Predicting the number of infant deaths per woman in ethiopia using complementary count regression and machine learning approaches

Abstract Infant mortality remains a major public health challenge in sub-Saharan Africa. Understanding factors associated with cumulative infant deaths and improving their prediction may help identify populations with higher infant-death burdens. This study examined maternal-level counts of infant deaths ever experienced by women in Ethiopia using complementary count-regression and machine-learning approaches. Data from the 2019 Ethiopia Mini Demographic and Health Survey were analyzed among 5,846 women with complete birth histories. The outcome was the cumulative number of infant deaths experienced by each woman. Because the outcome is mechanically related to the number of births, the logarithm of the number of children ever born was incorporated as an exposure offset in the count-regression models and excluded from the machine-learning predictor set. Poisson, negative binomial, zero-inflated Poisson (ZIP), and zero-inflated negative binomial models were compared using Akaike and Bayesian information criteria. Poisson Random Forest, Extreme gradient boost (XGBoost)-Poisson, and Light Gradient-Boosting Machine (LightGBM)-Poisson models were evaluated using 70% training and 30% independent testing split, with five-fold cross-validation for hyper-parameter tuning. Predictive performance was assessed using mean absolute error (MAE) and root mean square error (RMSE). The ZIP model provided the best fit among the evaluated count-regression models, with the lowest AIC (2503) and BIC (2544). In the count component, maternal age, household wealth, postnatal care, child vaccination status, maternal education, and place of delivery were statistically associated with the expected number of infant deaths. In the machine-learning analysis, XGBoost-Poisson achieved the best test-set predictive performance, with an MAE of 0.320 and RMSE of 0.570. Region had the largest relative predictive contribution (gain = 0.190), followed by marital status (0.167), maternal age (0.136), education level (0.125), and wealth index (0.117). These variable-importance measures represent predictive contribution and should not be interpreted as causal effects. Cumulative infant-death counts among women in Ethiopia were associated with demographic, socioeconomic, geographic, and maternal and child healthcare characteristics. The ZIP model provided interpretable adjusted associations, whereas XGBoost-Poisson provided the best out-of-sample prediction among the evaluated machine-learning models. The findings support the value of combining inferential count regression with count-specific machine-learning methods while emphasizing that predictive importance does not establish causality. Continued attention to maternal education, socioeconomic disadvantage, postnatal care, immunization, and access to appropriate delivery care may help inform child-survival strategies in Ethiopia.