Benefits and Risks of Tight Glucose Control in Critically Ill Adults:  A Meta-analysis FREE

In 2001 van den Berghe et al¹ published a randomized controlled trial of critically ill surgical patients showing that tight glucose control reduced hospital mortality by one-third. Since the greatest decrease in deaths occurred in the subgroup of patients with sepsis and multisystem organ failure, some speculated that the benefits of tight glucose control might extend to medical intensive care unit (ICU) patients as well.²

Because few interventions in critically ill adult patients reduce mortality to this extent, the results of this trial were enthusiastically received and rapidly incorporated into guidelines. In 2004, the Surviving Sepsis Campaign³ recommended glucose control for all patients with sepsis and explicitly stated, “There is no reason to think that these data are not generalizable to all severely septic patients.” This recommendation persists in the 2008 update, now endorsed internationally by 16 professional societies.⁴ In addition, the Institute for Healthcare Improvement,⁵ the Volunteer Hospital Association,⁶ the Michigan Health and Safety Coalition,⁷ the American Association of Clinical Endocrinologists,⁸ and the American Diabetes Association⁹ now recommend tight glucose control in all critically ill adults. These recommendations have led to worldwide adoption of tight glucose control in a variety of ICU settings.^10- 14

Subsequent large randomized controlled trials of tight glucose control in medical and mixed medical-surgical ICU settings,^15- 17 however, have failed to replicate this mortality benefit. Moreover, a recent cohort study¹⁸ of more than 10 000 critically ill adults showed a trend toward increased mortality with increasing use of tight glucose control after adjustment for disease severity. In addition, many studies have reported high rates of hypoglycemia with tight glucose control—some as high as 30% to 40%, as compared with the 5% rate found in the initial trial by van den Berghe et al.¹ Hypoglycemia is not benign in critically ill patients; it has been linked to serious neurologic events ranging from seizures to coma.^19- 20

Consequently, considerable controversy has emerged as to whether tight glucose control is warranted in all critically ill adults. We report the findings of a meta-analysis of randomized controlled trials examining the risks and benefits of tight glucose control as compared with usual care in critically ill adults. In addition to the overall analysis, we conducted subgroup analyses on 2 variables that have been debated in the controversy over tight glucose control: glucose goal (≤110 mg/dL or <150 mg/dL) and ICU setting (medical, surgical, or all critically ill patients).

We searched MEDLINE (1950-June 6, 2008) to identify studies in any language relevant to our research question. We used exploded Medical Subject Headings in the following search strategy: intensive care units or critical care or critical illness or postoperative care or sepsis or myocardial infarction or stroke or cardiovascular surgical procedures, or wounds and injuries; and blood glucose or insulin (administration and dosage, adverse effects, therapeutic use, therapy). We combined the findings of this search with phases 1 and 2 of a highly sensitive search strategy²¹ recommended by the Cochrane Collaboration for identifying all randomized controlled trials in MEDLINE. Using similar search terms, we also searched the Cochrane Library (issue 1, 2008) and multiple trial registries (all in August 2007) including clinicaltrials.gov (National Institutes of Health), the Current Controlled Trials registry (which has the capacity to search the International Standard Randomized Controlled Trial Number registry and 12 other trial registries), the Australian New Zealand Clinical Trials Registry, and Japan's University Hospital Medical Information Network Clinical Trial Registry. We manually searched abstracts from the conference proceedings of the American Thoracic Society (2001-2008) and the Society of Critical Care Medicine (2004-2008). In addition, we reviewed reference lists of relevant articles to identify any additional studies overlooked by our search.

Inclusion Criteria. We included randomized controlled trials that met each of the following criteria: (1) the setting was an adult ICU; (2) the intervention group received tight glucose control (glucose goal <150 mg/dL obtained using an insulin infusion during part or all of the ICU stay); (3) the comparison group received usual care (glucose goal and method of insulin administration could vary between studies); and (4) the primary or secondary end points included hospital or short-term mortality (≤30-day), septicemia, new need for dialysis, or hypoglycemia. To convert glucose values to mmol/L, multiply by 0.0555.

Exclusion Criteria. Studies were excluded if the intervention was conducted primarily during the intraoperative period rather than during the ICU stay or if we were unable to obtain adequate details of study methodology or results from the article or study investigators.

Missing Data. We contacted the investigators of all unpublished studies as well as any published studies in which data were missing to confirm eligibility and obtain additional study details.

Two unblinded reviewers (R.S.W. and D.C.W.) independently assessed and abstracted pertinent data from trials in duplicate using a standardized, predefined form (available from authors). Abstracted data included each study's methodology, setting, baseline patient characteristics, intervention, outcomes, and follow-up. We formally assessed the methodologic quality of each trial using the Jadad scale,²² which incorporates randomization, blinding, and attrition to derive a score of 0 to 5, with higher scores indicating higher quality. Any discrepancies between the 2 reviewers were resolved through discussion. For 4 studies that were presented at meetings but not yet published, the authors provided either the unpublished data or manuscripts.^23- 26 For 2 additional studies that were presented at meetings^17,27 and for 2 unpublished studies, the authors completed a standardized data abstraction form (J. R. A. Azevedo et al, January 2008, and R. P. C. Chan et al, July 2007).

Primary Outcome Measure: Hospital Mortality. We considered a reduction in hospital mortality to be the most important potential benefit of tight glucose control. Hospital mortality was defined as death occurring during the hospital stay or within 30 days following admission.

Secondary Outcome Measures: Septicemia, New Need for Dialysis, and Hypoglycemia. We compared the association of tight glucose control vs usual care with 2 additional potential benefits of tight control: rates of septicemia and new need for dialysis. Quiz Ref IDThese outcomes were chosen because they have biological plausibility, given the association of uncontrolled hyperglycemia with recurrent infection and chronic renal insufficiency in diabetic patients, and because they were shown to be reduced in the initial trial by van den Berghe et al.¹ We defined septicemia to encompass the terms sepsis, septicemia, bacteremia, or a description of positive blood cultures; a general description of infection did not qualify. New need for dialysis referred specifically to patients without a preexisting dialysis requirement who subsequently developed acute renal failure that required dialysis. Specific criteria for determining the need for dialysis were not reported; however, those patients with an increase in serum creatinine without the need for dialysis were not included in this definition.

Hypoglycemia is the major potential harm of tight glucose control. We defined hypoglycemia to include patients with 1 or more blood glucose measurements of 40 mg/dL or lower and recorded whether any associated symptoms were reported. Of note, our definition is well below the glucose level that the American Diabetes Association considers to represent hypoglycemia (glucose <70 mg/dL).⁹ We chose this strict definition to capture hypoglycemic events severe enough to have potential clinical relevance, whether or not concurrent symptoms occurred, and because a glucose level of 40 mg/dL or lower was the most common definition of hypoglycemia used in the included trials.

All outcome measures were calculated on a per-patient basis; for example, a patient with several episodes of hypoglycemia would only count as 1 occurrence for that outcome.

A priori we identified 2 variables for subgroup analysis based on the main controversies in the debate surrounding tight glucose control: glucose goal and ICU setting.

Glucose Goal in the Tight Control Group. Differing opinions exist about the optimal level of tight glucose control. Based on the 2008 recommendations for glucose control in critically ill patients from the American Diabetes Association⁹ (as close to 110 mg/dL as possible) and the Surviving Sepsis Campaign⁴ (<150 mg/dL), we stratified studies by glucose goal in the tight glucose control group into 2 categories: very tight control (upper limit of glucose goal ≤110 mg/dL); and moderately tight control (upper limit of glucose goal 111-150 mg/dL).

ICU Setting. Because of the concern that the pathophysiological effect of hyperglycemia may differ between surgical and medical critically ill patients, we stratified trials by ICU setting into 3 categories: (1) surgical (including general surgical, cardiothoracic, neurosurgical, and trauma ICUs); (2) medical (including general medical, cardiac, and neurologic ICUs); or (3) mixed medical-surgical ICUs. For those trials that did not specify the ICU setting,^28- 31 we categorized the setting as medical-surgical.

We performed sensitivity analyses based on 3 prespecified clinically relevant variables: proportion of diabetics, use of insulin-only infusions, and achieved mean glucose level in the study groups. Since individuals with diabetes vs those without diabetes may differ in whether hyperglycemia is a maladaptive response that should be treated, we restricted analysis to trials in which one-third or less individuals had diabetes (an arbitrary cut point based on a natural break in the distribution of proportion of those with diabetes in the included trials). We restricted analysis to trials using insulin-only infusions (as opposed to glucose-insulin-potassium infusions), because these interventions may have different effects. We restricted analysis to studies in which the mean glucose level achieved in the tight glucose control and usual care groups differed by at least 20 mg/dL (a difference that we specified a priori to be clinically meaningful), because studies that failed to achieve a clinically significant difference in glucose levels between study groups might have biased our results toward the null. In addition, for the subgroup analysis of very tight vs moderately tight glucose control, we performed an analysis in which studies were categorized based on actual mean glucose level achieved rather than target glucose goal because in many studies, the target glucose goal and achieved glucose level in the tight control group were disparate.

We used the analytic approach and software provided by the Cochrane Collaboration for all analyses (Review Manager [RevMan] version 4.2, Nordic Cochrane Centre, Copenhagen, Denmark). This software calculates relative risks (RRs) for studies with at least 1 occurrence in either study group for each outcome. Trials with missing outcome data or zero occurrences in both groups were excluded from the meta-analysis of that outcome. We calculated a pooled RR and 95% confidence interval (CI) for each outcome and considered findings to be statistically significant if the test for overall effect had a P value of less than .05.

For each outcome, we assessed for important variability among the trial results contributing to each summary estimate using 2 thresholds based on the χ² test. We considered a P value of less than .10 to indicate statistically significant heterogeneity. Because some heterogeneity is inevitable in meta-analysis, some argue that rather than assessing its statistical significance, investigators should assess its effect. One such method is to assess I², which quantifies the proportion of the variability in trial results that is due to heterogeneity rather than chance and uses a value greater than 50% to indicate meaningful heterogeneity. In our meta-analysis, if either threshold for variability was met, we identified each responsible trial and reviewed its clinical and methodological characteristics to determine whether an explanation for the outlying results existed. For each such case we report 2 summary estimates: (1) an estimate based on all studies with usable data, including the outlying trial(s); and (2) an estimate based on the largest group of studies with usable data that passed both the P value and I² thresholds.

There is disagreement about whether fixed- or random-effects models are preferred when calculating summary estimates for meta-analyses (Cochrane Handbook for Systematic Reviews of Interventions version 5.0; available at http://www.cochrane-handbook.org). While fixed-effects models typically result in narrower CIs, the 2 models tend to provide similar results unless heterogeneity is present among the included studies. We believe the random-effects model is more appropriate for meta-analyses that evaluate the efficacy of an intervention, particularly for analyses with important downsides, because it reduces the risk of a type I error. Moreover, when heterogeneity is present, the random-effects model is recommended by the Cochrane Collaboration because its assumptions account for the presence of variability among included trials. Therefore, we report the results of the random-effects model for all outcomes. In addition, if no heterogeneity existed among studies, we have provided results of the fixed-effects model for situations in which the 2 models yielded substantially different findings.

We visually assessed a funnel plot of study size vs effect size for our primary outcome of hospital mortality to seek evidence of publication bias.

We initially identified 1358 potentially eligible studies (Figure 1), the majority of which were excluded because they were ongoing, were not randomized controlled trials, or tested an intervention other than tight glucose control. After detailed review of the remaining 115 randomized controlled trials, 34 (including the 2 unpublished studies) met all inclusion criteria and were considered as potentially appropriate for inclusion in our meta-analysis.^{1,15- 17,23- 50} Subsequently, we excluded 1 trial³⁴ because we were unable to confirm full study details despite multiple attempts to contact the investigators, and another 4 trials^28,32,40,42 because they reported zero events in both study groups for all outcomes relevant to our analysis. This left 29 randomized controlled trials (19 full publications,^{1,15- 16,29- 31,33,35- 39,41,43- 46,48- 49} 8 published in abstract form only,^{17,23- 27,47,50} and 2 unpublished studies) including 8432 patients with usable data for our meta-analysis.

Table 1 provides the characteristics of the 34 randomized controlled trials that met our inclusion criteria. Trials were conducted in a diverse array of countries, most often at a single center. Study sizes ranged widely (10->1500 patients) with 21 trials enrolling fewer than 100 patients and 7 trials with more than 500 patients. The study participants encompass a broad distribution of adult ICU patients, as indicated by the variety of mean ages (46-75 years), distributions by sex (31%-95% men), proportions of patients with diabetes (0%-100%), and degree of disease severity as measured by mean Acute Physiology and Chronic Health Evaluation (APACHE II) score (9-32). Only 2 studies^29,46 had discrepant baseline patient characteristics in the intervention vs control groups. In both of these trials, disease severity was lower in the tight glucose control than usual care group (mean APACHE II score 14 vs 17 in the trial by Wang et al²⁹; and 19 vs 22, P < .01 in the trial by Mitchell et al⁴⁶). All trials had follow-up rates of 80% or greater. Because none of the trials attempted to double-blind study group assignment, no trial could receive a Jadad quality score higher than 3 out of 5. Target glucose goals, as well as mean achieved glucose levels, varied between trials in both the tight control and usual care groups (Table 2).

Twenty-seven trials, including the 2 unpublished ones, provided usable data on hospital mortality.^{1,15- 17,23- 27,29- 30,35- 39,41,43- 50} Among these trials, there was no significant difference in hospital mortality between tight glucose control and usual care strategies (21.6% vs 23.3%; RR, 0.93; 95% CI, 0.85-1.03; Figure 2).

We also performed subgroup analyses stratifying trials by ICU setting and by glucose goal in the tight control group. There was no significant difference in hospital mortality when we stratified by surgical (8.8% vs 10.8%; RR, 0.88; 95% CI, 0.63-1.22), medical (26.9% vs 29.7%; RR, 0.92; 95% CI, 0.82-1.04), and medical-surgical ICU setting (26.1% vs 27.0%; RR, 0.95; 95% CI, 0.80-1.13; Figure 2). Similarly, there was no significant difference in hospital mortality between tight glucose control and usual care strategies when we stratified by glucose goal in the tight control group (very tight [23.2% vs 25.2%; RR, 0.90; 95% CI, 0.77-1.04]; and moderately tight [17.3% vs 18.0%; RR, 0.99; 95% CI, 0.83-1.18]; Figure 3). Tests for heterogeneity identified the trial by Wang et al²⁹ as having outlying results for both the subgroups of trials in the medical-surgical ICU (P = .03, I² = 48%) and trials of very tight glucose control (P = .02, I² = 49%), which appeared to be explained by the previously mentioned discrepancy in baseline disease severity. Exclusion of the outlying trial²⁹ resolved this heterogeneity but did not significantly change the findings of either subgroup analysis (medical-surgical ICU [RR, 1.00; 95% CI, 0.90-1.11], very tight control [RR, 0.94; 95% CI, 0.84-1.05]).

The associations of tight glucose control on all of the secondary outcomes analyzed and stratified by glucose goal in the tight control group are shown in Figure 3. The associations between tight glucose control and all outcomes, stratified by ICU setting are shown in Figure 4 (data supporting these outcomes are available from the authors on request).

Rates of septicemia were reported in 9 trials.^{1,15,23,26,30,35- 37,48} Tight glucose control was associated with significantly reduced risk of septicemia as compared with usual care (10.9% vs 13.4%; RR, 0.76; 95% CI, 0.59-0.97). When stratified according to ICU setting, this reduction in septicemia was limited to surgical ICU patients^1,23,35- 37 (4.6% vs 8.4%; RR, 0.54; 95% CI, 0.38-0.76) and was not observed in medical¹⁵ or medical-surgical ICU patients^26,30,48 (Figure 4). When stratified by glucose goal in the tight control group (Figure 3), there was a nearly significant reduction in septicemia that was limited to studies using moderately tight glucose control (8.8% vs 14.6%; RR, 0.64; 95% CI, 0.41-1.00). The test for heterogeneity was significant only for the subset of trials using very tight glucose control (P = .04; I² = 64%). We identified the surgical ICU trial by van den Berghe et al as the outlying study,¹ but could not find an obvious reason for the outlying results. Analysis limited to the studies with homogeneous results did not change our finding of nonsignificant reduction in septicemia for the subset of trials on very tight glucose control.

New need for dialysis was reported in 8 published trials^{1,16,23,25- 26,30,35,50} and in 1 that was unpublished (Azevedo). There was no significant association between tight glucose control and a new need for dialysis overall (11.2% vs 12.1%; RR, 0.96; 95% CI, 0.76-1.20). Subgroup analyses stratifying by glucose goal (Figure 3) and ICU setting (Figure 4) also showed no significant association of tight glucose control with new need for dialysis. The test for heterogeneity was significant (P = .06; I² = 55%) only for the subgroup of trials evaluating very tight glucose control; again, the van den Berghe surgical ICU trial¹ was the outlying study. Regardless of whether the outlying trial was excluded or not, the findings were not statistically significant (Figure 3). Conversely, while the subgroup of trials^1,23,35 conducted in the surgical ICU (Figure 4) exceeded both thresholds of heterogeneity, the summary estimates differed based on the model used. While the findings of the random-effects model were not significant (RR, 0.69; 95% CI, 0.38-1.26), using a fixed-effects model resulted in a significant reduction in new need for dialysis (RR, 0.64; 95% CI, 0.45-0.92).

Hypoglycemia was reported in 14 published trials^{1,15- 17,23,25- 26,31,33,37- 38,46- 47,50} and in 1 that was unpublished (Azevedo). Tight glucose control was associated with an increased risk of hypoglycemia (13.7% vs 2.5%; RR, 5.13; 95% CI, 4.09-6.43). As would be expected, when compared with usual care, the risk of hypoglycemia was higher with patients receiving very tight glucose control than for those with moderately tight glucose control (Figure 3). The increased risk of hypoglycemia was fairly consistent across ICU settings (Figure 4). Trials that were conducted in the medical ICU indicated heterogeniety (I², 51%), but with only 2 trials^15,38 in this subgroup, we cannot determine which trial is the outlier. Nonetheless, we would judge the trial³⁸ of 8 patients reporting no increased risk of hypoglycemia to be less reliable than the larger study,¹⁵ which found a significantly increased risk. While most trials reported that few or none of the hypoglycemic events were associated with overt symptoms, some studies found that patients who experienced hypoglycemia had a higher risk of death.^15- 16,26

For each of our sensitivity analyses (restricting to trials with 34% of patients with diabetes or fewer, trials using insulin-only infusions, trials that achieved mean glucose levels that differed by at least 20 mg/dL between study groups, and stratifying trials by actual glucose level achieved in the tight glucose control group), the point estimates for most outcomes changed minimally. Those point estimates with moderate changes remained within wide confidence intervals. Only 2 findings changed in statistical significance when stratified by actual glucose level achieved: the reduction in septicemia became statistically significant in the subgroup of trials of very tight glucose control (RR, 0.58; 95% CI, 0.42-0.80); whereas in trials of moderately tight glucose control, the reduction in septicemia was no longer significant (RR, 0.87; 95% CI, 0.63-1.21).

Upon visual inspection of the funnel plot for hospital mortality, we found no evidence of publication bias (data not shown).

Quiz Ref IDIn this meta-analysis of randomized controlled trials of tight glucose control vs usual care in critically ill adults, we found no significant difference in hospital mortality or new need for dialysis. Although tight glucose control was associated with a significant reduction in septicemia overall, subgroup analysis suggested this benefit was limited to surgical ICU patients. On the other hand, we found clear evidence of the main harm of tight glucose control: hypoglycemia increased roughly 5-fold, regardless of the ICU setting, and was more common with patients receiving very tight than moderately tight glucose control. In short, our meta-analysis does not support the benefits of tight glucose control reported in the initial trial by van den Berghe et al,¹ yet it suggests a much higher risk of hypoglycemia.

Our study has several limitations. Since we have pooled results from individual trials, our analysis is limited by any flaws in the methodology of these underlying trials. Quiz Ref IDAlthough none of the included trials attempted to double-blind study group assignments, which could have introduced bias if patients were treated differently based on knowledge of their assignment, all trials achieved a good balance in the relevant baseline characteristics except as noted, and all had greater than 80% follow-up. The potential for differential treatment in these unblinded studies is most relevant for the outcome of new need for dialysis, which is likely to be determined at least in part subjectively, by the treating physicians.

Although several of the included studies were small, the main limitations of such trials (lack of power and narrow generalizability) would be attenuated by inclusion in a meta-analysis, and exclusion of such studies could introduce bias. Despite the increased power derived from pooling many studies, our meta-analysis may still be underpowered to detect small differences in outcomes between tight glucose control and usual care strategies. For example, for the primary outcome of hospital mortality, our meta-analysis is powered (assuming 2-sided alpha = .05 and power = 0.8) to determine statistical significance of a 2.6% absolute difference in mortality (20.7% vs 23.3%) but not the 1.7% difference in mortality that we actually identified (21.6% vs 23.3%). To be powered to establish statistical significance of a 1.7% difference between groups would require an estimated 19 146 patients.

Trials that were included in our meta-analysis varied widely with regard to baseline patient characteristics and insulin infusion protocols. However, this diversity, which did not appear to influence tests for heterogeneity except as noted, allowed us to capture the full scope of critically ill adults and ICU processes of care. Furthermore, we did not find major changes in our results when we performed sensitivity analysis based on variables of potential clinical relevance, which suggests further support for combining the broadly representative studies. Nonetheless, because it remains reasonable to expect that our overall negative findings might contain important subgroups that would benefit from tight glucose control, we present the findings stratified by the most widely debated variables—glucose goal in the tight control group and ICU setting.

Our meta-analysis shows that subsequent trials have not borne out the impressive results of tight glucose control promised by the initial trial by van den Berghe et al.¹ Tests of heterogeneity identified this trial as having outlying results when compared with subsequent randomized controlled trials reporting outcomes of septicemia and new need for dialysis. There are at least 3 reasons the results of this trial may differ from subsequent studies: bias, chance, and atypical clinical practices. Although this trial was not blinded, which could lead to bias, none of the other trials were blinded and we doubt this explains the discrepancy between study results. Quiz Ref IDThe initial trial by van den Berghe et al,¹ reported unusually high mortality in the usual care group based on the disease severity, a finding which may be due to chance. Moreover, several aspects of this trial have been criticized^51- 53 for using atypical clinical practices. Specifically, the use of early glucose infusion and parenteral nutrition, both of which may artificially induce hyperglycemia, may have contributed to the outlying results seen in this trial.

The surgical population or even more specifically cardiac surgery patients who comprised the bulk of the patients in the initial trial by van den Berghe et al,¹ may represent the group most likely to benefit from tight glucose control. However, our meta-analysis demonstrates that subsequent randomized controlled trials of this intervention in surgical patients have not confirmed a significant reduction in mortality, a finding supported by a subgroup analysis of surgical patients (n = 6431) in a recent large cohort study.¹⁸ Furthermore, subsequent randomized controlled trials of tight glucose control during⁵⁴ or after (unpublished data from Chan et al) cardiac surgery have also failed to confirm a reduction in mortality with tight glucose control.

Practical problems implementing tight glucose control have occurred both inside and outside of clinical trial settings. Actually achieving the target glucose goal can be difficult; even under the close supervision of a clinical trial, 4 published and 2 unpublished of the 29 studies (21%) in our meta-analysis did not achieve a mean glucose level within 5 mg/dL of the stated goal in the tight control group^{17,24- 25,41} (and unpublished data from Chan et al and Azevedo et al). Quiz Ref IDIn practice, there has been substantial resistance to full adherence with tight glucose control by nursing staff, due both to the increased workload stemming from the need for frequent glucose monitoring and changes in infusion rate and to concerns about risk of hypoglycemia.^53,55- 56 These concerns of hypoglycemia appear to be warranted, as indicated by the significant increase in risk of hypoglycemia in our meta-analysis. Whether these hypoglycemic events are a causal factor in these patients' deaths or simply a marker of disease severity is unknown.

Overall, we believe the 29 trials included in our meta-analysis allow us to draw conclusions about the benefits and risks of tight glucose control in the broad spectrum of critically ill adults. We found that tight glucose control was not associated with a significant reduction in hospital mortality or in new need for dialysis, but was associated with a markedly increased risk of hypoglycemia. Although we found a statistically significant association with reduction in septicemia, the reduction may have been in less severe episodes of septicemia, given the lack of an associated reduction in hospital mortality. Moreover, when stratified by ICU setting, the significant association with reduced risk of septicemia was limited to trials conducted in the surgical ICU. Given the overall findings of this meta-analysis, it seems appropriate that the guidelines recommending tight glucose control in all critically ill patients should be reevaluated until the results of larger, more definitive clinical trials are available.