Forecasting Maternal Mortality in Sub-Saharan Africa to 2035: Evidence from the Global Burden of Disease 2023 Study

Technical Summary: Forecasting Maternal Mortality in Sub-Saharan Africa to 2035

Problem Statement
Maternal mortality remains a critical indicator of global health disparity, with Sub-Saharan Africa (SSA) bearing a disproportionate burden. While historical trends show a reduction in the Maternal Mortality Ratio (MMR), the region is projected to miss the Sustainable Development Goal (SDG) 3.1 target of fewer than 70 deaths per 100,000 live births by 2030. Existing literature has largely relied on descriptive, retrospective analyses of Global Burden of Disease (GBD) data or limited forecasting studies using older datasets (e.g., World Bank data up to 2015) and narrow methodological comparisons (often limited to a single machine learning algorithm like Prophet). There is a lack of rigorous, comparative studies utilizing the most recent GBD 2023 estimates (up to 2023) to evaluate whether modern machine learning (ML) and deep learning (DL) techniques outperform classical statistical methods for short-term annual health time series in this specific context.

Methodology
This study employed a secondary time-series analysis of GBD 2023 MMR estimates for Sub-Saharan Africa, covering 44 annual observations from 1980 to 2023. The core methodological innovation was a unified comparative framework evaluating eight forecasting models across three families:

1. Statistical Methods: Holt's damped-trend exponential smoothing, ARIMA (1,1,1), and log-linear regression.
2. Machine Learning (ML): Random Forest, Gradient Boosting, XGBoost, and Support Vector Regression (SVR).
3. Deep Learning (DL): Long Short-Term Memory (LSTM) networks.

To ensure robust model selection and avoid overfitting, the study utilized a walk-forward, expanding-window out-of-sample validation framework. The last ten years of data (2014–2023) served as the testing period. Models were trained on an increasing window of historical data and evaluated one step ahead at each iteration. Performance was measured using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The best-performing model was re-estimated on the full 1980–2023 dataset to generate forecasts and 95% prediction intervals through 2035. All ML and DL models utilized a lag-feature approach (using the previous three years' MMR values as predictors) and were regularized to account for the small sample size.

Key Results

• Historical Trend: Between 1980 and 2023, the MMR in SSA decreased from 674.5 to 396.2 deaths per 100,000 live births (a 41.3% reduction). The Estimated Annual Percentage Change (EAPC) for the decade 2013–2023 was -2.17%.
• Model Performance: Classical statistical models significantly outperformed ML and DL approaches.
  • Best Performers: ARIMA (RMSE = 2.74, MAPE = 0.55%) and Holt's damped-trend exponential smoothing (RMSE = 2.75, MAPE = 0.52%) achieved the lowest errors.
  • Machine Learning: ML models (Gradient Boosting, XGBoost, Random Forest, SVR) performed moderately well but with errors an order of magnitude higher than the best statistical models (RMSE ranging from 10.15 to 16.61).
  • Deep Learning: The LSTM model failed catastrophically, producing an RMSE of 388.33 and a MAPE of 88.52%, effectively failing to learn any meaningful pattern from the limited data.
  • Log-Linear Regression: This simple trend extrapolation performed poorly (RMSE = 46.57) due to its inability to capture the deceleration of the historical trend.
• Future Projections: Based on the selected ARIMA (1,1,1) model, the MMR is forecasted to decline to 344.7 (95% CI: 301.7–387.7) by 2030 and 313.0 (95% CI: 231.6–394.4) by 2035.

Significance and Claims
The paper claims three primary contributions:

1. Empirical Model Selection: It provides the first unified comparison of statistical, ML, and DL methods for SSA maternal mortality using GBD 2023 data. The study demonstrates that for short, low-dimensional annual time series (approx. 40–50 observations), classical statistical models that account for trend and autocorrelation are empirically superior to data-hungry ML and DL methods. The catastrophic failure of the LSTM model is highlighted as evidence that deep learning requires larger datasets than available in this context to generalize effectively.
2. Updated Trajectory Assessment: The study confirms that, even with the most recent data and robust modeling, Sub-Saharan Africa is not on track to meet the SDG 3.1 target. The projected 2030 MMR is approximately five times higher than the target of 70, indicating a persistent and wide gap.
3. Methodological Lesson: The paper argues that the choice of forecasting model family should not be assumed a priori based on the popularity of ML/DL in other fields. Instead, model selection must be an empirical exercise driven by rigorous out-of-sample validation specific to the indicator and data characteristics.

Limitations
The author acknowledges several limitations: the small sample size (44 observations) limits the reliability of deep learning applications; the forecast is at the regional aggregate level and may not reflect subregional disparities; the projection extends statistical patterns without simulating the impact of specific health interventions; and the prediction intervals reflect uncertainty in the forecasting model but do not fully encompass the uncertainty inherent in the GBD estimation process itself.