Confronting the Twin Divides: Data Bias and Infrastructure Gaps in AI-Driven Nutrition for the Global South.
文献摘要
Nutrition science is being rapidly transformed by artificial intelligence (AI). Yet, its application in the Global South, here defined as low- and middle-income countries (LMICs) primarily across Sub-Saharan Africa, South Asia, and Latin America, is constrained by two interrelated challenges: data bias (e.g., non-diverse datasets, imbalanced samples, and inadequate attention to demographic, cultural, and socioeconomic variation) and infrastructure gaps (e.g., intermittent connectivity, unreliable electricity, and limited device access). These issues are not merely technical but structural, ranging from non-diverse training sets and culturally mismatched food recognition systems to limited population-specific validation, which produce biased outputs when deployed in settings with substantially different dietary patterns and food environments. Weak digital infrastructure, the absence of data governance, and limited workforce capacity further undermine equitable deployment. The Food Recognition Assistance and Nudging Insights (FRANI) application, validated in Ghana, Vietnam, and Sri Lanka, demonstrates what is technically possible but also highlights the boundaries of current progress: more than 90% of AI models for malnutrition go unused in day-to-day clinical practice. Existing reviews map these domains separately. This commentary argues that integrating data bias and infrastructure gaps within a single analytical framework is essential for equitable AI-driven nutrition in the Global South, and it calls for actionable policy responses that prioritize local data sovereignty, infrastructure investment, and governance frameworks before algorithmic deployment.