11]. Although for which got the necessary data from a patient data-manage-
many studies without CDSS implementation results were ment system (PDMS), generating reminders during care at
acceptable [12, 13], frequent measurements are crucial in the bedside computer, and logging the data into the elec-
TGC. Furthermore, a paper-based TGC protocol cannot tronic medical record. This system provided an active
remind the users about the time for the next measurement. reminder but it is unknown what happens with a reminder if
For these reasons, CDSSs may improve glycemic control the user was logged out or was away from the computer.
and help implementing TGC. To the best of our knowl- Altogether more research is necessary to investigate the
edge this is the first review exclusively dedicated to most appropriate implementation site, target user, and time
supporting the TGC protocol with a CDSS. Existing of advice. This will also depend on local circumstances
reviews on TGC focused on the range of TGC, the effects, such as technical infrastructure and responsibilities of TGC
and advantages, regardless of the implementation strategy implementation.
of the TGC protocol, such as a CDSS [12, 13]. The
remainder of this section will discuss the findings on
study design, CDSS characteristics, and TGC quality TGC quality indicators
indicators and will provide recommendations on these
subjects for future research. We found no uniform indicator set of glycemic controls was
used in the studies reviewed. Most indicators differed in their
definitions among the studies although they are all meant to
Study design measure the same underlying concept. For example the
sampling frequency of measurements plays a crucial role in
Most of the studies used a before–after design and TGC but it was reported in nine out of 17 studies. In these
introduced the CDSS together with a new protocol. nine articles, six different ways were used to calculate the
1516

sampling frequency (e.g., mean or median per patient or per CDSS is implemented, we cannot be sure that the users
patient treatment day or mean sampling interval). Thus ask for support on time. In such a case usage of the system
reproducibility and comparability of research results are (frequency) should be clearly described. If an active
hampered by this lack of unambiguous definitions. The CDSS is implemented based on the content of these
choice of quality indicators used was not explicitly men- highly controlled protocols, the reminders will be shown
tioned in any of the studies. We could not find an association many times. Showing many pop-up reminders usually
between indicator selection and patient population, diseases, irritates users, resulting in the alerts being ignored [5, 6].
or specification of the designed protocols. Also, in an active CDSS it is not clear how many times
Because hypoglycemia is the main risk of TGC the messages are actually seen on time and not too late,
implementation, almost all studies reported on at least one because of users being logged off or away from the
indicator related to hypoglycemia. The number of hypo- computer. Therefore, in future research it is necessary to
glycemic events before and after TGC implementation report the usage, usability, and satisfaction with such
and/or the management of these events form the main CDSSs.
safety-related indicators of TGC. However, we found Data verification is very important, especially when
several definitions and ambiguous terminology for ren- the values are entered manually in information systems.
dering a BGL measurement (or a set thereof) as a Button et al. [21] implemented double data entry as a
hypoglycemic event. Most of the studies reported no data-verification tool. Morris et al. [22] mentioned this
change or even a reduction in the number of hypoglyce- issue, although they did not implement any data verifi-
mia events. This finding is in contrast with some studies, cation techniques. They are of the opinion that enabling
including a recent meta-analysis [18], showing that TGC bedside nurses to use CDSS with minimum effort out-
implementation increased the risk of hypoglycemia weigh the theoretical advantage of additional security and
events. TGC, semantically, excludes the occurrence of double data entry. Similar to others, in their study the
hypos because it is ‘‘tight’’. Thus, if one is applying a accuracy of manually entered values and the possibility of
TGC protocol but hypoglycemia occur then the culprit introducing new kind of errors were not reported or dis-
resides in the implementation of the regulatory process. cussed. Using CDSSs without a data-verification process
CDSSs are used to improve adherence to TGC protocols; might be a critical safety issue and hence authors ought to
therefore, if the protocol is sound a CDSS contributes to address this issue when discussing their results.
keeping BGL in range and hence to reducing the risk of
hyperglycemia and hypoglycemia.
The strong relationship between hyperglycemia and
mortality and morbidity is well known from the literature. Conclusion
Hence hyperglycemia reduction is the main objective of
TGC. Surprisingly, only six studies explicitly defined a While most studies evaluating the effect of CDSS on the
hyperglycemic event and/or hyperglycemia index, but quality of the TGC process found improvement when
they used different definitions of such an event. evaluated on the basis of the quality indicators used, it is
The increased number of blood-glucose measure- impossible to define the exact success factors. This is
ments, especially in well regulated patients, may increase mainly because of the lack of standard agreed-upon
the number of measurements with low glucose levels and indicators of glycemic control, the simultaneous imple-
may thus reduce the mean and median plasma glucose mentation of the CDSSs with new treatment protocols,
level of the whole group, giving a false impression of and the various solutions used for integrating the CDSS
improvement of the studies. into the users’ workflow. This systematic review provided
There is almost no literature that compares different key recommendations and information for researchers and
glycemic metrics to relevant clinical outcomes, such as ICU managers who want to develop and evaluate CDSSs
severity-associated mortality [19]. It is, therefore, hard to for glucose regulation or other highly controlled clinical
evaluate and qualify these indicators. Deciding upon a practices.
common glycemic vocabulary is hence essential.
Applying the results of highly controlled clinical trials Open Access This article is distributed under the terms of the
like those evaluating TGC to everyday practice is difficult Creative Commons Attribution Noncommercial License which
permits any noncommercial use, distribution, and reproduction in
[20]. The need for frequent measurements and pump any medium, provided the original author(s) and source are
adjustments increases the need for CDSSs. If a passive credited.
 1517

References
1. Van den Berghe G, Wilmer A, Hermans 11. Thomas AN, Marchant AE, Ogden MC, 22. Morris AH, Orme J Jr, Truwit JD,
G, Meersseman W, Wouters PJ, Milants Collin S (2005) Implementation of a Steingrub J, Grissom C, Lee KH, Li
I, Van Wijngaerden E, Bobbaers H, tight glycaemic control protocol using a GL, Thompson BT, Brower R, Tidswell
Bouillon R (2006) Intensive insulin web-based insulin dose calculator. M, Bernard GR, Sorenson D, Sward K,
therapy in the medical ICU. N Engl Anaesthesia 60:1093–1100 Zheng H, Schoenfeld D, Warner H
J Med 354:449–461 12. Nazer LH, Chow SL, Moghissi ES (2008) A replicable method for blood
2. van den Berghe G, Wouters P, Weekers (2007) Insulin infusion protocols for glucose control in critically ill patients.
F, Verwaest C, Bruyninckx F, Schetz critically ill patients: a highlight of Crit Care Med 36:1787–1795
M, Vlasselaers D, Ferdinande P, differences and similarities. Endocr 23. Pachler C, Plank J, Weinhandl H,
Lauwers P, Bouillon R (2001) Intensive Pract 13:137–146 Chassin LJ, Wilinska ME, Kulnik R,
insulin therapy in the critically ill 13. Meijering S, Corstjens AM, Tulleken Kaufmann P, Smolle KH, Pilger E,
patients. N Engl J Med 345:1359–1367 JE, Meertens JH, Zijlstra JG, Pieber TR, Ellmerer M, Hovorka R
3. Shortliffe EH (1987) Computer Ligtenberg JJ (2006) Towards a feasible (2008) Tight glycaemic control by an
programs to support clinical decision algorithm for tight glycaemic control in automated algorithm with time-variant
making. JAMA 258:61–66 critically ill patients: a systematic sampling in medical ICU patients.
4. Musen MA, Shahar Y, Shortliffe EH review of the literature. Crit Care Intensive Care Med 34:1224–1230
(2006) Biomedical informatics: 10:R19 24. Saager L, Collins GL, Burnside B,
computer applications in health care 14. Kawamoto K, Houlihan CA, Balas EA, Tymkew H, Zhang L, Jacobsohn E,
and biomedicine. Springer, New York, Lobach DF (2005) Improving clinical Avidan M (2008) A randomized study
pp 698–736 practice using clinical decision-support in diabetic patients undergoing cardiac
5. Eslami S, Abu-Hanna A, de Keizer NF systems: a systematic review of trials to surgery comparing computer-guided
(2007) Evaluation of outpatient identify features critical to success. glucose management with a standard
computerized physician medication BMJ 330:765 sliding scale protocol. J Cardiothorac
order entry systems: a systematic 15. Boord JB, Sharifi M, Greevy RA, Vasc Anesth 22:377–382
review. J Am Med Inform Assoc Griffin MR, Lee VK, Webb TA, May 25. Hermayer KL, Neal DE, Hushion TV,
14:400–406 ME, Waitman LR, May AK, Miller RA Irving MG, Arnold PC, Kozlowski L,
6. Eslami S, Keizer NF, Abu-Hanna A (2007) Computer-based insulin infusion Stroud MR, Kerr FB, Kratz JM (2007)
(2007) The impact of computerized protocol improves glycemia control Outcomes of a cardiothoracic intensive
physician medication order entry in over manual protocol. J Am Med care web-based online intravenous
hospitalized patients—a systematic Inform Assoc 14:278–287 insulin infusion calculator study at a
review. Int J Med Inform 77:365–376 16. Vogelzang M, Zijlstra F, Nijsten MW Medical University Hospital. Diabetes
7. Davidson PC, Steed RD, Bode BW (2005) Design and implementation of Technol Ther 9:523–534
(2005) Glucommander: a computer- GRIP: a computerized glucose control 26. Toschlog EA, Newton C, Allen N,
directed intravenous insulin system system at a surgical intensive care unit. Newell MA, Goettler CE, Schenarts PJ,
shown to be safe, simple, and effective BMC Med Inform Decis Mak 5:38 Bard MR, Sagraves SG, Rotondo MF
in 120, 618 h of operation. Diabetes 17. Vogelzang M, Loef BG, Regtien JG, (2007) Morbidity reduction in critically
Care 28:2418–2423 van der Horst IC, van Assen H, Zijlstra ill trauma patients through use of a
8. Rood E, Bosman RJ, van der Spoel JI, F, Nijsten MW (2008) Computer- computerized insulin infusion protocol:
Taylor P, Zandstra DF (2005) Use of a assisted glucose control in critically ill a preliminary study. J Trauma 62:1370–
computerized guideline for glucose patients. Intensive Care Med 34:1421– 1375 discussion 1375–1376
regulation in the intensive care unit 1427 27. Laha SK, Taylor R, Collin SA, Ogden
improved both guideline adherence and 18. Wiener RS, Wiener DC, Larson RJ M, Thomas AN (2008) Glucose control
glucose regulation. J Am Med Inform (2008) Benefits and risks of tight in critical illness using a web-based
Assoc 12:172–180 glucose control in critically ill adults: a insulin dose calculator. Med Eng Phys
9. Dortch MJ, Mowery NT, Ozdas A, meta-analysis. JAMA 300:933–944 30:478–482
Dossett L, Cao H, Collier B, Holder G, 19. Eslami S, de Keizer NF, Schultz MJ, 28. Meynaar IA, Dawson L, Tangkau PL,
Miller RA, May AK (2008) A de Jonge E, Abu-Hanna A (2008) A Salm EF, Rijks L (2007) Introduction
computerized insulin infusion titration systematic review on quality indicators and evaluation of a computerised
protocol improves glucose control with for tight glycaemic control in critically insulin protocol. Intensive Care Med
less hypoglycemia compared to a ill patients—need for an unambiguous 33:591–596
manual titration protocol in a trauma indicator reference subset. Crit Care 29. Shulman R, Finney SJ, O’Sullivan C,
intensive care unit. JPEN J Parenter 12:R139 Glynne PA, Greene R (2007) Tight
Enteral Nutr 32:18–27 20. Rothwell PM (2005) External validity glycaemic control: a prospective
10. Juneja R, Roudebush C, Kumar N, of randomised controlled trials: ‘‘to observational study of a computerised
Macy A, Golas A, Wall D, Wolverton whom do the results of this trial decision-supported intensive insulin
C, Nelson D, Carroll J, Flanders SJ apply?’’. Lancet 365:82–93 therapy protocol. Crit Care 11:R75
(2007) Utilization of a computerized 21. Button E, Keaton P (2006) Glycemic
intravenous insulin infusion program to control after coronary bypass graft:
control blood glucose in the intensive using intravenous insulin regulated by a
care unit. Diabetes Technol Ther computerized system. Crit Care Nurs
9:232–240 Clin North Am 18:257–265