Enteral versus parenteral nutrition in critically ill patients: an updated systematic review and meta-analysis of randomized controlled trials

A literature review was conducted to identify all relevant RCTs published between 1980 and January 2016 in MEDLINE, Embase, CINAHL, the Cochrane Central Register of Controlled Trials, and the Cochrane Database of Systematic Reviews. The following keywords or medical subject headings were used: “randomized”, “clinical trial”, “nutrition support”, “artificial feeding”, “enteral nutrition”, “parenteral nutrition”, “intensive care”, “critical illness”, and “critically ill”. The literature search was not confined to articles written only in English. The authors’ personal files and reference lists of relevant review articles were also reviewed. Neither ethics board approval nor patient consent was required due to the nature of a systematic review.

Trials were included only if they met the following characteristics:

The methodological quality of the included trials was assessed in duplicate by two independent reviewers using a data abstraction form with a scoring system from 0 to 14 according to the following criteria as previously described [8, 17]:

Disagreement concerning the individual score of each of the defined categories was resolved by consensus between the two reviewers. Moreover, attempts were made to contact the authors of the included trials in order to request further information not contained in the published article, if needed.

Statistical analyses were performed using RevMan 5.3 (Cochrane IMS, Oxford, UK) with a random effects model. All trial data were combined to estimate the pooled risk ratio (RR) with 95 % confidence intervals (CIs) for mortality and infections and overall weighted mean difference (WMD) with 95 % confidence intervals for LOS and mechanical ventilation data. We calculated pooled RRs using the Mantel-Haenszel estimator, and WMDs were estimated by the inverse variance approach. The random effects model of DerSimonian and Laird was applied to estimate variances for the Mantel-Haenszel and inverse variance estimators [18]. In case RRs were undefined they were excluded for studies with no event in either arm. Heterogeneity testing was performed using a weighted Mantel-Haenszel χ² test and quantified by the heterogeneity I² statistic as implemented in RevMan. Differences between subgroups were analyzed using the test of subgroup differences described by Deeks et al. [19], and the results expressed using the P values.

Generating funnel plots and testing asymmetry of outcomes using the method proposed by Rücker et al. [20] addressed possible publication bias. We considered P < 0.05 to be statistically significant.

A predefined subgroup analysis was performed to further explore whether the treatment effect of either route is associated with significant differences in the caloric intake across the study groups (EN compared to PN). A priori, we hypothesized that a possible negative treatment effect of PN on mortality, infectious complications and length of stay is related to a higher caloric intake. We used the reported significance level on caloric intake across groups from each study to determine the allocation to the subgroup. We also assessed the effect of trial publication date and trial quality on the outcome based on the hypothesis that older or methodologically weaker studies tended to yield a more negative treatment effect of PN. For this purpose, we designated trials with a publication date later than 1995 (median year 1994–1995) as newer and trials with a methodological score of more than 7 (median score) as methodologically stronger, respectively. In addition, we conducted a sensitivity analysis excluding two studies [21, 22] we were uncertain if the patients were truly critically ill [21] and about the trial results [22] and authors did not respond to attempts to obtain clarification.

The literature search identified 83 potentially eligible randomized controlled trials, of which 65 were excluded for the following reasons (see Table A1 in Additional file 1): (a) patients not considered to be adult critically ill patients (N = 42), (b) no relevant clinical outcomes meeting inclusion criteria reported (N = 4), (c) being duplicate studies, reviews of published trials or subgroups of included studies (N = 11), (d) non-randomized or pseudo-randomized study design (N = 7), and/or (e) control group received a non-standard enteral formula (N = 1).

Thus, 18 RCTs with a total number of 3347 critically ill adult patients were finally included in the meta-analysis, whereof 1681 patients were treated with EN and 1666 patients with PN.

The median methodological score of the included 18 RCTs was 7 (range, 2–12) of which 10 RCTs were rated with a score ≤7 and 8 RCTs with a score >7. The median year of publication was 1994–1995 with 10 RCTs published before or in 1995 and 8 RCTs after 1995. All results were based on the individual trial data shown in Tables 1 and 2.

Aggregating data from all studies reporting on mortality (N = 16) there was no difference in overall mortality between the groups receiving EN or PN (RR 1.04, 95 % CI 0.82, 1.33, P = 0.75, heterogeneity I² = 11 %) (Fig. 1). In the subgroup analysis where trials were aggregated according to the caloric intake across groups, no effect on mortality was seen in trials (N = 4) where the PN group received significantly more calories than the EN group (RR 1.58, 95 % CI 0.75, 3.35, P = 0.23, heterogeneity I² = 48 %) or in the nine trials where the caloric intake was reported to be non-significantly different across groups (RR 1.03, 95 % CI 0.93, 1.14, P = 0.55, heterogeneity I² = 0 %; test for subgroup differences: P = 0.55) (Fig. 1, Panels a and b). Three RCTs did not report the caloric intake across study groups (Fig. 1, Panel c). In the sensitivity analysis excluding two studies [21, 22], there was still no difference in mortality between groups (RR 1.08, 95 % CI 0.83, 1.39, P = 0.57, heterogeneity I² = 14 %).

Eleven trials were aggregated which reported on infectious complications. EN compared to PN was associated with a significant reduction in the incidence of infectious complications (RR 0.64, 95 % CI 0.48, 0.87, P = 0.004, heterogeneity I² = 47 %; Fig. 2). The significant difference was also seen in the subgroup analysis of five aggregated trials in which the PN group had a significantly higher caloric intake than the EN group (RR 0.55, 95 % CI 0.37, 0.82, P = 0.003, heterogeneity I² = 0 %) but not when the five trials were aggregated where caloric intake was similar between EN and PN groups (RR 0.94, 95 % CI 0.80, 1,10, P = 0.44, heterogeneity I² = 0 % [test for subgroup differences: P = 0.003]) (Fig. 2, Panels a and b). One trial with data on infectious complications did not report the caloric intake across the EN and PN group (Fig. 2, Panel c). EN compared to PN was still associated with a significant reduction in infectious complications (RR 0.58, 95 % CI 0.41, 0.8, P = 0.001, heterogeneity I² = 29 %) in the sensitivity analysis excluding the two studies with inconclusive status of critical illness [21, 22].

Only four studies reported on ICU LOS (in mean and standard deviation [SD]) and when the data were aggregated, the use of EN was associated with a significant reduction in ICU LOS (WMD −0.80, 95 % CI −1.23, −0.37, P = 0.0003, heterogeneity I² = 0 %, Fig. 3, Panel a). In the subgroup analysis aggregating trials according to the caloric intake across groups, the significant difference was not observed in the two trials where caloric intake was similar between EN and PN groups (RR −0.47, 95 % CI −2.23, 1,29, P = 0.60, heterogeneity I² = 8 %) (see Figure A1 in Additional file 2). A total of seven RCTs reported on hospital LOS (with mean and standard deviation) where no significant difference was found between EN and PN (WMD −0.67, 95 % CI −1.57, 0.24, P = 0.15, heterogeneity I² = 2 %; Fig. 3, Panel b). The non-significant difference remained in the subgroup analysis where trials were aggregated according to the caloric intake across groups (RR −0.67, 95 % CI −1.57, 0.24, P = 0.15, heterogeneity I² = 2 %; test for subgroup differences: P = 0.08) (see Figure A2 in Additional file 2).

A total of four RCTs reported on length of mechanical ventilation (in mean and standard deviation) with no overall effect observed (WMD −0.38, 95 % CI −0.98, 0.21, P = 0.21, heterogeneity I² = 0 %, Fig. 3, Panel c).

According to the subgroup analyses, there was no effect of either route of nutrition on overall mortality in high-quality trials (N = 7 RCTs; RR 1.05, 95 % CI 0.94, 1.16; P = 0.38, heterogeneity I² = 0 %) compared to low-quality trials (N = 9 RCTs; RR 1.00, 95 % CI 0.62, 1.60; P = 1.00, heterogeneity I² = 30 %; test for subgroup differences: P = 0.85). This also applied to the comparison of older (N = 8 RCTS ≤ 1995; RR 1.01, 95 % CI 0.56, 1.83; P = 0.98, heterogeneity I² = 33 %) versus more recent publications (N = 8 RCTs > year 1995; RR 1.05; 95 % CI 0.94, 1.16; P = 0.38, heterogeneity I² = 0 %; test for subgroup differences: P = 0.90). With respect to infectious complications, the positive treatment effect of EN remained significantly independent of methodological trial quality (N = 4 RCTs with score ≤ 7; RR 0.42; 95 % CI 0.23, 0.77; P = 0.005; heterogeneity I² = 6 % and N = 7 RCTs with score > 7; RR 0.76; 95 % CI 0.58, 0.99; P = 0.04; heterogeneity I² = 37 %; test for subgroup differences: P = 0.08) (see Figure A3 in Additional file 3) and independent of the trial publication date (N = 6 RCTs > 1995; RR 0.61; 95 % CI 0.37, 0.99; P = 0.05, heterogeneity I² = 55 % and N = 5 RCTs ≤ 1995; RR 0.62; 95 % CI 0.39, 0.98; P = 0.04, heterogeneity I² = 39 %; test for subgroup differences: P = 0.94) (see Figure A4 in Additional file 4). Results regarding the treatment effect of EN versus PN on ICU and hospital LOS in both subgroup analyses also remained concordant compared with the primary analyses. Funnel plots for all outcome measures were created to further test for potential publication bias. The test for asymmetry was found to be significant for the reported endpoint infectious complications (P = 0.003) (Fig. 4). No significant differences were found with respect to the remaining endpoints of overall mortality (P = 0.61), ICU LOS (P = 0.34), hospital LOS (P = 0.30) or mechanical ventilation (P = 0.65) (data not shown).

Six previous meta-analyses comparing the effect of EN versus PN on clinical outcomes of critically ill adults reported no significant overall difference in mortality between the two routes of nutrition [6, 7, 9, 23‐25]. The meta-analysis by Simpson and Doig [25] including 11 RCTs only showed a mortality benefit of PN in a predefined subgroup analysis where PN was compared to delayed EN (odds ratio [OR] 0.29, 95 % CI 0.12, 0.70, P = 0.006; heterogeneity I² = 0 %, statistical heterogeneity P = 0.60). In the absence of an overall mortality effect, all of these meta-analyses showed significant reductions in the rate of infectious complications with the use of EN and one meta-analysis also showing a significant reduction of hospital LOS (WMD = 1.20 days; 95 % CI 0.38, 2.03; P = 0.004) [9].

In contrast to these previous meta-analyses, our updated results include the data of the recent multicenter RCT (“Calories trial”) by Harvey and coworkers [16]. In this pragmatic trial including 2400 critically ill patients, the use of total PN was compared with EN for a duration of 5 days after ICU admission. The calorie and protein intake was similar, albeit low with respect to the predefined nutrition target in both groups, while initiation of oral feeding was allowed if clinically indicated during the intervention period. Neither were there significant differences in the 90-day mortality primary endpoint nor in the rate of infectious complications or other secondary outcome measures including LOS variables and duration of mechanical ventilation. Why the increased infectious morbidity seen with PN in the former trials and meta-analyses was not observed in this latest largest RCT may have been related to the equally hypocaloric delivery of macronutrients via both routes.

In contrast to the results of the Calories trial [16], the overall significant positive treatment effect of EN on infectious morbidity and ICU LOS remained in our updated meta-analysis. On the one hand, this overall treatment effect may reflect the positive non-nutritional effects of EN in terms of preservation of gut integrity and intestinal microbial diversity as well as promotion of gut-mediated immunity, all of which may support the systemic immune response [2, 26]. Presumably, the duration of the intervention (5 days) in the Calories trial [16] was likely too short or not intense enough in a population with a mortality rate of 34 % that the beneficial non-nutritional, trophic effects of EN became evident with respect to infectious morbidity and ICU LOS. On the other hand, our subgroup findings support the hypothesis that not the route per se but rather likely the dissimilar amount of calories delivered may have influenced the treatment effect on infectious complications. While there was a significant treatment effect difference in the subgroup of five RCTs where caloric intake was reported to be significantly higher in the PN group, no treatment effect on infectious complications was observed in the five trials with similar caloric intake across the study groups. This latter subgroup observation was largely driven by the Calories trial contributing 23.8 % of the overall estimate [16]. Furthermore, studies showing differences in infectious complications were associated with publication bias. Except for one RCT [27], the caloric intake in the PN group was gradually increased to target in all aggregated trials. Still, PN as compared to EN poses a higher risk of providing macronutrients in excess of the metabolic capacity, particularly in the early phase of illness and in the absence of appropriate metabolic control [28]. Caloric overfeeding per se is regarded to negatively influence outcomes with an increased risk of infectious complications while a restricted delivery of macronutrients or the avoidance of overfeeding may preserve autophagy and thus likely positively influence outcome [29, 30]. Unfortunately, owing to the limited number of trials in which the relation of nutritional intakes and predefined targets were reported, we were neither able to further explore the treatment effect of nutritional adequacy nor hyperglycemia on infectious complications in more detail. With respect to the effect of EN-specific complications such as vomiting, aspiration or diarrhea on clinical outcomes, we were unable to complete another subgroup analysis due to inconsistent reporting in the included RCTs. The Calories trial revealed that PN as compared to EN was associated with a lower rate of vomiting and diarrhea but a higher rate of constipation [16]. However, the impact of EN-specific complications remains inconclusive given the non-significant differences in major clinical outcomes in the Calories trial and two other large RCTs comparing the effect of different EN feeding strategies [31, 32].

Most of the RCTs that were aggregated in our meta-analysis were single-center trials reporting small total number of patients, and ten RCTs were published more than 20 years ago. Based on the hypothesis that older or methodologically weaker studies per se tended to yield the more negative treatment effect of PN, we performed additional subgroup analyses as well as funnel plots to assess risk of publication bias. In the early meta-analysis by Braunschweig et al., the positive treatment effect of EN compared to PN on infection rates was independent of the publication date and trial quality score [6]. While the positive treatment effect of EN on infectious morbidity accordingly appeared to be independent of the publication date and methodological trial quality in our meta-analysis, the funnel plot analysis revealed significant asymmetry with respect to infectious complication rates. This bias is apparently driven by the published smaller trials showing a larger treatment effect weakening the strength of the inference we can make about the overall effect of the route of nutrition on infectious complication rates.

The strengths of our meta-analysis are the comprehensive and most up-to-date search of the worldwide literature without restriction to only English-written articles, the inclusion of data from the largest and most recent RCT by Harvey and coworkers [16], the duplicate data abstraction and specific criteria for searching and analysis, and the analysis of trial quality, publication year, and publication bias. Limitations of our meta-analysis include the missing outcome data points in some of the included trials, and the small number of aggregated trials with data on the clinical endpoints ICU LOS and duration of mechanical ventilation. Further limitations are the variation in reporting the caloric intake, the time of nutrition intervention, and the definitions used for infections with or without adjudication across the trials included in the subgroup analysis. We were also unable to separate the effect of protein intake via both routes on the reported clinical endpoints among the included trials. Thus, there may be other covariates driving the observed findings that were not adjusted for in our meta-analysis. This may also pertain to possible treatment effect differences of EN versus PN among specific subpopulations of critically ill patients, such as those with a different nutritional risk upon ICU admission. The results of our meta-analysis may not be applicable to patients with a relative short-term or absolute contraindication for EN where the treatment effect of PN (as compared to standard care or no nutrition) on clinical outcomes may differ, as shown in two recent large RCTs [33, 34]. Moreover, the effect on long-term functional outcomes were not studied in the aggregated RCTs but are currently viewed as the more appropriate endpoints to be influenced by different nutritional interventions including the route and amount of nutrition [14, 35]. Lastly, we did not examine cost-effectiveness of the two strategies of nutrition due to the inconsistency of reported data in the trials. It is likely that EN will incur less cost as compared to PN as shown by a secondary analysis of the Calories trial [36].