Introduction

Blood glucose management has been considered important in critically ill patients [1], since the pioneering study by van den Berghe et al. reported that intensive insulin therapy (IIT; target blood glycemic range, 80–110 mg/dL) reduced mortality and morbidity in surgical intensive care unit (ICU) patients (63 % were post-cardiac surgical patients) compared with a more liberal target range (180–200 mg/dL) [2]. Subsequently, several studies reported that tight glycemic control reduced mortality and morbidity at certain centers and in selected populations [3, 4]. A previous meta-analysis also reported that IIT decreased mortality and the risk of septicemia in surgical ICU patients [5, 6]. Furthermore, hyperglycemia might be an independent risk factor for ICU-acquired weakness [7]. Therefore, lowering the blood glucose level might be beneficial in critically ill patients.

On the other hand, IIT is frequently reported to increase the risk of hypoglycemia [2–5]. Although there is no evidence to suggest that such acute hypoglycemia would directly worsen patient outcomes, the increased risk of hypoglycemia during tight glycemic control in critically ill patients might result in undesirable adverse effects, as even moderate hypoglycemia has been associated with poor outcomes [8]. In 2009, the landmark Normoglycemia in Intensive Care Evaluation and Surviving Using Glucose Algorithm Regulation (NICE-SUGAR) trial reported that IIT (target blood glycemic range, 80–108 mg/dL) increased mortality among critical ill patients compared with an intermediate level of glycemic control (144–180 mg/dL) [9]. Therefore, several clinical guidelines do not recommend IIT, with some, including the Surviving Sepsis Campaign guidelines, recommending that insulin therapy be commenced for blood glucose levels of >180 mg/dL [10], while others have recommended that insulin therapy be initiated at blood glucose levels of >150 mg/dL [11].

This divergence among guidelines may be due to the lack of a meta-analysis comparing the risks or benefits of diverse glycemic targets, including <110, 110–144, 144–180, and >180 mg/dL. Notably, all prior meta-analyses compared IIT with insulin therapy targeting higher glycemic levels [5, 6, 12]. Recently, the methodology for network meta-analyses was established and it was made statistically possible to compare groups even when a direct comparison was not originally performed [13, 14]. The integration of direct evidence with indirect evidence increases the precision of the estimates and produces a relative ranking of all treatments for the studied outcome [15]. The assumption of consistency underlies the methodology and, if it holds, a network meta-analysis can provide valuable information [15]. Therefore, we considered that a network meta-analysis could provide comparisons among four glycemic targets, which was not possible using a conventional meta-analysis. Accordingly, we performed a network meta-analysis to identify the optimal blood glucose target in terms of balance of the benefit (the reduction of mortality and incidence of infection) and the risk (incidence of hypoglycemia) of insulin therapy among critically ill patients.

Methods

Search strategy

For the network meta-analysis, we considered all studies from three recent systematic reviews [5, 6, 12]. In addition, we searched the PubMed and Cochrane Library databases (6 June 2008 to 25 February 2016) for studies that were published after the meta-analysis by Wiener et al. [5], using the following search terms: “blood glucose” or “insulin” (administration and dosage, adverse effects, therapeutic use, therapy) AND “intensive care unit” or “critical care” or “critical illness” or “postoperative care” or “sepsis” or “myocardial infarction” or “stroke” or “cardiovascular surgical procedures” or “wounds and injuries”. We also evaluated the reference lists of the relevant clinical trials to identify additional studies.

Study selection

We included randomized controlled trials (RCTs) that fulfilled the following criteria: (1) full-text publication in English; (2) included adult patients who were treated in the ICU; (3) compared different glycemic targets; and (4) the outcome measures included the incidences of hypoglycemia, sepsis and bloodstream infection, or mortality. Studies were excluded if they did not meet the above four criteria. We also reviewed the full texts of all studies obtained from our search of the PubMed and Cochrane databases and the studies from the three systematic reviews.

Data abstraction and quality assessment

Three reviewers (TY, SI, and ME) independently abstracted the data and assessed the methodological quality of the eligible studies. Disagreements were resolved via consensus after a group discussion. The abstracted data for each study included the first author’s name, year of publication, number of study sites, number of patients, ICU setting, patients’ age, history of diabetes, target blood glucose levels, mean glucose levels during the study period, presence of a protocol, method of glucose measurement, duration of the study, and the mean length of stay in the ICU. Methodological quality was evaluated using the Cochrane risk of bias assessment tool, which assesses randomization, allocation concealment, blinding of the study participants and personnel, blinding of the outcome assessments, incomplete outcome data, selective outcome reporting, and other potential sources of bias [16].

Outcome measures and categories of target blood glucose levels

For the analyses, we defined the primary outcome as hospital mortality. The secondary outcomes were defined as the risk of sepsis or bloodstream infection, and hypoglycemia (defined as blood glucose levels of <40 mg/dL or hypoglycemia-associated symptoms). If hospital mortality data were not available, we substituted the 90-day mortality for hospital mortality. Cases of sepsis or bloodstream infection were defined as patients with sepsis, bacteremia, or positive blood cultures.

Blood glucose levels were categorized on the basis of the upper limit of the blood glucose target: <110, 110–144, 144–180, and >180 mg/dL. For example, if the target blood glucose range was 80–130 mg/dL, it was categorized as 110–144 mg/dL.

Subgroup analyses and additional analysis using actually achieved mean glycemic levels

We conducted subgroup analyses according to the ICU setting, proportion of diabetic patients, and observed mortality. We also performed an additional analysis using achieved mean glycemic levels during the study period.

ICU setting

We divided studies according to the ICU setting. First, we classified the ICU setting in each study into three categories, according to a previous systematic review [5]: (1) surgical ICU, including general-surgical, cardiothoracic-surgical, neurosurgical, and trauma ICUs; (2) medical ICU, including general medical, cardiac, and neurological ICUs; and (3) mixed medical-surgical ICUs. Next, we stratified the mixed medical-surgical ICUs by the proportion of surgical patients. If surgical patients constituted ≤50 % of the total cohort in a mixed medical-surgical ICU, it was defined as a medical ICU; if surgical patients in a mixed medical-surgical ICU constituted >50 % of the total cohort in the ICU, it was defined as a surgical ICU. Finally, we performed a subgroup analysis of the ICU setting for surgical ICUs and medical ICUs separately.

Proportion of diabetic patients

We stratified studies according to the proportion of diabetic patients: ≤50 or >50 %.

Observed mortality

We performed a subgroup analysis of the patients with >20 % observed mortality.

Additional analysis using actual mean glycemic levels

We performed an additional analysis based on the actually achieved mean glycemic levels.

Statistical analysis

We performed the network meta-analysis within a Bayesian framework using JAGS (version 4.1.0), R software (version 3.1.1), and the rjags and gemtc packages [17, 18]. Comparative odds ratios (OR) are reported with their 95 % credible intervals (CrI). Both fixed-effect and random-effect models were used, and the deviation information criterion (DIC) was used to compare the treatment effects of the two models. The DIC measures the fit of the model, while penalizing it for the number of effective parameters; the most appropriate model has the lowest DIC value [19]. We selected the random-effect model based on the DIC values for the two models. Furthermore, a Bayesian framework meta-analysis provided a rank probability for each glycemic target range and outcome. We used these results for our interpretation. In addition, factors such as the total number of trials in a network, the number of trials with more than two comparison arms, heterogeneity (i.e., clinical, methodological, and statistical variability within direct and indirect comparisons), inconsistency (i.e., a discrepancy between direct and indirect comparisons), and bias may influence effect estimates obtained from network meta-analyses [20]. The inconsistency was assessed by using Bayesian P values by the node splitting analysis of rjags the gemtc packages. The value of I 2 ≥ 50 % would be considered substantial heterogeneity [21, 22].

Results

We identified 39 studies through the re-analysis of the three systematic reviews [5, 6, 12], 431 references through the PubMed search, and 539 references through the Cochrane search. Fifty-seven studies were considered potentially eligible (Supplementary appendix 1). We excluded 22 studies after the full-text screening (4 conference abstracts, 7 non-English studies, and 11 studies that did not fulfill the eligibility criteria) (Supplementary appendix 2). Accordingly, the current network meta-analysis included 18,098 patients from 35 studies [2, 4, 9, 23–54].

As Table 1 shows, there was a large heterogeneity in the characteristics of the included studies. The mean age of the patients varied from 38.5-years-old to 74.9-years-old. The proportion of patients with diabetes varied from 1.3 to 100 %. Table 1 also provides the mean glucose levels during the study period for each group of studies. The risk of bias for each study is shown in the Supplementary appendix 3.

Table 1 Characteristics of included randomized control trials
Full size table

The network of eligible comparisons for the meta-analysis is shown in Supplementary appendix 4. The Bayesian P value and I 2 value in each comparison are shown in Supplementary appendix 5. The forest plot for hospital mortality is shown in Fig. 1a. Although the target glucose concentration of 144–180 mg/dL was associated with a significantly lower hospital mortality compared with the target concentration of >180 mg/dL in a direct comparison (OR 0.1; 95 % CrI 0.01–0.71), there were no significant differences among all comparisons of the network meta-analysis (<110 vs. 110–144 mg/dL, <110 vs. 144–180 mg/dL, <110 vs. >180 mg/dL, 110–144 vs. 144–180 mg/dL, 110–144 vs. >180 mg/dL, and 144–180 vs. >180 mg/dL) (Fig. 1a; Table 2). The rank probabilities for the four blood glucose targets are shown in Fig. 1b. This analysis suggested that target blood glucose levels of 110–144 and 144–180 mg/dL had a 48.4 and 43.6 % probability, respectively, of being the best treatment in view of reducing hospital mortality. In contrast, <110 and >180 mg/dL had a 7.0 and 68.4 % probability, respectively, of being the worst treatment in view of reducing hospital mortality. The Bayesian P value between 144–180 and >180 mg/dL was 0.034 (Supplementary appendix 5).

Fig. 1
figure 1

a Forest plot for hospital mortality. b Rank probabilities among all blood glucose targets for hospital mortality based on the network meta-analysis. a The upper row indicates the pairwise meta-analysis result (blue) and the lower row indicates the network meta-analysis result (black). OR odds ratio, 95 % CrI 95 % credible intervals, no direct study there were no studies directly comparing these blood glucose targets, number of patients the number of patients per target category

Full size image
Table 2 Results of subgroup analysis
Full size table

The forest plot for the risk of sepsis or bloodstream infection is shown in Fig. 2a. There were no significant differences among all comparisons of the network meta-analysis (Fig. 2a; Table 2). The rank probabilities indicated that target blood glucose levels of 110–144 and 144–180 mg/dL had a 61.6 and 31.7 % probability, respectively, of being the best treatment in view of lowering the risk of sepsis or bloodstream infection. In contrast, a target blood glucose concentration of >180 mg/dL had an 83.3 % probability of being the worst treatment in view of reducing the risk of infection (Fig. 2b).

Fig. 2
figure 2

a Forest plot for sepsis or bloodstream infection. b Rank probabilities among all blood glucose targets for sepsis or bloodstream infection based on the network meta-analysis. a The upper row indicates the pairwise meta-analysis result (blue) and the lower row indicates the network meta-analysis result (black). OR odds ratio, 95 % CrI 95 % credible intervals, no direct study there were no studies directly comparing these blood glucose targets, number of patients the number of patients per target category

Full size image

The forest plot for the risk of hypoglycemia is shown in Fig. 3a. Target blood glucose levels of <110 and 110–144 mg/dL were associated with a significantly higher risk of hypoglycemia compared with target blood glucose levels of 144–180 and >180 mg/dL. There were no significant differences in the risk of hypoglycemia between the 144–180 and >180 mg/dL target blood glucose levels (Fig. 3a; Table 2). The rank probability indicated that target blood glucose levels of >180 mg/dL had a 91.9 % probability of being the best treatment in view of lowering the risk of hypoglycemia. Furthermore, target blood glucose levels of <110 and 110–144 mg/dL had a 36.9 and 63.1 % probability, respectively, of being the worst treatment in view of reducing the risk of hypoglycemia (Fig. 3b).

Fig. 3
figure 3

a Forest plot for hypoglycemia. b Rank probabilities among all blood glucose targets for the risk of hypoglycemia based on the network meta-analysis. a The upper row indicates the pairwise meta-analysis result (blue) and the lower row indicates the network meta-analysis result (black). OR odds ratio, 95 % CrI 95 % credible intervals, no direct study there were no studies directly comparing these blood glucose targets, number of patients the number of patients per target category

Full size image

Subgroup analyses

ICU setting

There were no significant differences in the risk of hospital mortality and infection for all comparisons in the network meta-analysis performed for studies in both surgical and medical ICUs (Table 2). In terms of the risk of hypoglycemia, target blood glucose levels of <110 mg/dL were associated with a significantly higher risk compared with target blood glucose levels of 144–180 and >180 mg/dL, irrespective of the ICU setting. In addition, a target level of 110–144 mg/dL was associated with a significantly higher risk of hypoglycemia compared with target levels of 144–180 and >180 mg/dL in medical ICUs.

Proportion of diabetes patients

We could not perform a subgroup network meta-analysis for the >50 % diabetic patients subgroup, as there were only six such studies. In the <50 % diabetic patients subgroup, there were no significant differences in the risk of hospital mortality and infection in all comparisons (Table 2). Target blood glucose levels of <110 and 110–144 mg/dL were associated with significantly higher risks of hypoglycemia compared with 144–180 and >180 mg/dL.

Observed mortality >20 %

We performed a subgroup network meta-analysis for studies wherein the observed mortality was >20 %. There were no significant differences in the risk of hospital mortality and infection in all comparisons (Table 2). Target blood glucose levels of <110 and 110–144 mg/dL were associated with significantly higher risks of hypoglycemia compared with 144–180 and >180 mg/dL.

Actually achieved mean glycemic levels

We performed an additional network meta-analysis based on the actual mean glycemic levels. Although the actually achieved glucose concentration of <110 mg/dL was associated with a significantly higher hospital mortality compared with the concentration of 110–144 mg/dL in a direct comparison (OR 4.17; 95 % CrI 1.04–17.86), there were no significant differences among all comparisons of the network meta-analysis (Table 2; Supplementary appendix 6). Actually achieved blood glucose levels of <110 and 110–144 mg/dL were associated with significantly higher risks of hypoglycemia compared with 144–180 mg/dL (OR 7.4; 95 % CrI 2.3–25, OR 6.3; 95 % CrI 3.9–10, respectively).

Summary of findings regarding comparison between target blood glucose levels of 110–144 and 144–180 mg/dL

Although several guidelines do not recommend target blood glucose levels of both <110 and >180 mg/dL, it remains ambiguous whether target blood glucose levels of 110–144 or 144–180 mg/dL are superior because there were few studies which compared these two ranges directly [11, 55]. Therefore, we summarized the findings from the network meta-analysis for each outcome and subgroup for target glucose concentrations of 110–144 and 144–180 mg/dL in Supplementary appendix 7. There was no significant difference in the risk of death and infection between the two target concentrations. However, the risk of hypoglycemia was significantly higher in the 110–144 mg/dL group compared with the 144–180 mg/dL group, except for the surgical ICU subgroup.

Discussion

We conducted the current network meta-analysis to compare four different targets of acute glycemic levels in critically ill patients. In critically ill patients, we found that there were no significant differences in the risks of hospital mortality and infection for all comparisons of the four target glucose levels. However, target blood glucose levels of <110 and 110–144 mg/dL were associated with a four to ninefold increase in the risk of hypoglycemia compared with the targets of 144–180 and >180 mg/dL. There was no significant difference in the risk of hypoglycemia between the 144–180 and >180 mg/dL target glucose levels. Although this is a hypothesis-generating study, the current study is the first network meta-analysis to compare different target glucose levels for acute glycemic control. Therefore, our findings may be novel and, thus, require a detailed discussion.

Comparisons with prior studies

The Surviving Sepsis Campaign guidelines recommend that target blood glucose concentrations should be <180 mg/dL, rather than <110 mg/dL, on the basis of the results of the NICE-SUGAR study [10]. Although the NICE-SUGAR study reported that IIT increased mortality and the risk of hypoglycemia compared with targeting blood glucose levels between 144 and 180 mg/dL, there were few direct comparisons of 144–180 with 110–144 or >180 mg/dL as the target glycemic range. Therefore, our network meta-analysis provided the first evidence for comparisons between these targets.

Interpretation

Lowering blood glucose levels to <110 mg/dL was reported to reduce the mortality in a single-center study [2]; however, this finding was not reproduced in multicenter studies [9]. Our network meta-analysis also revealed that targeting blood glucose levels to <110 mg/dL was associated with no reduction in the risk of mortality and infection and, therefore, it may not be considered the best treatment for reducing the risk of mortality and infection. Furthermore, targeting blood glucose levels to <110 mg/dL was associated with a four to eightfold higher risk of hypoglycemia. This finding is consistent with a previous meta-analysis [5], and with the recent recommendation reported in international guidelines [10, 11].

Our analysis included 23 RCTs with >180 mg/dL as the upper limit of the blood glucose target. Although all direct and indirect comparisons did not show statistically significant differences in terms of hospital mortality and the incidence of infection, our rank probability suggested that a target glucose level of >180 mg/dL was the worst treatment in view of these two outcomes. This finding is consistent with the recent recommendation reported in international guidelines [10, 11]. However, we should note that only two RCTs achieved actual glycemic levels of >180 mg/dL. Therefore, our findings with regard to the >180 mg/dL target might not be truly reflective of the impact of this glycemic target. Furthermore, we should note that, as we categorized all studies in which the upper limit of their target was >180 as “>180 mg/dL”, the risk or benefit of a more liberal target (e.g., <200 mg/dL) could not be assessed in our analysis and, thus, requires further investigation.

Although the current guidelines do not recommend target blood glucose levels of both <110 and >180 mg/dL, it remains ambiguous whether target blood glucose levels of 110–144 or 144–180 mg/dL are superior [11, 55]. In our network meta-analysis, there were no significant differences in the reduction of hospital mortality and infection between the target blood glucose levels of 110–144 and 144–180 mg/dL, which is summarized in Supplementary appendix 7. Furthermore, the probability of being the best treatment for 110–144 and 144–180 mg/dL was similar, especially in view of reducing hospital mortality. Therefore, we were unable to deduce whether target blood glucose levels of 110–144 or 144–180 mg/dL were superior, as the order of 110–144 and 144–180 mg/dL in the rank probability was highly imprecise in terms of reducing hospital mortality and the incidence of infection. However, our results also showed that target blood glucose levels of 110–144 mg/dL were associated with a greater risk of hypoglycemia compared with targets of 144–180 mg/dL in critically ill patients, medical ICUs, the non-diabetic cohort, and overall mortality >20 % cohort.

Limitations

Although this is the first network meta-analysis on this issue and to study heterogeneous populations of critically ill patients, there are several limitations of this study. First, the majority of the included studies were retrieved from three recent systematic reviews [5, 6, 12]. However, we did also search the PubMed and Cochrane databases for additional studies that had been published after the meta-analysis by Wiener et al. (which used the same search terms) [5].

Second, we limited the included studies to full-text English publications, as we assumed that conference abstracts and non-English reports would not provide sufficient methodological or outcome data. Nevertheless, we carefully searched for full-text publications that were based on conference abstracts, and considered these studies.

Third, we performed a network meta-analysis in several subgroups to identify the impact of heterogeneity among the included RCTs. Notably, the proportion of surgical patients influenced the result in terms of the risk of hypoglycemia. Therefore, it should be noted that, as well as the different glycemic targets, there were other differences among the studies, including the glycemic control methodology, the glucose monitoring tools used, the feeding regimen, and the duration of intervention. This heterogeneity could also have influenced our results. Therefore, future studies are required to refute or confirm our results.

Fourth, our study was conducted using predefined four categories of blood glucose targets. Thus, we could not assess the impact of other blood glucose thresholds, such as 130, 160, 200, and 220 mg/dL. Accordingly, future studies are necessary to assess different blood glucose bands.

Fifth, the results of our network meta-analysis had some heterogeneity and inconsistency: in particular, the comparison between 144–180 and >180 mg/dL showed inconsistency in terms of the hospital mortality (Bayesian P value = 0.034). In addition, the comparison between <110 and 144–180 mg/dL in terms of the risk of hypoglycemia analysis showed heterogeneity (I 2 = 85.95). Therefore, it is necessary to consider this point in the interpretation of our results.

Finally, we used target blood glucose levels according to the methodology of previous systematic reviews [5, 6, 12]. However, this methodology did not consider the quality of blood glucose control during the study period. Therefore, we performed an additional network meta-analysis based on the actual mean glycemic levels, as we hypothesized that our results might be influenced by how accurately blood glucose control was achieved. Although these results were the same as those based on the target blood glucose level; the 95 % CrI in the comparison of >180 mg/dL was large because there were only two studies in which the actual blood glucose level of >180 mg/dL was reached. In this regards, future studies should consider the influence of the method of blood glucose level control when designing their own insulin protocol.

Conclusions

The current network meta-analysis revealed that there was no significant difference in the risk of mortality and infection among the four blood glucose ranges in critically ill patients. Our results also indicated that targeting blood glucose levels to <110 and 110–144 mg/dL was associated with a higher risk of hypoglycemia compared with the blood glucose targets of 144–180 and >180 mg/dL. These findings may shed light on the optimal blood glucose target in critically ill patients. To refute or confirm our findings, further studies should be conducted.