Nationwide evaluation of district-informed bottom-up and direct national approaches for one-week-ahead forecasting of dengue hospital admissions in Bangladesh

Abstract Dengue places recurrent pressure on surveillance and hospital services in Bangladesh, yet most Bangladesh-focused forecasting studies have used national or temporally aggregated data. Whether retaining district-level information before national aggregation provides additional value for short-range forecasting remains uncertain. This study evaluated district-informed and direct national approaches for one-week-ahead forecasting of reported dengue hospital admissions across all 64 districts of Bangladesh. A nationwide surveillance–meteorological panel for 2022–2025 comprised 93,504 district-day observations aggregated into 13,312 complete district-week records. Admissions were organized by reporting district, not patient residence. A single pooled district-week Light Gradient Boosting Machine (LightGBM) generated district-specific forecasts that were summed to obtain national predictions. This bottom-up approach was compared with direct national LightGBM, a negative binomial generalized additive model, one-week persistence, and hierarchical forecasting benchmarks. Model development was temporally ordered: 2024 was used for validation and model selection, finalized specifications were refitted using all pre-2025 model-ready data, and models remained fixed throughout the 52-week 2025 evaluation. Because model specifications were revised during peer review after initial 2025 results were known, 2025 was held out from model fitting but is not treated as an entirely untouched test set. Bottom-up LightGBM achieved the lowest observed national point error in 2025 (MAE = 239.37; RMSE = 380.27; wMAPE = 12.10%; R² = 0.9625), compared with the direct national negative binomial generalized additive model (wMAPE = 14.85%), one-week persistence (15.65%), and direct national LightGBM (18.67%). At district-week level, improvement over persistence was modest (wMAPE = 28.34% vs. 29.16%). All three primary 95% circular block-bootstrap confidence intervals for the primary national wMAPE differences crossed zero, and adjusted predictive-accuracy tests did not establish statistical superiority. Hierarchical reconciliation did not reduce national point error below the unreconciled bottom-up forecast. SHapley Additive exPlanations (SHAP) identified recent admissions as the dominant predictive signal, while additional analyses showed geographic heterogeneity in forecast-error magnitude. The empirical 95% residual-based uncertainty interval covered 49 of 52 held-out weeks (94.23%). District-informed bottom-up forecasting achieved the lowest observed national point error while retaining subnational forecast information, but its advantage over strong comparators remained statistically uncertain. Prospective multi-season validation is required before routine early-warning use.