Missingness-Aware TabNet: Handling Structural Missing Data for the Interpretable Prediction of Global Maternal Mortality

Reliable, explainable prediction of the maternal mortality ratio (MMR) is challenging in global health because country-level indicators are heterogeneous and missingness is often informative rather than random. This study aims to develop and validate a missingness-aware TabNet (MA-TabNet), an attention-based framework that treats absence patterns as learnable signals while maintaining a stable feature space for country-level MMR forecasting and interpretation. We build a country–year panel from a publicly available global nutrition and health dataset and predict MMR using socioeconomic and health indicators to test whether missingness patterns add predictive signal beyond observed covariates. The model applies a distribution-aware selective masking strategy, adding missingness indicators only for variables with high missing rates; remaining gaps are handled by median imputation, with indicators retained to explicitly encode reporting uncertainty. Country codes and regional groupings are encoded as learnable embeddings, and entmax-based sequential attention is used to improve feature selection via sparse, competition-style masks under correlated determinants. Hyperparameters are tuned using Bayesian optimization, and evaluation follows a temporally realistic protocol (train on earlier years; test on a future held-out year). MA-TabNet achieves a mean absolute error (MAE) of 21.05, root mean squared error (RMSE) of 36.24, and R2 of 0.9739, outperforming strong tree-based baselines and improving on the original TabNet while avoiding the training instability observed in some transformer-style tabular models. For transparency, we report attention-derived global and local importance, compare original versus missing-mask features in model importances, and complement these with permutation-based Shapley additive explanation summaries, permutation importance (MAE drop), partial dependence plots for top drivers, and continent-stratified residual analyses to clarify how structural reporting gaps shape predictions and to support trustworthy maternal health monitoring. Overall, these findings suggest that modeling missingness as a measurable reporting signal can yield accurate, auditable forecasts that are better aligned with temporally realistic SDG 3.1 monitoring than “fill-and-forget” preprocessing.