Spatial Machine Learning for Exploring the Variability in Low Height-For-Age From Socioeconomic, Agroecological, and...

Area of Science:

• Public Health
• Spatial Epidemiology
• Child Nutrition

Background:

• Childhood stunting remains a critical public health issue in Rwanda.
• A granular understanding of local stunting determinants is lacking.
• Geographic variations in stunting prevalence require detailed investigation.

Purpose of the Study:

• To explore spatial heterogeneity in childhood stunting prevalence in Rwanda's Northern Province.
• To identify local risk factors and their nonlinear effects on height-for-age.
• To compare the performance of spatial regression models in explaining stunting variations.

Main Methods:

• Cross-sectional study of 615 height-for-age observations.
• Application of Ordinary Least Squares (OLS), Geographically Weighted Regression (GWR), and Multiscale Geographically Weighted Regression (MGWR) models.
• Integration of generalized additive models and explainable machine learning.

Main Results:

• 27% of children were stunted, with higher prevalence in Musanze, Gakenke, and Gicumbi districts.
• MGWR model (R²=0.89) significantly outperformed GWR (R²=0.84) and OLS (R²=0.25).
• Stunting risk increased with child isolation, elevation, and rainfall; decreased with land surface temperature. Nonlinear associations found for NDVI, slope, soil fertility, and urbanicity.

Conclusions:

• MGWR effectively captures spatial heterogeneity in stunting risk factors.
• Identifying high-stunting areas and their specific drivers is crucial for targeted interventions.
• Local-level data and advanced spatial modeling are essential for reducing childhood undernutrition.