A Framework for Quantifying the Interacting System of Determinants of Child Malnutrition: A Yemen Case Study

Abstract Background: Several machine learning models, seeking to assess and predict rates of child malnutrition, have been developed in recent years in response to a growing burden of under-five child malnutrition in conflict and crisis settings. These models aim to either identify the strongest predictors of malnutrition among children in a given context, provide a framework for predicting the nutritional status of a given child (classifying them as either malnourished or not) based on individual information about various health-related, socioeconomic, and geographical factors, or do a combination of both. However, these models are often limited in their potential to provide practical or actionable insights due to the often limited interpretability of complex mathematical models, the “black box” nature of some machine learning models, and the complex interactions that exist between various drivers of child malnutrition, which make it difficult to fully conceptualize this system in the form of a model. Methods: In order to reconcile the complex, highly accurate but hard-to-interpret outputs a machine learning model can provide with conceptual understandings of malnutrition and its interconnected contextual drivers, we sought to understand whether quantitative analysis of the determinants of child malnutrition aligned with conceptual understandings of the causal pathways of child malnutrition. In this study, we take Yemen as a case study for this analysis and using Bayesian analysis, explore the roles of diarrheal disease, food insecurity (quantified through food consumption score) and the interactions between them in driving child malnutrition in Yemen. Results: Our analysis revealed that, contrary to intuitive reasoning, a progressive increase in food consumption did not necessarily lead to progressive decrease in a child’s risk of malnutrition. We also found that the presence of both diarrhea and poor food consumption together significantly increased the risk of severe malnutrition, as compared to either one on its own. Results of all analyses revealed that exploring the impact of any given factor on child malnutrition risk independently using standard modeling techniques is insufficient for holistically considering the pathways to malnutrition and their complex interactions. Conclusion: By aiming to quantify the conceptual framework for child malnutrition this way, we can more holistically consider the pathways to malnutrition and their interactions and consider the dependencies and interactions that exist between various health-related interventions in order to maximize their impacts, rather than simply knowing that specific factors may be independently associated with malnutrition risk through quantitative analysis, without fully understanding why this is true or how this can be translated into program-level insights.