Abstract

Child acute malnutrition remains a critical challenge in South Sudan, where prevalence rates have persistently exceeded the WHO emergency threshold of 15%. Despite decades of large-scale humanitarian interventions, relapse rates remain alarmingly high: up to 63% of children either relapse or die within six months of treatment. Existing nutrition information systems provide only periodic, national- or subnational-level snapshots of prevalence and assumed immediate drivers. These approaches lack granularity, struggle to capture local and seasonal dynamics, and ultimately constrain the design of effective, preventive interventions. Responses therefore continue to rely heavily on food assistance, nutrition education, and clinical treatment, perpetuating a cycle of recurrent malnutrition rather than addressing systemic causes.
The urgency of this challenge has been exacerbated by significant reductions in funding for nutrition information and response systems in fragile and crisis contexts in 2025. Resources for data collection and analysis, already insufficient, are now shrinking further, undermining the capacity of governments and humanitarian actors to generate reliable evidence. This reinforces existing weaknesses in traditional models and further limits their potential to protect children. These dynamics make clear the pressing need for innovative, cost-effective, and sustainable systems that integrate diverse data sources, which may help reveal the spatiotemporal dynamics of malnutrition, and provide actionable insights for anticipatory and preventive action.

In this study, we aim to identify, model the spatiotemporal drivers and triggers of child acute malnutrition embedded in local livelihood systems in South Sudan.

To this end, we developed a multi-scale relational geospatial database that integrates environmental, socioeconomic, and human security indicators with food security and nutrition survey data on child anthropometry and household-level characteristics. The consolidated dataset spans 2015–2024 and covers 150,799 children under five years of age. By combining survey-derived information with Earth observation data, geospatial indicators, and curated qualitative insights from literature and local media, the system harmonizes fragmented evidence into a coherent structure that allows for fine-scale, multi-level analysis.
Statistical modeling is conducted using Bayesian hierarchical logistic regression, incorporating spatial and temporal random effects and their interaction to account for seasonal and geographic variability. Explanatory variables capture drivers at multiple scales. At the child level, these include sex, age, and morbidity. At the household level, indicators encompass food security, production systems, and access to land and water. At the community and systemic levels, drivers include conflict, displacement, market dynamics, infrastructure, resource availability, and both climate- and human-induced shocks. This multi-level design allows for a more nuanced understanding of how malnutrition pathways operate across space, time, and livelihood systems.
A novel feature of the approach is the integration of a Large Language Model (LLM)–based AI assistant that synthesizes statistical model outputs with curated qualitative evidence. The assistant follows an augmented intelligence paradigm, amplifying rather than replacing human expertise. Its interactivity allows users to query, for example, the main drivers of acute malnutrition in a given area, and to receive contextualized, evidence-based insights that combine statistical outputs with complementary information. This design strengthens interpretation and usability, bridging statistical analysis with the practical needs of decision-makers.

Preliminary findings confirm seasonal and spatial heterogeneity in both the prevalence of child acute malnutrition and the constellation of associated drivers. Seasonal Peaks in malnutrition are observed, and the timing and amplitude vary across livelihood systems. The analysis also reveals fine-scale spatial variability in malnutrition risk, identifying hotspots and high-risk areas that are often obscured in national or administrative-level averages. Key drivers include household food insecurity, conflict intensity, and vegetation dynamics, but the strength and direction of their association vary significantly across space, season, and production systems. These patterns highlight the deeply context-dependent nature of malnutrition risk, shaped by the interaction of environmental, institutional, and livelihoods systems. The AI assistant enhances the detection and interpretation of such dynamics, making outputs more directly usable for anticipatory planning and policy dialogue.

This study highlights the limitations of centralized, sector-specific, and rigid frameworks that dominate current nutrition analysis. Over-reliance on analyst assumptions and periodic cross-sectional data constrains the ability to capture variability and design effective preventive responses. By systematically integrating multiple, including non-traditional, data sources into a relational geospatial information system, our approach demonstrates how spatiotemporal analysis can strengthen early warning capacities, strengthen existing systems such as the IPC Acute Malnutrition classification, and inform locally relevant, cost-effective interventions.
For South Sudan, this work demonstrates that AI-supported, spatiotemporal analysis can fill critical evidence gaps at a time when traditional monitoring systems face severe funding constraints. For the wider field, it offers a scalable and replicable methodology applicable to other crisis-affected settings where child malnutrition persists. Ultimately, these innovations have the potential to enable earlier, better targeted, and more dignified responses that align with community needs, reduce costs, and prevent children from falling into acute malnutrition, thus contributing to stronger, more resilient livelihoods and communities.