Modeling and forecasting neonatal mortality in Ethiopia: a comparative study using statistical, machine learning, and deep learning approaches

Abstract

Ethiopia faces alarming stagnation in neonatal mortality rates (NMR) at approximately 27 deaths per 1,000 live births, jeopardizing Sustainable Development Goal (SDG) 3.2 targets. Predictive analytics using advanced computational approaches remains underutilized for guiding interventions in low-resource settings. Using national NMR data (1977–2023) from the WHO Global Health Observatory, we compared five forecasting models, ARIMA, Prophet, Random Forest, XGBoost, and LSTM, under unified walk-forward validation. Data were transformed as required by model assumptions. Performance was assessed using RMSE and MAE on the original NMR scale (deaths/1,000 live births). LSTM achieved the lowest errors (RMSE: 0.72, MAE: 0.58 deaths/1,000 live births), outperforming statistical and machine learning approaches. Forecasts indicate marginal decline to 27.7 (2030) and 27.1 (2034), far exceeding the SDG 3.2 target of ≤ 12. Structural barriers, including rural healthcare access limitations, substandard perinatal care quality, and diagnostic delays likely contribute to this stagnation. Ethiopia requires a tenfold acceleration in NMR reduction to meet 2030 targets. LSTM-based forecasting could enable proactive resource allocation (e.g., pre-positioning supplies in high-risk districts) but should be interpreted alongside subnational data and uncertainty bounds. This study establishes LSTM as a promising tool for neonatal mortality prediction in data-constrained settings, while acknowledging the need for validation with longer time series and exogenous predictors.