For health systems in sub-Saharan Africa, the ability to predict perinatal mortality months in advance could redirect resources toward the highest-risk periods before deaths occur — making the choice of forecasting model far more than an academic exercise. Ghana's perinatal mortality rate remains a persistent challenge, and understanding which analytical tools best capture its temporal patterns has direct implications for maternal and newborn care planning.
This retrospective hospital-based time series study analyzed 192 monthly observations spanning January 2010 through December 2025, drawing on 46,108 live births and 1,152 perinatal deaths from a Ghanaian facility, yielding an observed mortality rate of roughly 25 per 1,000 births — notably below the country's pooled national estimate of 44.8 per 1,000 derived from prior meta-analysis. Four forecasting architectures were tested: the classical ARIMA model, a backpropagation neural network (BPNN), a deep learning neural network (DLNN), and a generalized regression neural network (GRNN). Evaluated on a held-out 2025 validation period, ARIMA(3,0,0) achieved the lowest root mean square error at 11.74, edging out DLNN (12.87) and GRNN (13.33), with BPNN performing worst at 14.98. Ecologically, greater antenatal care coverage correlated with lower monthly mortality rates, while other contextual factors showed opposing associations.
The finding that a parsimonious autoregressive model outperforms more computationally demanding neural network architectures is consistent with a recurring pattern in epidemiological forecasting: machine learning methods tend to underperform classical time series models when training datasets are relatively small or when the signal lacks the nonlinear complexity that deep architectures are designed to exploit. With roughly 192 data points, this study sits at the lower bound of what typically benefits neural network approaches. Critically, this is a single-hospital, observational study, limiting causal inference and generalizability. The ecological association between antenatal care and mortality also cannot rule out confounding. Nonetheless, the methodological comparison provides a practical, replicable framework that low-resource health systems across the region could adopt to anticipate mortality spikes and allocate clinical resources more proactively.