Predicting the burden of acute malnutrition in drought-prone regions of Kenya: a statistical analysis

BackgroundIn drought-prone regions, timely and granular predictions of the burden of acute malnutrition could support decision-making. We explored whether routinely collected and/or publicly available data could be used to predict the prevalence of global and severe acute malnutrition, as well as the mean weight-for-height Z-score and middle-upper-arm circumference for age Z-score, in arid- and semi-arid regions of Kenya, where drought is projected to increase in frequency and intensity. MethodsThe study covered six counties of northern Kenya and the period 2015-2019, during which a major drought occurred. To validate models, we sourced and curated so-called SMART anthropometric surveys covering one or more sub-counties for a total of 79 explicit survey strata and 44,218 individual child observations. We associated these surveys predictors specified at the sub-county or county level, and comprising climate food security, observed malnutrition, epidemic disease incidence, health service utilisation and other social conditions. We explored both generalized linear or additive models and random forests and quantified their out-of-sample performance using cross-validation. ResultsIn most counties, survey-estimated nutritional indicators were worst during the October 2016-December 2019 drought period; the drought also saw peaks in insecurity and steep vaccination declines. Candidate models had moderate performance, with random forests slightly outperforming generalised linear models. The most promising performance was observed for global acute malnutrition prevalence. DiscussionThe study did not identify a model that could very accurately predict malnutrition burden, but analyses relying on larger datasets with a wider range of predictors and encompassing multiple drought periods may yield sufficient performance and are warranted given the potential utility and efficiency of predictive models in lieu of assumptions or expensive and untimely ground data collection.