Causal impact of malnutrition on mortality among adults hospitalized for medical illness in sub-Saharan Africa: what is the role of severe sepsis?

Setting and study population

We analyzed data from an observational cohort study of hospitalized adults. The study occurred on the general medicine ward of Mbarara Regional Referral Hospital in south-Western Uganda. The setting and study population have been described elsewhere [6]. All patients being admitted for any medical illness during the study period (April to June 2011) were eligible for the study, which was set up to perform multiple nutritional assessments on each patient. However, those who were less than 18 (considered unable to provide informed consent) (N = 34) and those who declined to participate in the study (N = 3) were excluded. As we were able to perform the planned nutritional assessments on all consenting patients, no patient was excluded on this basis.

Ethical approval

All participants provided written informed consent. The study was approved by the Institutional Review Committee of Mbarara University of Science and Technology.

Measurement of malnutrition

For this report, we measured malnutrition at admission using the mid-upper arm circumference (MUAC). MUAC has been validated in adults for diagnosing malnutrition and shown to be highly correlated with the body mass index while predicting outcomes better [19–21]. MUAC is in general highly specific (up to 96 %), but can be poorly sensitive depending on the chosen cut-offs [22]. We used a non-stretchable MUAC tape to obtain MUAC as the circumference of either of the patient’s arm at the midpoint between the olecranon and the acromion, with the patient seated or lying supine, and the rested arm flexed at 90⁰ at the elbow. We defined malnutrition as a MUAC <19 cm for females and <20 cm for males [19].

Other predictor measurements

We defined sepsis as the presence of suspected infection plus 2 or more of the systemic inflammatory response syndrome (SIRS) criteria (pulse ≥90 beats/min; respiratory rate ≥20 cycles /min; a temperature ≥38 °C or ≤36 °C; and white blood cell count ≥12,000 cells/cc or <4,000 cells/cc); severe sepsis was defined as the presence of sepsis plus at least 1 organ dysfunction (Glasgow coma score [GCS] < 15, systolic blood pressure <90 mmHg, mean arterial pressure <70mmhg, or platelet counts <100,000 cells/cc) [23]. All temperature measurements were axillary. Diagnoses for infections and focus of infection were based on clinical suspicion and physical examination. We measured at admission, a random blood glucose (RBS) using a hand-held glucometer (Roche Diagnostics, Basel, Switzerland), and defined hypoglycemia as RBS <4.5 mmol/l [24]. We also obtained a complete blood count (Beckman Coulter, Villepinte, France) defining (severe) anemia as a hemoglobin concentration < 8 g/dl in both males and females [25]. HIV status was determined using a standard 3-test rapid testing algorithm (Determine, Abbott Laboratories, Abbott Park, IL; Statpak, Chembio Diagnostics, Medford, New York; and Unigold, Trinity Biotech, Bray, Ireland), and tuberculosis was defined as sputum smear positive pulmonary tuberculosis, or microscopically confirmed extra-pulmonary tuberculosis, or sputum smear negative tuberculosis as diagnosed through a consensus of at least two physicians [26]. All laboratory tests were performed at the Mbarara University Clinical Research Laboratory which participates in external quality assurance by the National Health Laboratory Service (Johannesburg, South Africa).

Severity of underlying illness

We measured severity of underlying illness using the Modified Early Warning Score (MEWS) which incorporates blood pressure, pulse, temperature, respiratory rate, and the neurological status at admission and strongly predicts mortality in hospitalized adults [27]. MEWS gives: 3 points for each of systolic blood pressure <70 mmhg, pulse rate ≥130 beats/min, respiratory rate ≥ 30 cycles/min, and GCS score ≤8; 2 points for each of SBP 70–80 or ≥200, pulse 111–129, respiratory rate <9 or 21–29, temperature <35 or ≥38.5 Celsius, and GCS 9–13; and a point each for SBP 80–100, pulse 40–50 or 101–110, respiratory rate 16–20, and GCS score of 14. The accrued total score can be incorporated into analyses, usually as a categorical variable [28].

Outcome

The outcome was 30-day mortality. All patients were followed in hospital to death or discharge, and after discharge, they were further followed using mobile telephone calls and out-patient clinic appointments to determine 30-day vital status.

Data analysis

We tested the association between malnutrition and patient characteristics at admission using chi-squared or Mann–Whitney-U tests as appropriate.

For the causal analysis, we identified confounders of the relationships between malnutrition, severe sepsis, and 30-day mortality using a directed acyclic graph (DAG) [8] (Fig. 1). In order to build the final DAG and the final adjusted model, we started by drawing a large conceptual framework including all variables that we considered to be associated with malnutrition, severe sepsis, and mortality, along with the directions of hypothesized relationships. From the large conceptual framework, we selected the final adjustment variables (shown in the DAG) based on their ability to qualify as confounders of the relationships between malnutrition, severe sepsis, and mortality. Variables which in the larger framework appeared to be ancestors of malnutrition (e.g., food intake) or of severe sepsis (e.g., sepsis) but with no hypothesized independent link to mortality, as well as variables which appeared to be mediators of the association of malnutrition with mortality (e.g., anemia, with the exception of severe sepsis which was the mediator of interest), were eliminated from the final DAG (Fig. 1) and were not included in the adjusted analysis. As severity of underlying illness and socio-economic status are unmeasured, we used the MEWS as a proxy for severity of underlying illness, and education status as a proxy for socioeconomic status.

We first used a logistic regression model adjusting for the confounders to estimate the overall causal odds-ratio for the effect of malnutrition on mortality. We then used the medeff package in Stata (Stata Corp, College Station Texas) to estimate the indirect effect of malnutrition mediated through severe sepsis as well as its direct effect via other pathways [29, 30]. In contrast to conventional logistic regression, this analysis summarizes effects as causal risk differences, with the interpretation of those differences being similar to the classical interpretation of risk comparisons on the additive scale [10, 31]. The medeff package uses repeated simulation of unobserved counterfactual values of the mediator and outcome to obtain point estimates and confidence intervals for the mediation effect, and is thus similar to some forms of bootstrapping. As required for the validity of mediation analyses, we adjusted for confounders of the effects of both malnutrition and severe sepsis on mortality, as identified in the DAG [32]. In addition, we tested for an interaction between malnutrition and severe sepsis, which is also recommended for mediation analyses [33]. On the hypothesis that relatively mild cases of sepsis are less likely to result in death in hospital settings [1], we focused on mediation of the effects of malnutrition by severe sepsis rather than any sepsis. In assessing possible positivity violations [33], we checked overlap of each covariate as well as propensity scores for malnutrition in the malnourished and normally nourished groups.