Protocol and guidance
The study protocol was prepared following PRISMA-P guidelines [18] and was registered at PROSPERO (CRD42017074039). The methods of the systematic review and meta-analysis followed PRISMA guidelines [19]. Reporting of statistical data in the study followed SAMPL guidelines [20].
Study selection
Inclusion criteria
We included RCTs that met each of the following criteria: the setting was a PICU, and the patient was child (age < 16 years); the intervention group received TGC (glucose goal < 140 mg/dl obtained using insulin treatment during part or all of the PICU stay); the comparison group received usual care (method of insulin administration and glucose goal could vary between trials); and the primary or secondary outcomes included hospital mortality, hypoglycemia (any, severe), new need for dialysis, sepsis, or seizures.
Exclusion criteria
Trials were excluded if the intervention was conducted primarily during the intraoperative period rather than during the PICU stay, or if we were unable to obtain adequate details of the study methodology or results from the article or study investigators.
Missing data
We contacted the investigators of all unpublished RCTs as well as any published RCTs in which data were missing to confirm eligibility and obtain additional study details.
Duplicate publications
If separate articles from the same RCT were published, the article with the most updated data was selected. In the case of duplicate publications, only one publication was included.
Information sources and search strategy
Medline, Embase, and the Cochrane Library at the CENTRAL Register of Controlled Trials were systematically searched. Gray literature was searched through appropriate databases (British Library Thesis Service, Database of Abstracts of Reviews of Effects, OpenGrey). We also consulted databases of clinical trial registries (ClinicalTrials.gov, World Health Organization International Clinical Trials Registry Platform, European Union Clinical Trials Register, ISRCTN Registry). The last electronic search was on May 1, 2017. We also hand searched the references to the retrieved articles and meta-analyses.
For the search strategy, we used a combination of keywords and MeSH terms for “child” AND “insulin”, using the sensitive search filters for therapeutic interventions (Additional file 1: Supplemental Digital Content).
Study selection
Two reviewers (YZ and LC) independently screened the titles and abstracts of retrieved reports for potential eligibility. They then screened the full text of potentially relevant trials. Disagreements were resolved by discussion and consensus or by consulting a third reviewer (TL).
Data collection process
Following removal of duplicate articles, two reviewers (YZ and LC) independently extracted data from the included RCTs using a standardized electronic form. Disagreements between the two reviewers were resolved by discussion and consensus or by consulting a third reviewer (TL). Another reviewer (FF) double-checked the extracted data.
Outcomes and prioritization
The primary outcome was hospital mortality because we considered a reduction in hospital mortality to be the most important potential benefit of TGC. Hospital mortality was defined as death occurring during the hospital stay or within 30 days following admission. In cases in which both in-hospital and 30-day outcomes were reported, the former was used for analysis.
The secondary outcomes were hypoglycemia (any, severe), sepsis, new need for dialysis, and seizures. We defined severe hypoglycemia as a blood glucose level below 40 mg/dl and any hypoglycemia as a blood glucose level below 60 mg/dl. We defined sepsis to encompass the terms septicemia, bacteremia, or a description of positive blood cultures; a general description of infection did not qualify.
Subgroup and sensitivity analyses
We performed subgroup analyses based on three variables prespecified clinically relevant to analysis outcomes: glucose goal in the tight control group, cardiac surgery, and continuous glucose monitoring. Subgroup analyses were performed only if there were at least two RCTs in each subgroup or a trial’s report permitted a comparison within the trial.
Differing opinions exist on the optimal level of TGC. The 2018 recommendations from the American Diabetes Association recommend targeting blood glucose levels of 140–180 mg/dl in critically ill patients [21,22,23]. We stratified studies by glucose goal in the TGC group into two categories: very tight control (upper limit of glucose goal < 110 mg/dl); and moderately tight control (upper limit of glucose goal 110–140 mg/dl).
Because of the concern that the pathophysiological effect of hyperglycemia may differ between patients with and without cardiac surgery, we stratified trials by PICU setting into two categories: cardiac surgery and not cardiac surgery. For trials involving mixed populations but not presenting separate data for patients with cardiac surgery, we included the pooled results in the cardiac surgery subgroup only if ≥ 50% of patients underwent cardiac surgery.
Continuous glucose monitoring has been shown to be safe and effective in children and adults, and may assist in the safer provision of tight glycemic control, with less hypoglycemia [24]. Thus, we stratified trials by whether they used continuous glucose monitoring to control blood glucose.
We conducted sensitivity analyses to examine the impact of using alternative effect measures (odds ratio vs relative risk), pooling methods (Peto vs Mantel–Haenszel (M–H) or inverse variance), statistical models (fixed vs random effects), and removing one study at a time.
Risk of bias and quality of evidence
Two reviewers (YZ and LC) independently assessed risk of bias (low risk of bias, high risk of bias, or unclear risk of bias) using the Cochrane risk of bias instrument, which deals with random sequence generation and allocation concealment (selection bias), blinding of study participants and personnel (performance bias), blinding of outcome assessment (detection bias), incomplete outcome data (attrition bias), selective reporting (reporting bias), and other bias. They resolved any disagreements by discussion and consensus or by consulting a third reviewer (TL). We judged trials with more than two high-risk components as having a moderate risk of bias, and trials with more than four high-risk components as having a high risk of bias. We used the GRADE approach to rate the quality of evidence and generate absolute estimates of effect for the outcomes [25].
Data synthesis
Computations were performed with RevMan 5.3.3 software (freeware available from The Cochrane Collaboration). We used the M–H method as the primary analysis to estimate the odds ratio (OR) and 95% confidence intervals (CIs). Two-tailed P < 0.05 was considered a criterion for statistical significance. We report the results of the random-effects model for all outcomes. We assessed heterogeneity with the Cochran Q test and the I2 test, with I2 values exceeding 25%, 50%, and 75% representing low, moderate, and high heterogeneity, respectively [26]. If an analysis included 10 or more RCTs, we planned to use a funnel plot to explore the possibility of published bias.