Introduction

Stress-induced hyperglycemia during intensive care unit (ICU) admission has a strong andconsistent relation with mortality [1–3]. Nevertheless, hyperglycemia in these populations of patients was not alwaystreated with insulin infusion until the publication of a landmark single-center study in2001 [4]. This trial demonstrated reductions in mortality when continuous intravenousinsulin was used to achieve blood glucose (BG) from 80 to 110 mg/dl, compared withconventional therapy. Although these findings were corroborated in a large single-centercohort study [5], they were not confirmed by subsequent randomized trials [6–10].

One possible explanation for the divergent results among such trials may relate to theincidence of severe hypoglycemia sustained by patients in the interventional arms ofrandomized trials [6–11]. Data from observational [12–17] and interventional studies [4, 6, 11] demonstrated a strong and independent relation between hypoglycemia andmortality, even at milder thresholds, such as BG <70 mg/dl. Glycemic variability, notconsidered in the design or implementation of these trials, has also been independentlyassociated with mortality in observational [18–24] and prospective [25] investigations. These findings have led to the emergence of the concept thatthree domains of glycemic control in the critically ill (hyperglycemia, hypoglycemia,and glycemic variability [26, 27]) must be addressed to optimize glycemic control.

These factors, however, may not apply to all patients and, in particular, to those withthe diagnosis of diabetes, presumably related to adaptive mechanisms developed in thesetting of chronic hyperglycemia [28]. Observational cohort studies demonstrated that the relation betweenhyperglycemia and mortality is much stronger among patients without diabetes than inthose with diabetes [3, 29–31], and other observational data suggested that diabetes is not independentlyassociated with increased risk of mortality and may actually have a modest protectiveeffect [32–36].

The purpose of this study was to assess how diabetic status modulates the relation ofthe three domains of glycemic control to mortality in a large and diverse group ofcritically ill patients. We hypothesized that an association would exist betweenmortality and each of the three domains of glycemic control, but that a premorbiddiagnosis of diabetes would attenuate the strength of these associations compared withthose observed in patients without diabetes.

Materials and methods

Patient cohorts and clinical settings

Table 1 provides an overview of the nine different patientcohorts (Amsterdam (AM), Austin (AU), BayCare (BC), Birmingham (BI), Geelong (GE),Okayama (OK), Stamford (ST), Tufts (TU), and Vienna (VI)), the organizationalstructure of the ICUs, and the glycemic-control practices of the differentcenters.

Table 1 Overview of cohorts
Full size table

Outcomes

The primary end point for this analysis was all-cause hospital mortality, defined asdeath before hospital discharge.

Definitions and statistical analysis

Patients were classified as having preexisting diabetes by documentation in theirmedical records. Disease severity was assessed by using APACHE II scores [37]. Descriptive statistics were calculated for all variables of interest.Continuous variables were summarized by using means and standard deviations, whereascategoric variables were summarized by using counts and percentages.

The primary outcome, mortality, was assessed in relation to the glycemic-controlmetric and control variables by using a logistic regression model adjusting forcorrelation among observations taken at the same center (that is, a generalizedestimating equation (GEE) model. Three models were run, one for each glycemicmeasure: hyperglycemia, hypoglycemia, and glycemic variability. The models included avariable denoting diabetic status, the glycemic measure, and the key interaction termof diabetic status and glycemic measure. Each model controlled for mean BG, age,APACHE II score, mechanical ventilation, ICU length of stay (LOS), as well asadjusting for center effects. The models on hyperglycemia and glycemic variabilityalso controlled for hypoglycemia (minimum BG <70 mg/dl). Each model was stratifiedby diagnostic category: medical or surgical. Patients admitted with trauma diagnoseswere included in the surgical cohort.

Before analysis, the set of variables was assessed for the presence ofmulticollinearity. A tolerance statistic less than or equal to 0.4 was considered toindicate the presence of multicollinearity, and in such cases, only one member of acorrelated set would be retained for the multivariable model.

The estimates of each model were presented by using odds ratios and their associated95% confidence intervals. A Bonferroni correction was applied to account for multipletesting. As the greatest number of pairwise comparisons presented for aglycemic-control variable was 10, the standard P value of 0.05 was adjustedto 0.005 to denote statistical significance for all analyses.

Analyses were run by using SAS Version 9.2 (SAS Institute, Cary, NC, USA) and MedCalcV12.4.0.0 (Ostend, Belgium).

The institutional review boards of the different centers approved the investigation.The requirement for informed consent was waived because of the retrospective natureof the study and because the data were deidentified.

Results

In Table 2a and b, we present the considerable heterogeneity inbaseline clinical characteristics of the nondiabetic and diabetic cohorts in the ninedifferent centers. The percentage of patients with diabetes in the different centersranged from 14.0% (AM) to 38.6% (BC).

Table 2 Baseline characteristics, selected outcomes, and details of glycemic control
Full size table

Glycemic control

Patients with diabetes had higher mean BG, higher CV, and higher rates ofhypoglycemia than did patients without diabetes. The nine centers demonstratedconsiderable variation in the frequency of BG monitoring as well as in the intensityof glycemic control, as reflected by mean BG.

Three domains of glycemic control: unadjusted mortality data, nine centers

Mean BG

Figure 1A and 1B displays theunadjusted relation between mean BG and mortality for the nine centers. Additionalfile 1, Table S1 in the online supplement delineates thenumber of patients in each "band" of mean BG, as well as their mean (95%confidence interval (CI)) mortality. Among patients without diabetes, mortalitywas lowest when mean BG was 80 to 110 and 110 and 140 mg/dl and increased athigher levels. The mortality rate of the 200 patients with mean BG <80 mg/dl(0.62% of the total of 32,084 patients without diabetes) was 47.0%. Among patientswith diabetes, the shape of the relation between mean BG and mortality wascharacterized as a shallow trough, with modestly higher mortality in the aggregatewith mean BG 80 to 110 and > 180 mg/dl than with mean BG in the 110- to180-mg/dl range. The mortality rate of the 71 patients with mean BG <80 mg/dl(0.55% of the total of 12,880 patients with diabetes) was 42.3%.

Figure 1
figure 1

Mean blood glucose (BG) and mortality. The relation of mean BG(milligrams per deciliter) during ICU stay to mortality in those without(A) and those with diabetes (B), for each of the ninecohorts as well as the entire population.

Full size image

Hypoglycemia

Figure 2A and 2B illustrates theunadjusted relation between hypoglycemia and mortality. Hypoglycemia wasassociated with increased mortality in patients with diabetes as well as inpatients without diabetes, although the relation was stronger among patientswithout diabetes.

Figure 2
figure 2

Minimum BG and mortality. The relation of minimum BG (milligrams perdeciliter) during ICU to mortality in nondiabetes (A) and diabetes(B) patients, for each of the nine cohorts as well as the entirepopulation. Cohorts with fewer than 20 patients in a particular "band" arenot reported.

Full size image

Glycemic variability

Figure 3A and 3B displays theunadjusted relation between CV and mortality. Among patients without diabetes, therelation between increasing CV and increasing mortality was steep, with more thana threefold higher mortality among the entire cohort with CV > 40%compared with those with CV <20%. This relation was similar, albeit attenuated,among patients with diabetes.

Figure 3
figure 3

Coefficient of variation and mortality. The relationp of coefficientof variation (%) to mortality in nondiabetes (A) and diabetes (B) patients for each of the nine cohorts as well as the entire population.Cohorts with fewer than 20 patients in a particular "band" are notreported.

Full size image

Cumulative derangements in the three domains of glycemic control and theirassociation with mortality

Figure 4A and 4B illustrates thecumulative impact of derangements in the three domains of glycemic control. Amongpatients without diabetes who had mean BG between 80 and 110, 110 and 140, and 140and 180 mg/dl, increasing CV and the occurrence of hypoglycemia were associated withincreased mortality, and their effect was cumulative. Among patients without diabeteswith mean BG > 180 mg/dl, no incremental impact was found of additionalderangements of glycemic control. Among patients with diabetes, hypoglycemia wasconsistently associated with increased mortality, but mean BG and CV did not have aconsistent, cumulative impact on mortality.

Figure 4
figure 4

Cumulative derangements of three domains. The relation of cumulativederangements of the three domains of glycemic control to mortality innondiabetes (A) and diabetes (B) patients. Patients arestratified first by mean BG during ICU stay, then by increasing coefficient ofvariation (CV), and then by the presence or absence of hypoglycemia, defined asminimum BG <70 mg/dl during ICU stay. "Bands" with fewer than 20 patientsare not reported.

Full size image

Multivariable analysis

Figure 5A through F displays the resultsof multivariable analysis, assessing the independent association of bands within eachdomain with mortality.

Figure 5
figure 5

Forest plots of bands of the independent association of mean BG,hypoglycemia, and coefficient of variation to mortality, for diabetes andnondiabetes patients. This figure illustrates the independentassociation of mean BG, hypoglycemia, and coefficient of variation tomortality, for diabetes and nondiabetes patients, including stratificationbased on medical versus surgical status. Pair-wise comparisons of odds ratio(95% CI) for each domain of glycemic control are presented.

Full size image

Mean BG

An effect of center was seen on the relation between mean BG and mortality. Amongpatients without diabetes, mean BG of 110 to 140 mg/dl was independentlyassociated with reduced risk of mortality compared with mean BG of 140 to 180 and> 180 mg/dl, and similar risk compared with mean BG of 80 to 110mg/dl.

The medical and surgical patients demonstrated different patterns. Among medicalpatients, bands of mean BG of 80- to 140-mg/dl range were independently associatedwith the lowest risk of mortality, with increased risk of mortality at higherbands. In contrast, among surgical patients, a mean BG of 80 to 110 mg/dl wasindependently associated with increased risk of mortality compared with bands ofmean BG of 110 to 180 mg/dl.

The relation of mean BG to mortality was somewhat different among patients withdiabetes. Among the entire cohort of patients with diabetes, as well as for bothmedical and surgical subpopulations, mean BG of 80 to 110 mg/dl was independentlyassociated with increased risk of mortality compared with the bands of mean BG of110 to 180 mg/dl, those with mean BG of 110 to 140, 140 to 180, and< 180 mg/dl had a reduced risk of mortality.

Hypoglycemia

Severe (minimum BG <40 mg/dl) and mild to moderate (BG of 40 to 69 mg/dl)hypoglycemia were independently associated with increased risk of mortality, forthe entire cohort, as well as for the medical and surgical subpopulations.

Glycemic variability

Among patients without diabetes, low glycemic variability (CV <20%) wasindependently associated with decreased risk of mortality compared with bands ofCV of 20% to 40% and > 40% for the entire cohort; this relation was morerobust in medical patients than in surgical patients. However, among patients withdiabetes, multivariable analysis demonstrated that increased CV was notindependently associated with increased risk of mortality.

Diabetes

Diabetes was independently associated with decreased risk of mortality for theentire cohort (OR (95% CI)) 0.93 (0.87 to 0.97); P = 0.0030. Figure 6 displays the results of multivariable analysis assessing theindependent association of diabetes with mortality, stratified by individual bandsof the three domains of glycemic control. Among patients with mean BG of 80 to 110mg/dl, diabetes was independently associated with increased risk of mortality forthe entire cohort and the medical subgroup of <80 to > 110 mg/dl.However, for all other bands of mean BG, diabetes was associated with decreasedrisk of mortality for the entire cohort and the medical subgroup. Diabetes was notindependently associated with mortality in the surgical subgroup. Similarly, amongthe entire cohort with hypoglycemia and in the medical subgroup with hypoglycemia,diabetes was independently associated with decreased mortality; diabetes was notindependently associated with mortality among hypoglycemic surgical patients.

Figure 6
figure 6

Forest plots describing the independent association of diabetes withmortality, for each of the three domains of glycemic control. Thisfigure illustrate the independent association of diabetic status withmortality associated with each of the three domains of glycemic control. Forexample, Figure 6a demonstrates that, among patients with mean BG 80 to 110mg/dl, diabetes was independently associated with increased risk ofmortality, but among patients with mean BG of 110 to 140 mg/dl, diabetes wasindependently associated with decreased risk of mortality.

Full size image

Finally, diabetes was independently associated with decreased mortality among theentire cohort and both subgroups in patients with increased glycemic variability,defined as CV > 20%.

Discussion

Salient findings

This multicenter investigation demonstrates clinically important differences betweencritically ill patients with diabetes and patients without diabetes in regard to therelation between the three domains of glycemic control and mortality. Among patientswithout diabetes, the lowest mortality occurred in patients with mean BG of 80 to 140mg/dl. In contrast, among patients with diabetes, mean BG of 80 to 110 mg/dl wasindependently associated with increased risk of mortality compared with patients witha mean BG of 110 to 140, 140 to 180, and even > 180 mg/dl. Hypoglycemia wasindependently associated with increased risk of mortality among patients withdiabetes as well as among those without diabetes. Increased glycemic variability (CV>20%), however, was independently associated with increased risk of mortality amongpatients without diabetes but not among patients with diabetes. Derangements in morethan one domain of glycemic control were associated with cumulative increase inmortality among nondiabetes patients but not among patients with diabetes. Finally,for the entire cohort of 44,964 patients, diabetes was independently associated withdecreased risk of mortality.

Relation to prior literature

Hyperglycemia is associated with increased mortality in the critically ill [2, 3, 14, 29–31]. Increments of mean BG levels above 80 mg/dl are clearly associated withincreasing mortality among patients without diabetes. In contrast, a blunted relationexists between increasing mean BG levels above 80 mg/dl and mortality among patientswith diabetes [3, 29–31]. It is likely that changes in glycemic-control practice over time havealtered the observed relation between mean BG and mortality. The currentinvestigation reflects contemporary practice; all patients were admitted to ICUspracticing at least "moderate" glycemic control; the range of mean BG values of thepatients without diabetes in the different centers (119 to 137 mg/dl) contrastssharply with the mean morning BG of the patients in the control arm of the firstLeuven trial of IIT (153 mg/dl) [4].

Hypoglycemia was the second of the three domains to be associated with increased riskof mortality in critically ill patients. Although most of the literature hasdescribed an independent association of severe hypoglycemia (minimum BG <40 mg/dl)with mortality [12–15, 22], recent observational studies [16, 17] and prospective trial data [11] have identified mild hypoglycemia (minimum BG <70 mg/dl) as beingindependently associated with increased risk of mortality. Our findings confirm theseobservations for patients with and without diabetes.

Glycemic variability was the third of the three domains to be independentlyassociated with mortality in the critically ill [18–25]. One observational study suggested that glycemic variability wasindependently associated with mortality only among critically ill patients withoutdiabetes [24]; our study confirms these findings.

Finally, the independent impact of diabetic status, without reference to glycemiccontrol, on the mortality of critically ill patients has been the subject of recentobservational studies that concluded that patients with diabetes did not experiencehigher mortality, and diabetes may, in fact, be protective [30–36]. We demonstrated here that diabetes is independently associated withdecreased risk of mortality.

Strengths and weaknesses

The clearest strength of this study is its size. The 44,964 patients include patientsadmitted with a large array of medical, surgical, and trauma diagnoses, treated witha variety of glycemic-control protocols, substantially enhancing the generalizabilityof the investigation. Moreover, this is a modern cohort of patients treated in an eracharacterized by attention to glycemic control. Each of the nine centers maintained arobust database characterized by prospective data collection, creating an additionalimportant strength of this investigation: the breadth of demographic, clinicaloutcome, and glycemic data available for analysis. The absence of information aboutinsulin therapy is an important limitation. It is likely that important differencesexist between insulin-treated and insulin-naive patients regarding the relation ofthe three domains of glycemic control to mortality.

Another potential limitation is that the identification of diabetic status was madeon clinical grounds, based on all information available at the time of ICU admission.It is likely that some patients designated as without diabetes may actually have haddiabetes; HgbA1c levels were not obtained routinely, and, of course,glucose-tolerance testing could not be performed. Furthermore, we are unable todetermine whether the diabetes patients were categorized as type I or type II.Although most were likely type II, important differences may exist between the twogroups in their response to derangements in the domains of glycemic control.Additionally, we cannot provide details of nutritional therapy and cannot thereforeanalyze the interactions among glycemic control, nutritional therapy, and insulintreatment of hyperglycemia. Moreover, many of the glycemia data from several of thecenters included in this study were derived from capillary blood measured onpoint-of-care devices, a method associated with increased analytic inaccuracy [38–41]. Nevertheless, any degree of measurement imprecision would only serve todampen the observed relations between glycemia and diabetic status.

Finally, we acknowledge that the observational nature of this investigation mandatesthat its conclusions must be considered to be hypothesis generating, rather thanproof of causality. Nevertheless, it would be unethical to randomize patients toinduced hyperglycemia, hypoglycemia, or increased glycemic variability.

Biological plausibility

Considerable evidence suggests that diabetes may alter the relation between glycemiaand mortality in critically ill patients [28]. Diabetes patients may develop a tolerance to hyperglycemia, and amoderate degree of hyperglycemia that might exert toxicity in a patient withoutdiabetes may be well tolerated in a patient with diabetes. This may explain thestrong relation seen between increasing mean BG levels and mortality in patientswithout diabetes, detailed in several large observational studies, but not amongthose with diabetes [3, 29–31, 36, 42]. In a recent study [43], diabetes patients with poor preadmission glycemic control, reflected byhigh HgbA1c levels, had higher mortality when mean BG was tightly controlled duringICU stay compared with patients with high premorbid HgbA1c levels who had a highermean BG during ICU stay. These intriguing data parallel the results of largeinterventional studies in outpatient populations with type II diabetes [44, 45]. An extensive body of literature has explored the physiological basis ofthe deleterious impact of hypoglycemia [46–51] demonstrated in interventional [4, 6, 11, 25] and observational [12–17] studies; none of these has focused explicitly on the different impact thathypoglycemia may exert on patients with diabetes compared with those withoutdiabetes. Similarly, although various physiological mechanisms underlying the harmfuleffect of increased glycemic variability detailed in interventional [4, 6, 25] and observational [18–24] studies have been proposed [52–56], the reasons that glycemic variability has no or a muted independentassociation with risk of mortality in patients with diabetes compared with thestriking relation seen in patients without diabetes requires furtherclarification.

Clinical implications

The central findings of the current investigation have important implications for thecare of critically ill patients. Hyperglycemia does not have the same associationwith mortality among critically ill patients without diabetes compared with thosewith diabetes. The euglycemic range was independently associated with the lowest riskof mortality among patients without diabetes but with higher mortality among patientswith diabetes. Additionally, important differences were noted when comparing medicaland surgical populations. These findings call into question the "one size fits all"strategy for glycemic control of critically ill patients. It may be most appropriateto establish lower glycemic target ranges for medical patients without diabetes thanfor patients with diabetes or for surgical patients without diabetes.

In addition, our observations call into question the appropriateness of recentlypublished glycemic-control guidelines that recommend a glycemic target range of 140to 180 mg/dl [57] or 140 to 200 mg/dl [58] for all critically ill patients. Furthermore, premorbid glycemic controlin diabetes may have an important impact on the consequences of glycemic control inthe ICU [43]. The optimal glycemic-control protocol may result not only fromstratifying patients by diabetic status, but also by additionally stratifyingpatients with diabetes based on the degree of preadmission glycemic control. Incontrast, the deleterious association of hypoglycemia with mortality, even atthreshold levels of <70 mg/dl, was observed in patients with diabetes and in thosewithout diabetes. Because hypoglycemia can never be the subject of a randomizedtrial, the data from this investigation, when combined with the findings fromprevious interventional [4, 6, 10, 11, 25] and observational [12–17] studies, provide the strongest evidence basis for the goal of avoidinghypoglycemia in all critically ill patients.

Finally, increased glycemic variability, defined as CV > 20%, was identifiedin the current study as having a strong independent association with increased riskof mortality in patients without diabetes. These data provide strong impetus for thecreation of insulin-dosing strategies and the development of new technologies [59] for accurate continuous or near-continuous BG monitoring, with the goal ofreducing glycemic variability in critically ill patients. Further investigationshould stratify patient outcomes by specific admitting diagnosis; importantdifferences may be found within the broad medical and surgical categories that thecurrent investigation was underpowered to assess.

The design of future trials of IIT should include consideration of all three domainsof glycemic control as well as recognition of the differences in their associationwith mortality based on premorbid diabetes status.

Conclusions

This large international cohort study evaluated the relation of diabetic status to theassociation of hyperglycemia, hypoglycemia, and increased glycemic variability in aheterogeneous population of critically ill patients. We found that diabetic statusmodulates the relation between the three domains of glycemic control and mortality inclinically important ways. Our findings suggest that patients with diabetes may benefitfrom higher glucose target ranges than those without diabetes. Additionally,hypoglycemia is independently associated with increased risk of mortality, regardless ofthe patient's diabetic status, and increased glycemic variability is independentlyassociated with increased risk of mortality among patients without diabetes. Thesefindings may inform the implementation of glycemic-control protocols in the intensivecare unit, as well as for the design of future interventional trials of intensivemonitoring and treatment of dysglycemia in the critically ill.

Key messages

  • Diabetic status modulates the relation between the three domains ofglycemic control (hyperglycemia, hypoglycemia, and glycemic variability) and mortalityin critically ill patients in clinically important ways.

  • The range of mean BG from 80 to 140 mg/dl is associated with thelowest severity adjusted mortality among nondiabetes patients. In contrast, among thosewith diabetes, a mean BG of 80 to 110 mg/dl is associated with higher mortality riskthan is the range of 110 to 180 mg/dl.

  • A single episode of hypoglycemia (BG <70 mg/dl) is independentlyassociated with increased risk of mortality among those without as well as those withdiabetes.

  • Increased glycemic variability, defined as CV > 20%, isindependently associated with increased risk of mortality among those without, but notamong those with diabetes.