Predictive Modelling for Stillbirths and Neonatal Deaths in Sub-Saharan Africa

Study (NCOPS), and Demographic and Health Surveys (DHS). The unified dataset contains 5,996,390 birth records from 66 countries. The pipeline implements a structured five-stage harmonisation framework: (1) ethical data acquisition and governance; (2) variable mapping across 13 harmonised domains using standardised domain-prefix naming conventions; (3) value standardisation and recoding using structured case-when logic with regular-expression pattern matching; (4) linkage of environmental and climate data from ERA5, CHIRPS, SRTM, ACAG and MODIS sources at 99.3% completeness; and (5) quality assurance including range validation, cross-tabulations and logical consistency checking.

The modelling pipeline benchmarks classical statistical methods (logistic regression, generalised estimating equations), ensemble machine learning (Random Forest, XGBoost, LightGBM, CatBoost) and exploratory deep learning (multilayer perceptrons) across four prediction scenarios and two primary outcomes. Interpretability analysis uses SHapley Additive exPlanations (SHAP) values throughout. The methodology is documented in Data Harmonisation Documentation Version 6.1 and Statistical Analysis Plan Version 1 (February 2026), both publicly deposited on the Open Science Framework prior to commencement of model development analyses.