Sedation of mechanically ventilated adults in intensive care unit: a network meta-analysis

Sedatives are commonly used for mechanically ventilated patients in intensive care units (ICU). However, a variety of sedatives are available and their efficacy and safety have been compared in numerous trials with inconsistent results. To resolve uncertainties regarding usefulness of these sedatives, we performed a systematic review and network meta-analysis. Randomized controlled trials comparing sedatives in mechanically ventilated ICU patients were included. Graph-theoretical methods were employed for network meta-analysis. A total of 51 citations comprising 52 RCTs were included in our analysis. Dexmedetomidine showed shorter MV duration than lorazepam (mean difference (MD): 68.7; 95% CI: 18.2–119.3 hours), midazolam (MD: 10.2; 95% CI: 7.7–12.7 hours) and propofol (MD: 3.4; 95% CI: 0.9–5.9 hours). Compared with dexmedetomidine, midazolam was associated with significantly increased risk of delirium (OR: 2.47; 95% CI: 1.17–5.19). Our study shows that dexmedetomidine has potential benefits in reducing duration of MV and lowering the risk of delirium.

Quality assessment of component trials. Only randomized controlled trials were included in our study.
Therefore, the qualities of trials were assessed in six aspects: (1) random sequence generation, (2) allocation concealment, (3) blinding of participants and personnel, (4) blinding of outcome assessor (5) incomplete data outcome, (6) selective reporting, and (7) other bias. These items were adapted from the Cochrane Collaboration's tool for assessing risk of bias 10 . Statistical analysis. There were two types of outcome data: continuous and binary outcomes. The former included ICU and hospital LOS, and duration of MV. RASS and RSS were ordinal variables. Binary outcomes included mortality, atrial fibrillation and delirium. Mean difference (MD) was reported for the comparison of continuous outcomes between interventions, and odds ratio (OR) was reported for binary outcomes.
Graph-theoretical methods, which have been routinely applied to electrical networks, were employed for network meta-analysis 11 . Direct comparisons between sedatives were derived from each of the two-arm trials, and were represented by edges in a network plot. Effect sizes from component trials were weighted by the inverse of the observed variance of the treatment effect. For a meta-analytic network, the node corresponds to a treatment strategy, and the edge represents the existing comparisons between treatments. Each two-arm study (e.g. randomized controlled trial) contributes to one comparison. The thickness of the edge is proportional to inverse standard error of random effects model comparing two treatments. Additionally, we employed net heat plot to highlight hot spots of inconsistency between specific direct and indirect evidence in the whole network 12 . The area of a gray square displays the contribution of the direct estimate of one design in the column to a network estimate in a row. The colors are associated with the change in inconsistency between direct and indirect evidence in row design after detaching the effect of column design. Forest plots were employed to show the effect size of each drug, by setting dexmedetomidine as the reference. Effect sizes and corresponding 95% confidence intervals were reported in the forest plots. All statistical analyses were performed using R (version 3.2.3) 13 .

Results
Included studies and characteristics. The initial search identified 598 citations. Another 9 studies were added from the references of relevant articles (Fig. 1). After removing duplicates, a total of 203 citations remained for further screening. The titles and abstracts were screened by hand and 125 were excluded because 18 were related to anesthesia, one was animal study, 57 were irrelevant studies, 12 were observational studies, 11 involved pediatric patients, and 26 were reviews. The full-text articles of the remaining 78 citations were screened. Twentyseven articles were excluded because 20 investigated sedation protocol that the type of sedative drugs could not be identified, 4 studies were secondary analysis of previous reports, and 3 were study protocols. As a result, a total of 51 citations comprising 52 RCTs were included in our analysis  . The article by Jakob and colleagues comprised two RCTs 38 . Characteristics of component trials are shown in Table 1. These articles were published between the year 1989 and 2016. Study populations included patients underwent major operations requiring ICU admission, and those requiring long-term MV. The sample sizes in included trials ranged from 20 to 500. Most of the trials were two-arm trials, and there were five three-arm trials 20,35,42,63,64 . Study endpoints included duration of MV, Richmond Agitation Sedation Scale (RASS), Ramsay Sedation Scale (RSS), ICU and hospital length of stay (LOS). Adverse events included bradycardia, hypotension and death.

Risk of bias.
Random sequence generation was adequately described in approximately half of included trials ( Fig. 2). In the remaining trials, they did not specifically describe the method of sequence generation. Allocation concealment was properly done in about 21 of the 52 trials. Blinding was difficult to perform because propfol was distinctive in appearance. Attrition bias and reporting bias were generally well performed in included trials. Risk of bias assessment of each trial is present in Supplemental Fig. 1.
Network graph for the duration of MV is shown in Fig. 3. Collectively, there were 8 sedatives being compared. They were midazolam, dexmedetomidine, propofol, clonidine, morphine, haloperidol, clonidine and lorazepam. The placebo meant that no sedative was given in that group. Two studies employed propofol and midazolam in combination as the control arm, and we denoted it as the standard 20 to inverse standard error of random effects model comparing two treatments. For example, the dex-propofol comparison appears to be thick, indicating a small standard error for the effect size of the comparison (Fig. 3). Multi-arm studies were highlighted with blue color. Clinical outcomes. MV duration was reported in most studies. Random effects model was employed to combine the results. Dexmedetomidine showed shorter MV duration than lorazepam (MD: 68.74; 95% CI: 18.2-119.3 hours), midazolam (MD: 10.2; 95% CI: 7.7-12.7 hours) and propofol (MD: 3.4; 95% CI: 0.9-5.9 hours). However, MV duration in dexmedetomidine group was longer than clonidine (MD: − 9.4; 95% CI: − 16.1-− 2.7 hours) and placebo (MD: − 5.2 95% CI: − 11.4-0.99 hours). There were no significant differences for MV duration in dexmedetomidine group as compared to that in haloperidol, morphine, sevoflurane and standard groups (Fig. 4). There were large changes in inconsistency between direct and indirect evidence in design dex:pro (shown in the row) after detaching the effect of corresponding column design (Fig. 5). As compared to dexmedetomidine, midazolam was associated with significantly increased risk of delirium (OR: 2.47; 95% CI: 1. 17-5.19). Propofol was associated with increased risk of delirium with marginal statistical significance (OR: 2.14; 95% CI: 0.94-4.89). There was no difference in the risk of delirium in other comparisons (Fig. 6). There was no difference in the ICU LOS in all comparisons (Suppl. Fig. 2), except that haloperidol was associated with longer ICU stay (MD: 5; 95% CI: 1.8-8.2 days). Dexmedetomidine was associated with shorter LOS in hospital than propofol (MD: 4.6; 95% CI: 1.2-8.1 days, Suppl. Fig. 3). There is no difference in mortality between dexmedetomidine and other sedatives (Suppl. Fig. 4). There were no differences in RSS, RASS or atrial fibrillation between dexmedetomidine and other comparators (Suppl. Figs 5 to 7).

Discussion
The present study showed that dexmedetomidine was able to reduce MV duration in critically ill patients, as compared to conventional sedatives such as lorazepam, midazolam and propofol. In addition, dexmedetomidine was associated with lower risk of delirium than that of midazolam and propofol. Dexmedetomidine was also associated with shorter hospital LOS than propofol. Propofol showed a shorter MV duration when compared to midazolam, and it has similar risk of delirium to midazolam. There were no significant differences between sedatives in other important outcomes such as mortality, ICU and hospital LOS.
Several meta-analyses of sedatives in critically ill patients have been conducted before our study [65][66][67] . Fraser's study included only 6 trials comparing Benzodiazepine versus nonbenzodiazepine-based sedation for the mechanically ventilated patients. Numerous studies in this area have been published since that time, and the evidence needs to be updated. In our study, we included 51 citations that were far more than that included in    Fraser's study. Furthermore, previous studies did not perform meta-analysis in network framework. By using conventional pairwise meta-analysis, many types of sedatives had to be combined as the control group, ignoring the fact that these sedatives were different in their pharmacological properties. In clinical practice, clinicians usually face with the choice between multiple alternative sedatives. The choice may become difficult when investigators  have undertaken head-to-head comparisons of only some of available sedatives. Chen's study highlighted dexmedetomidine, and other sedatives were used as the controls. Cruickshank's study also had the same limitation.
We believe that all other sedatives are different in their efficacy and safety profiles. In this situation, network meta-analysis is more appropriate because it allows for simultaneous comparisons of multiple interventions against each other. With respect to searching database, SCOPUS was not searched in Cruickshank's study. In the present study, many publications in non-English language were retrieved from SCOPUS. Missing these citations may result in biased estimates of pooled results, also known as publication bias. Although there are significant differences in included trials, some aspects of our findings are consistent with previous meta-analyses. For example, Cruickshank's study found that dexmedetomidine was effective in reducing time to extubation in the ICU patients, and risk of bradycardia but not of overall mortality is higher among patients treated with dexmedetomidine 67 . However, Fraser's study found a similar prevalence of delirium between patients treated with benzodiazepine versus nonbenzodiazepine sedatives 65 . This discrepancy can be partly explained by the limited number of component trials and the combination of different types of sedatives as the control group. Several limitations in our study need to be acknowledged. First, the study population involved critically ill patients that were heterogeneous in nature. Patients after major operation and medical ICU patients requiring long-term MV were enrolled. However, the common feature was that they all needed MV, and sedatives were used for the same purpose for them. Second, there were significant risks of bias in most of enrolled trials. For example, blinding to participants and investigators was not performed due to the appearance of propofol. Also, specific methods to generate random sequence were not explicitly reported in nearly half of the trials. Third, more than two thirds of component trials had a sample size of less than 100, which were typically small studies. As a result, the systematic review may be subject to small study effect bias 68 .
In conclusion, our study showed that dexmedetomidine had potential benefits in reducing duration of MV and lowering the risk of delirium. Propofol is considered superior to midazolam in terms of MV duration. Adverse events of hypotension and bradycardia should be closely monitored when sedatives are used for MV patients.