Statistical analyses were conducted according to a pre-specified, published statistical analysis plan written
prior to the interim analysis.43The final analyses were conducted using Stata/SE version 13.0.
Baseline characteristics
Baseline demographic and clinical data were summarised by treatment group. Statistical tests for differences between the groups were not reported, as these may be misleading. Discrete variables were summarised as numbers and percentages, which were calculated according to the number of patients for
whom data were available; when values were missing, the denominator was reported. Continuous variables were summarised by standard measures of central tendency and dispersion, either mean and standard deviation (SD) and/or median and interquartile range (IQR) as specified below:
l age, mean (SD) and median (IQR)
l sex, n (%)
l severe comorbidities (as defined by APACHE II34), n (%):
¢ severe liver condition
¢ severe renal condition
¢ severe respiratory condition
¢ severe cardiovascular condition
¢ immunocompromised
l acute severity of illness:
¢ SOFA score,33mean (SD)
¢ APACHE II Score,34mean (SD)
¢ APACHE II Acute Physiology score,34mean (SD)
¢ APACHE II predicted risk of death34(2013 UK recalibration), median (IQR)
¢ ICNARC Physiology Score,35mean (SD)
¢ ICNARC model predicted risk of death35(2013 recalibration), median (IQR)
l surgical status– surgery within 24 hours prior to critical care unit admission, n (%)
l ventilation status– mechanical ventilation at admission to the critical care unit, n (%)
l malnourished– yes/no (based on clinical judgement), n (%)
l actual/estimated BMI, mean (SD) and median (IQR)
l ulna length (cm), mean (SD) and median (IQR)
l mid-upper arm circumference (cm), mean (SD) and median (IQR)
l degree of malnutrition (high, BMI of< 18.5 kg/m2or weight loss of> 10%; moderate, BMI of < 20 kg/m2
or weight loss of> 5%; no malnutrition) (based on NICE definitions1), n (%). Adherence
Non-adherence with the allocated treatment was reported as the number and percentage of patients who:
l did not receive any nutritional support
l received first nutritional support via the opposite route to assigned
l received initiation of nutritional support more than 36 hours after admission to critical care
l received early nutritional support via assigned route and subsequently changed to opposite route
during first 120 hours, or
l received no nutritional support for at least a full 1-day period during the first 120 hours.
Delivery of care
Delivery of care was summarised by treatment group but not subjected to statistical testing. As with baseline characteristics, discrete variables were summarised as numbers and percentages. Percentages were calculated according to the number of patients for whom data were available; where values were missing, the denominator was reported. Continuous variables were summarised by mean (SD) and/or median (IQR), as specified below:
l time from critical care unit admission to commencement of nutritional support (hours), median (IQR)
l total calories and average calories per 24 hours received during intervention period (total calories and a
l total protein and average protein per 24 hours received during intervention period (total protein and a breakdown of the total protein received via the enteral and the parenteral route), mean (SD)
l total gastric residual volume aspirated (ml) and average per 24 hours during intervention period, if fed
via the enteral route, mean (SD)
l total gastric residual volume replaced and average per 24 hours during intervention period, if fed via
the enteral route, mean (SD)
l patients receiving additives during intervention period (glutamine, selenium and fish oils), if fed via the
parenteral route, n (%)
l patients receiving prokinetics during intervention period, if fed via the enteral route, n (%)
l patients receiving insulin, n (%), and total insulin received (IU), mean (SD), during intervention period
l patients receiving vasoactive agents during intervention period, n (%)
l incidence of diarrhoea and constipation, n (%)
l time from randomisation to commencement of exclusive oral feeding (days), median (IQR)
l daily SOFA score during the intervention period, median (IQR).
Primary outcome: clinical effectiveness
The number and percentage of deaths at 30 days following randomisation, due to any cause, were reported for each treatment group. The primary effect estimate was the relative risk of all-cause mortality at 30 days, reported with a 95% CI. The absolute risk reduction and 95% CI were also reported. Deaths at
30 days after randomisation were compared between the treatment groups, unadjusted, using Fisher’s
exact test. A secondary analysis of the primary outcome, adjusted for baseline variables, was conducted using multilevel logistic regression. Baseline variables adjusted for in the multilevel logistic regression model were age, ICNARC Physiology Score, surgical status, degree of malnutrition and a site-level random effect. Baseline variables were selected for inclusion in the adjusted analysis a priori according to anticipated relationship with outcome. The results of the multilevel logistic regression model were reported as an adjusted odds ratio with 95% CI. The unadjusted odds ratio was presented for comparison.
Secondary outcomes: clinical effectiveness
The mean (SD) of the number of days alive and free from advanced respiratory support, advanced
cardiovascular support, renal support, neurological support and gastrointestinal support, as defined by the
CCMDS36(see Appendix 4), up to 30 days, within each treatment group, were reported. Patients who died
within the first 30 days were assigned zero days alive and free of each organ support. Days of organ support were recorded only while the patient was in a critical care unit; any days outside of a critical care unit were assumed to be free of organ support. Differences between the treatment groups were tested using the t-test, using the non-parametric bootstrap to account for anticipated non-normality in the
distributions.44A total of 1000 bootstrap replications were taken, stratified by treatment group, with
bias-corrected and accelerated CIs reported.
The mean (SD) number of treated infectious complications per patient and the number and percentage of patients with each infectious complication (chest infection, central venous catheter infection, other vascular catheter-related infection, bloodstream infection, infective colitis, urinary tract infection, surgical site infection, other infectious complication) and each non-infectious complication (episodes of hypoglycaemia, elevated levels of liver enzymes, nausea requiring treatment, abdominal distension, vomiting, new
or substantially worsened pressure ulcers) within each treatment group were reported. Infectious complications and pressure ulcers were assessed while the patient was in the critical care unit; all other non-infectious complications were collected through adverse event reporting up to 30 days following randomisation. Differences between the treatment groups were tested using the t-test for means and Fisher’s exact test for percentages.
The median (IQR) of the length of stay in critical care and in acute hospital were reported for each treatment group. Differences in length of stay between the treatment groups were tested using the Wilcoxon
Kaplan–Meier curves by treatment group were plotted up to 90 days and 1 year after randomisation and compared using the log-rank test. An adjusted comparison was performed using a Cox proportional hazards model, which was adjusted for the same baseline variables as the primary outcome, including shared frailty within sites (gamma-distributed latent random effects). The appropriateness of the
proportional hazards assumption was assessed graphically by plotting–log[−log(survival)] against log(time)
within treatment groups. The number and percentage of deaths at critical care and acute hospital discharge, and by 90 days and 1 year post-randomisation, were reported for the treatment groups.
Differences in all-cause mortality at each time point were compared, unadjusted, using Fisher’s exact
test and, adjusted, using multilevel logistic regression, adjusted for the same baseline variables as the primary outcome.
Safety monitoring
The number and percentage of patients experiencing each serious adverse event (occurring between randomisation and 30 days) were reported for each treatment group. The total number of patients
experiencing one or more serious adverse events was compared between treatment groups using Fisher’s
exact test and summarised as a relative risk with 95% CI.
Subgroup analyses of the primary outcome
Subgroup analyses were conducted to test for a difference in treatment effect according to pre-specified subgroups. Differences in the primary outcome (30-day mortality) were analysed by age (in quartiles),
degree of malnutrition (high/moderate or none), acute severity of illness (APACHE II34and ICNARC model35
predicted risk of mortality– in quartiles), mechanical ventilation at admission to the critical care unit, presence of cancer and time from critical care unit admission to commencement of nutritional support (< 24 hours vs. ≥ 24 hours).
These analyses tested for an interaction between the subgroup categories and the treatment group in multilevel logistic regression models, adjusted for the same baseline variables as used in the primary analysis.
Secondary analyses of the primary outcome
Sensitivity analyses for missing data in the primary outcome
As the number of missing data was anticipated to be minimal, a sensitivity approach was taken when the primary outcome variable was missing. The primary analysis was repeated once, assuming that all patients in the enteral route group with missing outcomes survived and all patients in the parenteral route
group with missing outcomes did not survive. The analysis was then repeated again with the opposite assumptions. This approach gives the absolute range of how much the results could change if all of the data were complete.
Adherence-adjusted analysis
Although the intention-to-treat analysis provides the best estimate of the clinical effectiveness of early nutritional support via the parenteral route compared with the enteral route, it was also of interest to estimate what the efficacy of early nutritional support delivered via the parenteral route would be compared with the enteral route, if delivered as intended. In a randomised controlled trial, the allocated
treatment can be used as an‘instrumental variable’, that is, a variable associated with receipt of the
intervention and associated with the outcome only through its association with the intervention.45This
relationship enables us to estimate what the treatment effect would be for patients who are adherent to the protocol. The primary analysis was repeated adjusting for adherence using a structural mean model with an instrumental variable of allocated treatment, assuming a linear relationship between the degree of