esseyz/ethiopia-malnutrition-risk-pipeline

# Ethiopia Child Malnutrition & Stunting Risk Pipeline > An end-to-end machine learning pipeline for predicting child stunting severity > using the 2016 Ethiopia Demographic and Health Survey (EDHS). --- ## Overview Child stunting — defined by the WHO as a height-for-age Z-score (HAZ) below −2 standard deviations — affects nearly **38% of Ethiopian children under five**, one of the highest prevalence rates in Sub-Saharan Africa. This project builds a reproducible ML pipeline on real-world DHS microdata to: - Characterise the sociodemographic determinants of stunting - Validate feature-target associations through statistical inference (Chi-Square + Cramér's V) - Train and evaluate models that stratify malnutrition risk into three clinical classes - Export a production-ready pipeline and Tableau-ready reporting dataset --- ## Results | Model | CV Macro F1 | Test Macro F1 | ROC-AUC (OvR) | |-------|------------|---------------|---------------| | RF Baseline (7 features) | 0.412 ± 0.006 | 0.391 | 0.631 | | RF Tuned (10 features) | 0.439 ± 0.008 | 0.431 | 0.647 | | **XGBoost (10 features)** | **0.434 ± 0.013** | **0.434** | **0.648** | > **Note on performance:** Macro F1 of ~0.43 reflects the inherent difficulty > of three-class separation on survey microdata with small individual effect > sizes (Cramér's V: 0.10–0.14). The XGBoost model demonstrates a **+11% > improvement** over the baseline RF after feature expansion and tuning. > ROC-AUC of 0.648 confirms meaningful discriminative ability beyond chance. --- ## Key Findings ### 1. Child Age is the Dominant Predictor (SHAP: 0.312) Stunting is a cumulative nutritional deficit. Older children have had longer exposure to risk factors — this growth trajectory signal is 3× stronger than any other feature. ### 2. Maternal Age & Antenatal Visits are Critical (SHAP: 0.105, 0.085) Younger mothers with fewer antenatal care visits face compounded disadvantages. ANC visits rank 3rd in SHAP importance — a directly acti …