Types of studies

We considered published and unpublished randomised and quasi‐randomised controlled trials for this review.

Types of participants

Preterm infants (less than 37 weeks' gestation) receiving enteral feeding of human milk, within a hospital setting.

Types of interventions

Human milk, with or without additional protein supplementation. Micronutrient supplements were allowed in both groups.

Types of outcome measures

The primary and secondary outcomes for this review were aligned with the outcomes of the Cochrane Review, Multi‐nutrient fortification of human milk for preterm infants (Brown 2016).

Electronic searches

We conducted a comprehensive search including: Cochrane Central Register of Controlled Trials (CENTRAL 2019, Issue 8) in the Cochrane Library and MEDLINE via PubMed (2018 to 23 August 2019). We have included the search strategies for each database in Appendix 1. We did not apply language restrictions.

We searched clinical trial registries for ongoing or recently completed trials (ISRCTN Registry). The World Health Organization’s International Clinical Trials Registry Platform (ICTRP) (www.who.int/ictrp/search/en/) and the U.S. National Library of Medicine’s ClinicalTrials.gov (clinicaltrials.gov) were searched via Cochrane CENTRAL.

This search updates the searches conducted for previous versions of the review (Amissah 2018, Kuschel 2000c).

Searching other resources

We also searched the reference lists of any articles selected for inclusion in this review, in order to identify additional relevant articles. We did not search any additional conference proceedings.

Selection of studies

For the 2018 update, two review authors (EA and JB) independently screened the titles and abstracts of the records identified by the searches. We resolved conflicts by discussion, or by consulting with a third author (JH). We retrieved the full text of all potentially relevant articles, and linked reports of the same study. Two review authors (EA and JB) independently assessed the full‐text articles for inclusion, using the eligibility criteria. We resolved conflicts by discussion, or by consulting with a third author (JH). We had planned to correspond with investigators to clarify study eligibility and obtain missing results if needed. We used Covidence for the study selection and data collection processes (Covidence).

For the 2020 update, Cochrane Neonatal screened the titles and abstracts identified by the search, as well as potentially relevant full‐text articles, independently and in duplicate in consultation with a review author (JH).

Data extraction and management

We developed a data extraction form prior to data gathering, to enable two review authors to independently extract information from the studies. We extracted data such as source details, study eligibility, study design, participant characteristics, intervention and control details, and outcomes. We resolved conflicts in the data extraction and management process by discussion, or by consulting with a third review author. We then exported the data into Cochrane's review software, Review Manager 5 (Review Manager 2014).

Assessment of risk of bias in included studies

Two review authors (EA and JB) independently assessed the risk of bias (low, high, or unclear) of all included trials, using the Cochrane ‘Risk of bias’ tool, for the following domains (Higgins 2017):

• Sequence generation (selection bias)

• Allocation concealment (selection bias)

• Blinding of participants and personnel (performance bias)

• Blinding of outcome assessment (detection bias)

• Incomplete outcome data (attrition bias)

• Selective reporting (reporting bias)

• Any other bias

We resolved conflicts by discussion, or by consulting with a third review author. See Appendix 2 for a more detailed description of risk of bias for each item.

Measures of treatment effect

For dichotomous data, we used the number of events in the control and intervention groups of each study to calculate risk ratios (RRs) with 95% confidence intervals (CIs). For continuous data, we calculated mean differences (MDs) between treatment groups with 95% CIs, where outcomes were measured in the same way. We did not need to use standardised mean differences (SMD) in this update, but they will be used in future updates where outcomes from studies are the same, but different methods have been used to collect the data. We did not calculate numbers needed to treat for an additional beneficial outcome (NNTB) or the numbers needed for an additional harmful outcome (NNTH), due to insufficient data.

Unit of analysis issues

We did not identify any unit of analysis issues. In future updates, if we identify cluster‐randomised trials, we will undertake analysis at the individual level, taking clustering into account, as recommended in the Cochrane Handbook for Systematic Reviews of Interventions (Higgins 2017).

Dealing with missing data

We noted levels of attrition. We carried out analyses using an intention‐to‐treat basis, where possible, for all of the outcomes. Where possible, we analysed all participants in the treatment group to which they were randomised, regardless of the actual treatment received. We did not contact any of the trial authors. In future updates, if data are missing, we will make an attempt to contact the trial authors. We were unable to conduct sensitivity analyses, and were unable to address the potential impact of missing data on the findings of the review, due to insufficient data.

Assessment of heterogeneity

We considered whether the clinical and methodological characteristics of the included studies were sufficiently similar for meta‐analysis to provide a clinically meaningful summary. This was done by assessing statistical heterogeneity using the Chi² test and the I² statistic. We took an I² measurement greater than 50%, and P < 0.10 in the Chi² test for heterogeneity to indicate moderate‐to‐high heterogeneity. Where we detected moderate‐to‐high heterogeneity, we had planned to explore possible explanations for clinical heterogeneity via subgroup analyses or methodological heterogeneity via sensitivity analyses, or both. We had planned to take clinical and statistical heterogeneity into account when interpreting the results, especially if there was any variation in the direction of effect.

Assessment of reporting biases

Reporting biases arise when the dissemination of research findings is influenced by the nature and direction of results. Some types of reporting bias (e.g. publication bias, multiple publication bias, language bias) reduce the likelihood that all studies eligible for a review will be retrieved. If all eligible studies are not retrieved, the review may be biased. We aimed to conduct a comprehensive search for eligible studies, and were alert for duplication of data. We were unable to formally assess publication bias, as there were insufficient studies for any of the outcomes (10 or more studies required). In future updates, if we find 10 or more studies reporting an outcome, we will assess publication bias by visual inspection of a funnel plot.

Data synthesis

We performed meta‐analyses using Review Manager (Review Manager 2014). We used risk ratio (RR) for dichotomous data, and mean difference (MD) for continuous data, with their respective 95% confidence intervals (CIs). We used a fixed‐effect model to combine data where similar interventions, populations and methods were employed by the trials. We planned to explore potential causes of heterogeneity via sub‐group and sensitivity analyses and assessed the quality of evidence at the outcome level, using GRADE methodology.

Quality of evidence

We used the GRADE approach, as outlined in the GRADE Handbook, to assess the quality of evidence for the following (clinically relevant) outcomes: growth, neurodevelopment, duration of hospital admission, feeding intolerance that resulted in cessation or reduction in enteral feeding, and necrotizing enterocolitis (Schünemann 2013).

Two authors (EA and JB) independently assessed the quality of the evidence for each of the outcomes above. We considered evidence from randomised controlled trials as high quality, but downgraded the evidence one level for serious (or two levels for very serious) limitations, based upon the following: design (risk of bias), consistency across studies, directness of the evidence, precision of estimates, and presence of publication bias. We used GRADEpro GDT software to create a ‘Summary of findings’ table, to summarise results and report the quality of the evidence (GRADEpro GDT).

The GRADE approach leads to an assessment of the quality of a body of evidence at one of four levels:

• High: we are very confident that the true effect lies close to that of the estimate of the effect.

• Moderate: we are moderately confident in the effect estimate; the true effect is likely to be close to the estimate of the effect, but there is a possibility that it is substantially different.

• Low: our confidence in the effect estimate is limited; the true effect may be substantially different from the estimate of the effect.

• Very low: we have very little confidence in the effect estimate; the true effect is likely to be substantially different from the estimate of effect.

Subgroup analysis and investigation of heterogeneity

We considered whether an overall summary was meaningful by assessing clinical and methodological heterogeneity among trials (see section above on 'Assessment of Heterogeneity'). If we had found moderate‐to‐high heterogeneity, we had planned to perform subgroup and sensitivity analyses.We had planned to carry out the following subgroup analyses to evaluate differences in outcome between: gestational age subgroups (less than 30, versus 30 up to 34, versus 34 up to 37 completed weeks), birth weight subgroups (less than 1 kg versus 1 kg or above), male versus female, and types of protein supplements (bovine versus human, and with versus without micronutrients or minerals). However, there were insufficient data to allow us to conduct any subgroup analyses.

Sensitivity analysis

We had planned to conduct sensitivity analyses by examining only those trials considered to have a low risk of bias for allocation concealment and randomisation. We were unable to do this as all the included studies were judged to be of unclear risk of bias for both allocation concealment and randomisation.