In a population of malnourished children admitted to an MSF nutritional program, the results suggest that WH using WHO standards was the best indicator to predict mortality under treatment with Z score and percentage of median providing nearly the same performance. Regarding the NCHS reference, the results are consistent with previous studies finding that WH% (NCHS) performed better than WHZ (NCHS) to predict mortality [9
When stratifying by age classes, all the indicators performed more poorly in the youngest age group (). Children in this group are the most vulnerable and have the highest fatality rate. The variability in the causes of deaths is likely to be important, and the proportion of deaths related to other causes than malnutrition is higher among the youngest age group than among older children. This might explain the lower AUC to estimate the relation between nutritional status and mortality risk for any indicator. It is also possible that the youngest children responded more poorly to treatment.
Strengths and weaknesses exist among the different anthropometric indicators. For example, when focusing on even the highest AUC observed (AUC = 0.79 for WHZ [WHO] or WH% [WHO] for girls 6–59 mo) the sensitivity corresponding to a specificity of 80% is 40% and the specificity corresponding to a sensitivity of 80% is 35%. Thus, the choice of indicators must be evaluated in each specific context depending on what trade-off between sensitivity and specificity is tolerable. It is also important to emphasize that the choice of a cutoff depends on other factors besides sensitivity and specificity, for example, mortality risk without treatment, adverse events and risk of treatment, and response to treatment itself. This choice must take into account the means and objectives of the entire nutritional program.
Although the results of this study are informative, the results reported here were derived from a selected sample of the general population (i.e., children admitted to the MSF nutritional program). This sample selection has implications for the interpretation of results, which cannot be extrapolated to the general population (e.g., the cutoffs WH% [NCHS] = 80% or WHZ [NCHS] = 2.0 correspond to a sensitivity of 100% for the prediction of death, which would never occur in the whole population). Moreover, since all children in the database were included in a nutritional and medical program, it is impossible to predict how these children's clinical course would have evolved had they not been treated. A possible effect of this program may be to have lessened the observed performance of all the indicators, since we can suppose that children who would otherwise have died may have survived, which reduced the link between nutritional status on admission and mortality risk.
Rather than predicting death among admitted children, a far more important question would be to know who among all the malnourished children would benefit most from admission in the program in order to reduce mortality. Our results may help to answer this question, but do not bring sufficient evidence for extrapolation. Thus, using only our results to select children for admission to a nutritional program would not be appropriate. Also, it is important to point out that although all children in the program were diagnosed as malnourished, the cause of their death may be related to the interaction between malnutrition and other co-morbid conditions, such as malaria, or any other number of factors specific to the MSF program.
Several further limitations require note. The majority of children lost to follow-up were moderately malnourished at admission to outpatient treatment. These children are at low mortality risk and the most probable scenario is that the mother of these children, after a slight improvement of the nutritional status of the child, decided to leave the program because attendance at the outpatient treatment centers was too difficult in light of their daily activities. A further limitation could be evoked because of our choice to include defaulters in this study, since nonresponse to treatment is a negative outcome. However, nonresponse to treatment was not the principal objective of this study and we instead focused on probability of death. Excluding defaulters from the sample population would not qualitatively change the results (unpublished data). Finally, children with missing information on edema were assumed not to have bipedal edema. In a complementary analysis excluding these children, the results were similar and the rank of indicators remained unchanged.
It is also important to note that our database is constructed from program monitoring data, which have a different purpose than data collected for a clinical study. For example, the frequency of children aged exactly 12, 24, 36, or 48 mo was higher than what would be expected. This lack of accuracy may have direct consequences on our analyses, especially in the analyses stratified by age, but the similar results obtained in this analysis indicate a low impact of this potential bias. Obtaining accurate data on age is an ongoing problem. Ideally, comparison of the predictive value of different indicators for the whole age range for which an intervention is considered should be conducted and not simply comparing indicators for three age groups.
Another limitation relates to the admission criteria of the program. Children were included primarily due to the WH% (NCHS) <80% criteria, and not the MUAC <110 mm. Both the NCHS reference and the WHO standards present similar values of sensitivity for defining moderate malnutrition. Applying either of these criteria means similar populations would have been selected and therefore should not affect our results. On the contrary, the selection of children using the MUAC would have differed from the current sample. This fact may partly explain the unexpected weakness of MUAC to predict deaths, whereas it had previously been identified as a useful and important screening tool to rapidly assess nutritional status and risk of death of children, particularly in emergency settings [10
]. A complementary hypothesis could suggest that the link between MUAC and deaths in previous studies may sometimes be due to additional factors (comorbidities including infections), which were mostly treated in this program.
If these results were confirmed by a study including children regardless of their nutritional status, this may have important implications for the use of MUAC as a diagnostic tool for malnutrition in therapeutic programs. MUAC changes with age and using it as tool in an emergency clearly serves a purpose as a sufficient and rapid tool, but in a stable nutritional program it is potentially limited in its use as a unique criterion. Several attempts have already been made more than 20 y ago [15
] to correct MUAC in relation to height and age, which did not improve prediction (in practice, correcting MUAC for age or height leads to the selection of older children with a lower risk of dying and is not helpful). Further studies may nevertheless be needed to assess how the MUAC could be used in association with other anthropometric or clinical criteria to increase the ability to estimate the risk of death in such contexts.
In another study performed on the same database [18
], we used multivariate logistic regression to assess which other factors were significantly related to mortality risk in a population of children with moderate malnutrition (defined with the WHZ [WHO] < −2 cutoff instead of the WH% [NCHS] cutoff), regardless of WH indicators. The results suggest that weight, height, presence of edema, and systematic use of a short list of clinical signs could be used to help identify children at highest risk of death in community-based programs in order to orientate the child towards either hospitalization or ambulatory care. The two studies address two different questions: this study examines differences in the diagnosis of malnutrition when comparing the WHO standards and the NCHS reference and in prognostic accuracy for predicting death at admission; the companion study examines clinical signs other than WH indicators as indicators of mortality risk [18
]. However, the results of both studies are derived from a sample of admitted children, and cannot be considered as a screening tool to include children in nutritional programs.
These limitations point to the importance of future studies in a population more representative of all children being screened, examining both malnourished and healthy children when assessing the sensitivity and specificity of growth references. Because there were 11 outpatient and two inpatient centers, quality of care among these different centers could have varied, thereby confounding results. We also could have stratified the analysis into children admitted into inpatient or outpatient care, in essence, asking the question concerning how the indicators respond in two different populations. Grouping them together allowed us to not classify children a priori into two different groups, one with a higher expected mortality (inpatient) than the other.
Overall, this analysis of anthropometric data from over 60,000 children in a large-scale nutritional intervention program in Niger suggests that the new WHO growth standards are more accurate indicators for mortality risk in malnourished under-5-y populations during their stay in a nutritional program, compared with the older NCHS reference. This improved accuracy appears to hold whether using Z scores or percentage of the median to measure WH. Sensitivity and specificity in relation to survival should be reexamined taking a more representative sample of children, in order to assess who would benefit most from the nutritional program in order to reduce mortality.