3006002/Stunting_Prediction_Kenya

Predicting childhood stunting in Kenya using KDHS survey data. # Predicting Childhood Stunting in Kenya ## Overview A machine learning model that predicts stunting severity in Kenyan children under five from KDHS household, maternal, and child-health survey data. A stacking ensemble reached macro ROC-AUC 0.709 on held-out data, but detects only normal growth reliably and is far less accurate in several harder-to-reach regions. ## Problem Statement Predicting stunting severity (severe stunting, moderate stunting, or normal growth, per WHO height-for-age z-score thresholds of -3 SD and -2 SD) for Kenyan children under five, using household, maternal, and child-health variables from the Kenya Demographic and Health Survey Children's Recode — including region, residence type, mother's education, household wealth index, drinking water source, toilet facility, birth order and spacing, birth weight, breastfeeding duration, and antenatal care history. Primary metric: macro-averaged F1 score, reported alongside one-vs-rest ROC-AUC and per-class precision/recall. ## Dataset | Property | Details | |------------------|--------------------------------------------------------------------------| | Name | Kenya Demographic and Health Survey (KDHS) 2022, Children's Recode | | Source | The DHS Program (dhsprogram.com) — requires separate authorization to download | | Size | 19,530 children x 1,312 columns raw; 17,327 rows x 90 columns after cleaning and feature engineering | | Target variable | `haz_category`, derived from `hw70` (height-for-age z-score): Severe stunting (HAZ < -3), Moderate stunting (-3 ≤ HAZ < -2), Normal growth (HAZ ≥ -2) | | Licence | Restricted / research use — DHS Program authorization required; raw data files are not redistributed in this repo | ## Methods - **EDA**: examined the target distribution against WHO stunting thresholds, quantified missingness across predictors, and explored relationships …