Updated meta-analysis of comparison of mortality in enteral feeding (EN) vs. parenteral nutrition (PN) or other methods in gastrointestinal cancer patients

Cl inical trials and meta-analyses investigating the effect of enteral nutrition (EN), parenteral nutrition (PN) or other methods of nutritional support are at best confusing with multiple endpoints including major and minor complications as well as mortality. Most deal with a combination of diagnoses such as Crohn's disease, head trauma, liver disease, pancreatitis, various cancers, etc. and neglect to present separate results by diagnoses causing one to seek out the source documentation. There is no up to date comprehensive meta-analysis of EN vs. other methodologies in gastrointestinal (GI) cancer. The goal here is to compare EN vs. alternative methods in GI cancers only with respect to mortality endpoint. A systematic literature (clinical trials and meta-analyses) search was performed for EN for the past 20 years using Pubmed, Medline, Cancerlit, Embase Cochrane collaborative plus other sources. Randomized clinical trials comparing EN to PN or other methods such as "nil by mouth", intravenous crystalloid solutions, and other oral supplementation were identified. The focus of the search was on articles from the cancer and clinical nutrition literature as well as abstracts and proceedings. This yielded about 14 such publications (1995-2012) with adequate data used in this analysis. Four studies had no deaths in both groups and are not presented here. The primary endpoint for our meta-analysis was incidence of mortality given the raw death incidence and sample size. Many publications provide information on the reduced risk of death of one intervention vs. the other which could be put into a meta-regression with the accumulated data from other publications. We used a Bayesian random effects model to account for heterogeneity across the studies. However, considering the restricted diagnosis of the subject pool and the similarity of treatment approaches, heterogeneity was not anticipated to be significant. The priors used were non informative and conjugate on the prior mean and variance of the odds ratio (OR) yielding a distribution for both the odds ratio and the predictive odds ratio. We also investigated plots of the distributions of these statistics. Of the data analyzed , the odds ratio (OR) of mortality in GI cancer of EN vs. other methods is about 0.819 with 95% confidence interval CI, (0.571,1.174) and the predictive OR is about 0.893, with 95% CI (0.546, 1.230). The posterior distributional plots of the OR and predictive OR yield interesting patterns over the studies. The heterogeneity as anticipated was non-significant. Assuming a wide distributional effect of EN over other methodologies there appears to be a consistent trend that, although EN is preferred because of other reduced morbidities in many publications, the effect on overall mortality of the reduced risk evidenced by this analysis is not statistically convincing across studies. However, using a Bayesian analysis, the posterior probability of the OR being less than one (in favor of EN) ranges from 0.811 to 0.99 and the predicted OR being less than one ranges from 0.544 to 0.888 for realistic prior ranges of the OR over time. On a final note, although EN is superior to other methods with respect to mortality, the effect of this superiority has appeared to decrease over time as seen in the last plot of this manuscript.