effects of two different levels of computerized decision support on blood glucose regulation in...

8
i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60 j ourna l homepage: www.ijmijournal.com Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients Saeid Eslami a,, Nicolette F. de Keizer a , Dave A. Dongelmans b , Evert de Jonge c , Marcus J. Schultz b,d,e , Ameen Abu-Hanna a a Department of Medical Informatics, Academic Medical Center, University of Amsterdam, Meibergdreef 15, 1105 AZ Amsterdam, The Netherlands b Department of Intensive Care, Academic Medical Center, University of Amsterdam, Amsterdam, The Netherlands c Department of Intensive Care, Leiden University Medical Center, Leiden, The Netherlands d Laboratory of Experimental Intensive Care and Anesthesiology, Academic Medical Center, University of Amsterdam, Amsterdam, The Netherlands e HERMES Critical Care Group, Amsterdam, The Netherlands a r t i c l e i n f o Article history: Received 16 March 2011 Received in revised form 7 October 2011 Accepted 7 October 2011 Keywords: Computerized decision support system Evaluation Safety statistical process control Glucose regulation a b s t r a c t Introduction: Although the use of computerized decision support systems (CDSS) in glucose control in the ICU has been reported, little is known about the effect of the systems’ operating modes on the quality of glucose control. The objective of this study was to evaluate the effect of providing patient-specific and patient non-specific computerized advice on timing of blood glucose level (BGL) measurements. Our hypothesis was that both levels of support would be effective for improving the quality of glucose regulation and safety, with patient specific advice being the most effective strategy. Patients and methods: A prospective study was performed in a 30-bed mixed medical-surgical intensive care unit (ICU) of a university hospital. In phase 1 the CDSS provided non-specific advice and thereafter, in phase 2, the system provided specific advice on timing of BGL measurements. The primary outcome measure was delay in BGL measurements before and after the two levels of support. Secondary endpoints were sampling frequency, mean BGL, BGL within pre-defined targets, time to capture target, incidences of severe hypoglycemia and hyperglycemia. These indicators were analyzed over the course of time using Statisti- cal Control Charts. The analysis was restricted to patients with at least two blood glucose measurements. Results: Data of 3934 patient admissions were evaluated, which corresponded to 119,116 BGL measurements. The BGL sampling interval, delays in BG sampling, and percentage of hypoglycemia all decreased after introducing either of the two levels of decision support. The effect was however larger for the patient specific CDSS. Mean BGL, time to capture target, hyperglycemia index, percentage of hyperglycemia events and “in range” measurements remained unchanged and stable after introducing both patient non-specific and patient specific decision support. Corresponding author. Tel.: +31 20 5666252. E-mail address: [email protected] (S. Eslami). 1386-5056/$ see front matter © 2011 Elsevier Ireland Ltd. All rights reserved. doi:10.1016/j.ijmedinf.2011.10.004

Upload: saeid-eslami

Post on 05-Sep-2016

216 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

Es

SMa

Nb

c

d

Ne

a

A

R

R

7

A

K

C

s

E

S

G

1d

i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60

j ourna l homepage: www.i jmi journa l .com

ffects of two different levels of computerized decisionupport on blood glucose regulation in critically ill patients

aeid Eslamia,∗, Nicolette F. de Keizera, Dave A. Dongelmansb, Evert de Jongec,arcus J. Schultzb,d,e, Ameen Abu-Hannaa

Department of Medical Informatics, Academic Medical Center, University of Amsterdam, Meibergdreef 15, 1105 AZ Amsterdam, TheetherlandsDepartment of Intensive Care, Academic Medical Center, University of Amsterdam, Amsterdam, The NetherlandsDepartment of Intensive Care, Leiden University Medical Center, Leiden, The NetherlandsLaboratory of Experimental Intensive Care and Anesthesiology, Academic Medical Center, University of Amsterdam, Amsterdam, TheetherlandsHERMES Critical Care Group, Amsterdam, The Netherlands

r t i c l e i n f o

rticle history:

eceived 16 March 2011

eceived in revised form

October 2011

ccepted 7 October 2011

eywords:

omputerized decision support

ystem

valuation

afety statistical process control

lucose regulation

a b s t r a c t

Introduction: Although the use of computerized decision support systems (CDSS) in glucose

control in the ICU has been reported, little is known about the effect of the systems’ operating

modes on the quality of glucose control. The objective of this study was to evaluate the

effect of providing patient-specific and patient non-specific computerized advice on timing

of blood glucose level (BGL) measurements. Our hypothesis was that both levels of support

would be effective for improving the quality of glucose regulation and safety, with patient

specific advice being the most effective strategy.

Patients and methods: A prospective study was performed in a 30-bed mixed medical-surgical

intensive care unit (ICU) of a university hospital. In phase 1 the CDSS provided non-specific

advice and thereafter, in phase 2, the system provided specific advice on timing of BGL

measurements. The primary outcome measure was delay in BGL measurements before and

after the two levels of support. Secondary endpoints were sampling frequency, mean BGL,

BGL within pre-defined targets, time to capture target, incidences of severe hypoglycemia

and hyperglycemia. These indicators were analyzed over the course of time using Statisti-

cal Control Charts. The analysis was restricted to patients with at least two blood glucose

measurements.

Results: Data of 3934 patient admissions were evaluated, which corresponded to 119,116

BGL measurements. The BGL sampling interval, delays in BG sampling, and percentage of

hypoglycemia all decreased after introducing either of the two levels of decision support. The

effect was however larger for the patient specific CDSS. Mean BGL, time to capture target,

hyperglycemia index, percentage of hyperglycemia events and “in range” measurements

remained unchanged and stable after introducing both patient non-specific and patient

specific decision support.

∗ Corresponding author. Tel.: +31 20 5666252.E-mail address: [email protected] (S. Eslami).

386-5056/$ – see front matter © 2011 Elsevier Ireland Ltd. All rights reserved.oi:10.1016/j.ijmedinf.2011.10.004

Page 2: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

54 i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60

Adherence to protocol sampling rules increased by using decision support with a larger

effect at the patient specific level. This led to a decrease in the percentage of hypoglycemia

events and improved safety. The use of the CDSS at both levels, however, did not improve

the quality of glucose control as measured by our indicators. More research is needed to

investigate whether other socio-technical factors are in play.

A local ICU database identified all patients admitted to the ICU.The hospital information system and PDMS were searched for

Table 1 – Protocol characteristics.

Type of protocol Sliding scalesPresent in what form WrittenWho is responsible for glucose controlStart of insulin By nurseDosing of insulin By nurseCorrection of hypoglycemia By nurseProtocol thresholds and targetsStart of insulin (mg/dL) >144BGL targets (mg/dL) 90–144Timing of BGL-measurements

described in or mandated by theprotocol

Yes

Rules on stopping insulinThreshold to stop insulin infusion <63 or 63–144 with > 50%

reduction of BGLOther reason for stopping insulin Stop of feedingAction in case of hypoglycemiaBGL < 40 mg/dL 50 ml 20% glucose

1. Introduction

Studies have shown that critically ill patients could benefitfrom blood glucose regulation during their stay in the inten-sive care unit (ICU) [1,2]. Recent studies have shown thatfrequent and timely BGL measurements improved the qualityand safety of blood glucose regulation [3,4]. Continuous bloodglucose monitoring would have provided timely measurementof blood glucose, but devices to do this are expensive and arenot commonly available yet.

Clinical computerized decision support systems (CDSS) arecomputer programs that are intended to help healthcare work-ers in making decisions [5]. A CDSS can be characterized bythe level of support. The level of decision support varies fromnon-patient specific (general support) to patient specific sup-port. For the purpose of this study, non-patient specific CDSSis defined as showing the protocol without interpreting thepatient’s data. On the other hand, patient specific CDSS refersto making a conclusion or giving advice after interpreting thepatient’s data. The difference between the effects of these lev-els of support has not been studied yet, at least not in thisdomain.

ICU staff increasingly use patient data management sys-tems (PDMS) allowing them to access data, but these systemsoften do not provide active support for making decisions.Based on the literature and our recent systematic reviews[6,7] we hypothesized that both levels of support, patientnon-specific as well as patient specific, would increase theadherence to protocol sampling rules in terms of reducingblood glucose sampling delays, and that the patient specificsupport would give better results. The underlying idea is thatimproving adherence to the protocol leads to improving qual-ity and safety of glucose regulation. The purpose of this studywas to test this hypothesis and measure the effect of the twodifferent levels of support, patient non-specific and patientspecific support, embedded in a PDMS, on adherence to sam-pling rules of a locally developed glucose regulation protocol.

2. Methods

2.1. Study location

Collection of data was prospectively performed in a 30-bed“closed-format” mixed medical-surgical ICU of an academichospital in the Netherlands. For each patient there was oneICU nurse who stayed near to the patient bed. The ICU uses

a patient data management system (PDMS) (Metavision,ImdSoft Sassenheim, the Netherlands) since March 2002.There is a bedside computer available for each bed. TheICU teams used the PDMS to complete all patient charting

© 2011 Elsevier Ireland Ltd. All rights reserved.

and documentation such that no information had paper asits primary storage mechanism. The laboratory computerinterface conveyed BGL measurement results to the PDMS.Consequently, the PDMS processes and displays all valuesdirectly after their measurement with a maximum delay of1 min, introduced by the interface.

2.2. Local glucose control protocols

Table 1 shows the protocol’s characteristics. According to theprotocol when the latest BGL is in the normal range it shouldbe rechecked after 4 h. When the latest BGL is out of thepre-defined range it should be generally rechecked after 1 h.When the latest BGL is in a hypoglycemic range it should berechecked after 30 min.

2.3. Measurement of BGL

Arterial blood samples were used for BGL measurements.These measurements were obtained by blood gas analyzers(Rapidlab 865®, Bayer, Germany) in the ICU and a glucoseanalyzer (Modular P800® system, Roche Diagnostics GmbH,Germany) in the hospital’s central laboratory.

2.4. Patients

The term “sliding scale” here refers to a dynamic protocol forintravenous insulin infusion. The protocol is self-regulating, there-fore it does not consider baseline glucose value nor condition (e.g.diabetes).

Page 3: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

c a l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60 55

aacimfi

2

T

2citicp

uBmwnti

cBttsamoBoiw

2

DcsiT

2

SiSsoittWi

Table 2 – Outcome measures description.

Protocol related indicators (primary outcome measures)Sampling frequency Represented as the mean sampling

intervalDelay in blood glucose

monitoringRepresented as the mean delay time.When the next measurement was doneearlier than 10 min before the scheduledtime, delay time was considered −10

Safety-related indicatorsHypoglycemia Reported as percentage of overall BGL

measurements ≤40 mg/dl (severe) and≤63 mg/dl (moderate) and as percentageof patients with at least 1 episode of(severe) hypoglycemia. After ahypoglycemic event, to classify asubsequent BGL measurement as startinga new hypoglycemic event we requirethat the BGL first increase into the normalrange and then drop again below thehypoglycemia threshold in a subsequenthour

Hyperglycemia Defined as BGL > 144 mg/dl, and severehyperglycemia as BGL >180 mg/dl. Theseare reported as the percentage of overallBGL measurements above the giventhreshold

Hyperglycemia index The area between the BGL curve and the144 mg/dl (hyperglycemia) thresholddivided by time. We report on the mean ofthis value per patient

Effectiveness/efficiency-related indicatorsMean BGL The overall mean BGL during a defined

time interval (more focus on efficiencycan be obtained by shortening the

i n t e r n a t i o n a l j o u r n a l o f m e d i

ll records on BGL. The first BGL measured directly after ICUdmittance was excluded from the final analysis for furtheralculation, because they are hardly influenced by any ICU reg-men. Admitted patients who had in total more than two BGL

easurements, including admission BGL, were included in thenal analysis.

.5. Intervention

he intervention was carried out in three phases.The glucose regulation protocol was implemented in early

002 and was last revised in 2005 [8]. This nurse-driven proto-ol was in use and available on both paper and the hospitalntranet in 2002. In phase 0 in this study the glucose con-rol protocol was still only available on paper and the hospitalntranet. In this phase, lasting for 8 months, we prospectivelyollected the data that formed the baseline for the consequenthases.

In phase 1 of the study a patient non-specific CDSS wassed for 8 months. Every 5 min the system checked for newGL measurements. Only when there was a new BGL measure-ent the CDSS showed a snapshot of the protocol in a pop-upindow. This non-specific reminder appeared only if the ICUurse was logged in at the bedside computer and worked inhe PDMS. The pop-up window showed the whole BGL protocolncluding the general timing and pump position advice.

In phase 2, with duration of 10 months, the CDSS alsohecked for new BGL measurements every 5 min. When a newGL measurement was encountered the CDSS calculated theime for the next measurement according to the protocol andhe patient’s former BGLs. The time for the next expected mea-urement was then recorded in a database. The CDSS gave

reminder about 10 min before the next expected measure-ent was due. The pop-up window showed the planned time

f measurement and reminded the ICU nurse to measure theGL and to follow the protocol. The system showed a messagenly if the expected BGL was not timely performed and only

f the ICU nurse was logged in at the bedside computer andorked with the PDMS.

.6. Outcome measures

elay in blood glucose monitoring was the primary out-ome measure. Other common performance indicators wereelected as secondary outcome measures to show the qual-ty of glucose regulation [9]. These indicators are described inable 2.

.7. Statistical process control (SPC)

PC uses statistical methods to identify periods of time dur-ng which a process goes from “in control” to “out of control.”PC and its primary tool – the control chart – are a branch oftatistics that combines rigorous time series analysis meth-ds with graphical data presentation, often yielding insights

nto the data more quickly and in a more understandable way

han other statistical techniques [10]. Control charts can dis-inguish between common and special causes of variation [11].

ith common cause variation (noise), the variation is inherentn the process itself and the process is stable and predictable

interval)

within certain limits. Special cause variation signifies that theprocess is no longer stable or predictable and has changed(either for better or worse) [12]. A control chart includes a plotof the data over time with three additional lines – the centerline (usually reflecting the mean) and an upper and lower con-trol limits, typically set at ±3 sigma from the mean. When thedata points are, without any special pattern, within the con-trol limits then the process is “in control” and stable. Thereare several rules that indicate when a special cause variationor a special pattern has occurred on a control chart. We usedthe following four widely used rules to identify a special causevariation: one or more points beyond a control limit, a run ofseven or more points on one side of the center line, two outof three consecutive points appearing beyond 2 sigma on thesame side of the center line, or a run of seven or more pointsall trending up or down.

2.8. Power analysis

Previous studies have shown that CDSS could reduce delay inmeasurement by about 50% [13]. Power analysis showed thatat least 2000 BGL measurements were needed to demonstratean absolute difference of 25% in delay time between baselineand the most effective intervention, with an alpha of 0.05 and

beta of 0.8. Retrospective data analysis showed that at least 2weeks of data would be needed to meet these requirements.
Page 4: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

56 i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60

Table 3 – Patient characteristics.

Variable Phase 0 Patient non-specific CDSS Patient specific CDSS p-Value

Number 1014 1017 1470Age (years) 60 ± 15.5 60 ± 15.8 61 ± 15.9 0.62Male (%) 36% 36% 36% 0.91APACHE II 19.6 ± 7.8 21.1 ± 7.8* 20.6 ± 7.6* <0.01Surgical 62% 60% 57% 0.09ICU LOS (day) 4.9 ± 7.9 4.6 ± 6.8 4.4 ± 7.5 0.25ICU mortality (%) 12% 11% 11% 0.73Admission BGL 8.4 ± 4.5 8.4 ± 4.6 8.2 ± 4.1 0.70

All data are mean ± SD (medians).∗ Phase 0 and patient non-specific; phase 0 and patient specific.

2.9. Statistical analysis

We used the X-MR control chart [11] to plot and analyze theprocesses. We used the X-MR chart instead of attribute charts[14,15] due to the large subgroup size and hence the increasedchance of false positive results. The quality indicators thatwe choose (each represents either a mean or a proportion)were calculated per month and plotted as points on the X-MRchart. The mean of the points before glucose control imple-mentation was calculated along with the ±3 sigma limits. Todetermine whether a change in the process occurs furtheralong the time axis, the mean and control limits, as calcu-

lated based on the 8 months of phase 0, were extrapolatedover the entire study period. The ANOVA, Kruskal–Wallis andChi square tests were used to assess the statistical significance

Fig. 1 – Control charts of mean BGL sampling intervals and delaySeptember 2008 (introducing of phase II) and according to SPC ruline) all processes became out of control (here corresponding to alimits (UCL) and lower control limits (LCL) were calculated basedentire study period.

of differences among pre- and post-intervention periods andto compare these results with those of the SPC analysis. Allanalysis was performed with Systat 12.

3. Results

3.1. Patients

In total, data of 3934 patient admissions (from 01 May 2007to 30 June 2009) were evaluated, corresponding to 119,116

BGL measurements prospectively. Table 3 shows the patientbaseline characteristics. Only the APACHE II score showed astatistically significant higher value in the intervention periodcompared to the “before” period.

in BG monitoring (protocol-related indicator). Sinceles (run of seven or more points on one side of the center

significant reduction). The center lines (CL), upper control on the first 8 months (phase 0) and extrapolated over the

Page 5: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

c a l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60 57

3

I4a

3

OtviFppirpcaii(crgs

3

Cighv≤hsw

3

FettiC

tcr

4

Ibai

Res

ult

of

com

par

ing

thre

e

ph

ases

.

Mea

nd

elay

Mea

nd

elay

<80

Mea

n

del

ayta

rget

ran

geM

ean

del

ay>

144

BG

L

test

inte

rval

Med

ian

BG

LSD

Hyp

erin

dex

Tim

e to

cap

ture

targ

et

Seve

reh

ypog

lyce

mic

pat

ien

tsa

Mod

erat

eh

ypog

lyce

mic

pat

ien

tsa

≤40

≤63

80–1

44

>14

4

50

±

93

103

±

100

38

±

89

62

±

92

204

±

123

7.9

±

1.8

1.79

±

1.31

12.8

± 25

.2

483

±

414

2.7

12.6

.15

.96

58.4

33.6

1–4b

−11

to

4b3–

7b−2

to

4b3–

7bn

scn

scn

scn

scn

sdn

sdn

sd1.

06(.9

1,

1.23

)dn

sdn

sd

48

±

91

107

±

107

33

±

85

61

±

91

198

±

120

7.8

±

1.6

1.78

±

1.27

11.8

± 19

.7

481

±

435

2.6

12.7

.15

1.02

57.6

33.9

7–12

2–15

7–10

9–14

9–13

ns

ns

ns

ns

ns

ns

ns

.80(

.69,

.92)

ns

ns

37

±

81

98

±

97

25

±

76

49

±

81

188

±

110

7.8

±

1.7

1.82

± 1.

84

11.9

±

23.4

477

±

405

2.0

11.1

.10

.82

57.9

33.6

11–1

4

−1

to

12

12–1

5

10–1

5

14–1

8

ns

ns

ns

ns

ns

ns

ns

0.85

(.73,

.98)

ns

ns

rep

rese

nte

d

as

mea

n

±

SD

or

per

cen

tage

.

wit

h

at

leas

t

one

BG

L

40

mg/

dl (

seve

re) a

nd

≤63

mg/

dl (

mod

erat

e).

(95%

con

fid

ence

inte

rval

).–W

alli

s

H.

tio

(95%

con

fid

ence

inte

rval

of

odd

s

rati

o).

i n t e r n a t i o n a l j o u r n a l o f m e d i

.2. CDSS messages

n the two intervention periods, the message was shown1,875 times: 18,190 times in the patient non-specific phasend 23,685 times in the patient specific phase.

.3. Protocol related indicators

verall adherence to the part of the protocol concerningiming of measurements increased significantly. Mean inter-al between BGL measurements decreased significantly afterntroducing the CDSS in both levels of support (Table 4 andig. 1). However the reduction of the interval was larger whenatient specific decision support was provided. The otherrotocol related indicator, delay in blood glucose monitor-

ng, decreased and became “out of control,” according to SPCules, after CDSS introduction in both specific and non-specifichases. Becoming “out of control” means that a significanthange has been detected. Because this change is related to

decrease in the delay time, this is a good change signify-ng improvement. The reduction in the glucose measurementnterval was greater in the BGL measurements that were highhyperglycemic) and that were within target range (Fig. 1), inomparison to the measurements that were below the targetange (hypoglycemic). This reduction was also significantlyreater in the patient specific phase than in the patient non-pecific phase.

.4. Safety of glucose control

ontrol charts of safety-related quality indicators are shownn Figs. 2 and 3 (hypoglycemia, hyperglycemia, and hyper-lycemic index). The percentage of hyperglycemia andyperglycemic indices did not change after any of the inter-entions. The percentage of BGL measurements that were40 mg/dl (severe hypoglycemia) and ≤65 mg/dl (moderateypoglycemia) decreased after implementation of the patientpecific CDSS (Fig. 2). According to SPC rules, this reductionas statistically significant for moderate hypoglycemia.

.5. Effectiveness and efficiency of glucose control

ig. 4 shows the quality control charts for theffectiveness/efficiency-related indicators of glucose con-rol. Mean BGL, the percentage of BGL within locally definedargets, and time to reach targets did not change afterntroducing the patient non-specific and patient specificDSS.

The results of ANOVA, Kruskal–Wallis and chi squareests on the effect of introducing glucose control and relatedhanges on these indicators were concordant with the SPCesults (Table 4).

. Discussion

n this study we showed that a CDSS, integrated in a PDMS inoth patient specific and patient non-specific levels, improveddherence to a locally developed protocol for BGL regulationn terms of reducing blood glucose monitoring delay in ICU

Tabl

e

4

Phas

es

0 0

vs. 1

1 1

vs. 2

2 0

vs. 3

Dat

a

wer

ea

Pati

ents

bA

NO

VA

cK

rusk

ald

Od

ds

ra

Page 6: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

58 i n t e r n a t i o n a l j o u r n a l o f m e d i c a l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60

Fig. 2 – Control charts of percentage BGL ≤40 mg/dl and≤63 mg/dl (safety-related indicators), shown respectively atthe bottom and top of the figure. Since April 2009 (markedby * in the chart at the top) and according to the SPC rules(one point beyond lower control limit and two out of threeconsecutive points beyond 2 sigma on the same side ofcontrol line) the percentage of moderate hypoglycemiasignificantly reduced and the process became out ofcontrol. However, although severe hypoglycemiademonstrated reduction this was not detectable by SPCrules as a significant one. The center lines (CL), uppercontrol limits (UCL) and lower control limits (LCL) werecalculated based on the first 8 months (phase 0) andextrapolated over the entire study period.

Fig. 3 – Control charts of mean hyperglycemia index andpercentage of BGL > 144 mg/dl (safety-related indicators),shown respectively at the bottom and top of the figure.According to SPC rules all processes are stable and incontrol (i.e. there are no significant changes detected). Thecenter lines (CL), upper control limits (UCL) and lowercontrol limits (LCL) were calculated based on the first 8months (phase 0) and extrapolated over the entire study

patients. However, the effect was greater in the patient spe-cific CDSS. Interestingly, the improvement in adherence toprotocol sampling rules did not lead to improvement in thequality of BGL regulation. We found a slight improvement inthe safety aspect of BGL regulation only for the CDSS withpatient-specific advice, with a decrease in the percentage ofmoderate hypoglycemic events.

Recent studies [3,4] showed that when BGL measurement isperformed frequently and on time then the quality and safetyof glucose regulation improve. There are several reasons for

delays in glucose monitoring, and forgetting to perform a testis the most important one. A computerized reminder system,like the one which we developed and implemented, could be

period.

useful. There are several evaluation studies on the effect ofCDSS on BGL regulation [6]. In contrast to the majority of thepublished studies in this field, we used an active CDSS, mean-ing that it takes the initiative to act, and it was integrated inthe daily workflow of nurses. In our study the glucose regu-lation protocol was available on paper and in use for a longtime before we implemented the CDSS. Other studies imple-mented a BGL regulation protocol and CDSS simultaneouslywhich makes it hard to isolate the effect of the system alone[16]. In addition, to our knowledge, this is the first study thataimed to implement a protocol on sampling rules and com-pare two levels of support.

The CDSS used in the present study comprised of a patientnon-specific active system, which is only aware of the protocolbut does not interpret the patient’s data, and a patient-specific

active system. Developing and integrating a patient specificCDSS is more complicated and technically demanding thana patient non-specific CDSS. Therefore, if they were equally
Page 7: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

i n t e r n a t i o n a l j o u r n a l o f m e d i c a l

Fig. 4 – Control charts, from bottom to top, of percentage ofBGLs in protocol’s range, time to reach target range, andmean BGL (efficiency-related indicators). According to SPCrules all processes are stable and in control (no significantchange was detected). The center lines (CL), upper controllimits (UCL) and lower control limits (LCL) were calculatedbased on the first 8 months (phase 0) and extrapolated overt

epsis

Authors’ contributions

he entire study period.

ffective, the patient non-specific style would seem to be moreractical and generalizable. A potential advantage of patient

pecific systems, however, is that the users do not need tonterpret the protocol and make the mental calculations them-elves, as the CDSS does this instead. This is easier and less

i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60 59

time consuming for the clinicians and might increase theacceptance of the system by healthcare workers, thus increas-ing the effect of the CDSS.

In this study, the CDSS is triggered by a time-event (every5 min) thereby assuring that for all new BGL measurements thetime of next measurement is calculated in a timely manner.This also means that the reminder appears with a maxi-mum delay of 5 min from the desired launching time (whichis 10 min before due time of a measurement).

Delay in blood glucose monitoring significantly decreasedafter CDSS introduction. This reduction was larger in thehigher (hyperglycemic) and in the target range measurementsthan the measurements below the range (hypoglycemic). Thismeans that after CDSS implementation, when the BGL wasabove or in the target range, the nurses performed the nextBGL measurement with less delay according to the protocol.

The frequency of measurements increased and the delay inblood glucose monitoring decreased but the quality of glucoseregulation did not change according to the indicators used.Only a slight improvement was shown in moderate hypo-glycemia events. This finding could be explained in differentways. First we note that according to the protocol, the qual-ity of glucose regulation was already in acceptable range inour patients before using the CDSS. Hence it is fair to hypoth-esize that the effect of a CDSS in situations where qualityof glucose regulation is worse could be much higher than inour patients. Second, the system is just one part of a widersocio-technical environment where it interacts with people[17]. Factors such as trust of nurses in the protocol may explainwhy quality of BG regulation did not improve. It could be thecase that despite increasing the BGL measurement frequency,the nurses did not adjust the insulin pump accordingly. Third,external factors such as published negative studies [18] andmeta-analysis studies [19,20] showing no benefit of tight gly-caemic regulation could have indirectly affected the behaviorof the nurses.

A limitation of our study is that we did not investigate if theuse of this CDSS has any influence on insulin administrationnor clinically relevant endpoints such as survival, length ofstay and costs of treatment. However, as glucose regulationis shown to be an evidence based strategy that may decreasemortality, adherence to this strategy is commonly advocated.

5. Conclusion

Adherence to protocol sampling rules increased with the useof decision support, with a larger effect using patient-specificdecision support. This led to a decrease in the percentageof hypoglycemia events and improved safety. The use of theCDSS at both levels, however, did not improve the quality ofglucose control as measured by our indicators. More researchis needed to investigate whether other socio-technical factorsare in play.

All authors made substantial contributions to the study designand methods. Saeid Eslami extracted data and analyzed data

Page 8: Effects of two different levels of computerized decision support on blood glucose regulation in critically ill patients

60 i n t e r n a t i o n a l j o u r n a l o f m e d i c a

Summary pointsWhat is already known:

• CDSSs have been identified as key for improvingpatient’s safety and outcomes.

• CDSSs can increase adherence to the guidelines.

What this study added to our knowledge:

• A CDSS integrated in a patient data management sys-tem in both levels, patient non-specific and patientspecific, improved adherence to protocol with a largereffect at the patient specific level.

• The use of the CDSS at both levels, however, did notimprove the quality of care as measured by our indi-cators. More research is needed to investigate whethersocio-technical factors are at play.

• Statistical process control is a useful tool formonitoring the effect over time and captures within-institution significant and stable changes.

r

and drafted the manuscript. All authors interpreted the resultsand were involved in revising the final manuscript.

Competing interest

None of the authors has any conflict of interest in themanuscript. There were not any financial or other relationswith relevant parties that could have affected the results andconclusions of the study.

e f e r e n c e s

[1] G. Van den Berghe, A. Wilmer, G. Hermans, W. Meersseman,P.J. Wouters, I. Milants, E. Van Wijngaerden, H. Bobbaers, R.Bouillon, Intensive insulin therapy in the medical ICU, N.Engl. J. Med. 354 (2006) 449–461.

[2] G. van den Berghe, P. Wouters, F. Weekers, C. Verwaest, F.Bruyninckx, M. Schetz, D. Vlasselaers, P. Ferdinande, P.Lauwers, R. Bouillon, Intensive insulin therapy in thecritically ill patients, N. Engl. J. Med. 345 (2001) 1359–1367.

[3] R. Garg, A. Jarry, M. Pendergrass, Delay in blood glucosemonitoring during an insulin infusion protocol is associatedwith increased risk of hypoglycemia in intensive care units,J. Hosp. Med. 4 (2009) E5–E7.

[4] R. Juneja, C.P. Roudebush, S.A. Nasraway, A.A. Golas, J.Jacobi, J. Carroll, D. Nelson, V.J. Abad, S.J. Flanders,

Computerized intensive insulin dosing can mitigatehypoglycemia and achieve tight glycemic control whenglucose measurement is performed frequently and on time,Crit. Care 13 (2009) R163.

l i n f o r m a t i c s 8 1 ( 2 0 1 2 ) 53–60

[5] E.H. Shortliffe, Computer programs to support clinicaldecision making, JAMA 258 (1987) 61–66.

[6] S. Eslami, A. Abu-Hanna, E. de Jonge, N.F. de Keizer, Tightglycemic control and computerized decision-supportsystems: a systematic review, Intensive Care Med. 35 (2009)1505–1517.

[7] S. Eslami, N.F. Keizer, A. Abu-Hanna, The impact ofcomputerized physician medication order entry inhospitalized patients—a systematic review, Int. J. Med. Inf.(2007).

[8] S. Eslami, A. Abu-Hanna, N.F. de Keizer, R.J. Bosman, P.E.Spronk, E. de Jonge, M.J. Schultz, Implementing glucosecontrol in intensive care: a multicenter trial using statisticalprocess control, Intensive Care Med., 36, 1556–1565.

[9] S. Eslami, N.F. de Keizer, E. de Jonge, M.J. Schultz, A.Abu-Hanna, A systematic review on quality indicators fortight glycaemic control in critically ill patients: need for anunambiguous indicator reference subset, Crit. Care 12 (2008)R139.

[10] J.C. Benneyan, R.C. Lloyd, P.E. Plsek, Statistical processcontrol as a tool for research and healthcare improvement,Qual. Saf. Health Care 12 (2003) 458–464.

[11] M.A. Mohammed, P. Worthington, W.H. Woodall, Plottingbasic control charts: tutorial notes for healthcarepractitioners, Qual. Saf. Health Care 17 (2008)137–145.

[12] R.G. Carey, How do you know that your chart is improving?Part I: Basic concepts in statistical thinking, J. Ambul. CareManage. 25 (2002) 80–87.

[13] E. Rood, R.J. Bosman, J.I. van der Spoel, P. Taylor, D.F.Zandstra, Use of a computerized guideline for glucoseregulation in the intensive care unit improved bothguideline adherence and glucose regulation, J. Am. Med.Inform. Assoc. 12 (2005) 172–180.

[14] R.G. Carey, Improving Healthcare with Control Chart: Basicand Advanced SPC Methods and Case Studies, ASQ QualityPress, Milwaukee, 2003.

[15] D.J. Wheeler, Advanced Topics in Statistical Process Control,SPC Press, Knoxville, TN, 1995.

[16] R. Goud, A. Hasman, A.M. Strijbis, N. Peek, A parallelguideline development and formalization strategy toimprove the quality of clinical practice guidelines, Int. J.Med. Inf. 78 (2009) 513–520.

[17] M. Berg, J. Aarts, J. van der Lei, ICT in health care:sociotechnical approaches, Methods Inf. Med. 42 (2003)297–301.

[18] 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, J.J. Ronco,Intensive versus conventional glucose control in critically illpatients, N. Engl. J. Med. 360 (2009) 1283–1297.

[19] R.S. Wiener, D.C. Wiener, R.J. Larson, Benefits and risks oftight glucose control in critically ill adults: a meta-analysis,JAMA 300 (2008) 933–944.

[20] 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, D. Talmor, Intensive insulin therapy andmortality among critically ill patients: a meta-analysisincluding NICE-SUGAR study data, CMAJ 180 (2009) 821–827.