A Hybrid Bayesian Translational Epidemiology Framework for Precision Mop-Up Following Integrated Maternal, Newborn and Child Health Week in Kogi State, Nigeria
Year: 2026
Type: article
Abstract: Integrated Maternal, Newborn and Child Health Week (MNCHW) campaigns improve access to essential maternal, newborn and child health interventions; however, residual populations remain unreached, particularly in geographically, socioeconomically, and health-system constrained settings. Conventional post-campaign analyses predominantly describe coverage and provide limited capacity to identify who remains unreached, explain the mechanisms generating residual gaps, or spatially prioritize precision mop-up interventions. The objective of the study to develop and apply a Hybrid Bayesian Translational Epidemiology Framework provided the milieu for the integration of evidence synthesis, Bayesian machine learning, multilevel causal mediation, and Bayesian spatial epidemiology to identify, explain, predict, and spatially prioritize residual MNCHW service gaps in Kogi State, Nigeria. A sequential multi-phase hybrid translational epidemiology and implementation science design was applied to linked routine MNCHW, immunization, RMNCH, health-facility, contextual and geospatial datasets covering 21 Local Government Areas, 239 political wards and n=990 health facilities. PRISMA-ScR evidence synthesis informed selection of epidemiological, socioeconomic, environmental, accessibility and health-system predictors. Data management, descriptive epidemiology, Bayesian machine learning, causal mediation and statistical modeling were performed in R version 4.3.2. Bayesian Additive Regression Trees (BART) were used to estimate individual- and facility-level probabilities of residual underperformance, with Shapley Additive Explanations (SHAP) used to quantify predictor contributions; predictive performance was assessed using repeated cross-validation, discrimination, calibration and prediction-error measures. Multilevel Bayesian causal mediation analysis estimated direct, indirect and total intervention effects through community mobilization, CBHW activities, caregiver participation, outreach implementation, commodity availability, cold-chain functionality, geographic accessibility and health workforce capacity. Bayesian spatial analysis employed the Besag–York–Mollié 2 (BYM2) model implemented through R-INLA version 23.12, with queen-contiguity spatial structures, to estimate posterior relative risks, spatial dependence, exceedance probabilities and uncertainty. QGIS version 3.34 and ArcGIS Pro version 3.2 were used for spatial visualization, hotspot characterization and operational cartographic mapping. Outputs from BART, causal mediation and BYM2 were integrated into a composite Mop-up Priority Index incorporating predicted risk, spatial risk, service gaps, accessibility and facility readiness to guide precision microplanning and targeted mop-up. The hybrid framework demonstrated strong predictive performance, with a mean 10-fold AUROC of 0.968, mean classification accuracy of 96.6%, validated AUROC of 0.961, validated RMSE of 0.033, and Bayesian R² of 0.83. Posterior coefficients indicated that Health Equity Index (β=0.73, 95% CrI: 0.58–0.87) and flood vulnerability (β=0.42, 0.26–0.57) increased residual service-gap risk, whereas cold-chain uptime (β=−0.68, −0.82 to −0.53), outreach frequency (β=−0.54, −0.66 to −0.41), facility readiness (β=−0.46, −0.60 to −0.32), and HRH density (β=−0.39, −0.52 to −0.25) were protective, with posterior probabilities >0.99. Multilevel mediation demonstrated important indirect implementation effects through outreach (β=0.46; approximately 44.6% mediated), cold-chain uptime (β=0.31; 42.5%), caregiver participation (β=0.41; 41.6%), commodity availability (β=0.27; 37.2%), geographic accessibility (β=0.29; 35.5%), and session implementation (β=0.24; 34.7%). The integrated pathway showed a total effect of 1.25 (95% CrI: 1.09–1.41; posterior probability >0.999). BYM2 analysis demonstrated predominantly structured spatial variation (φ=0.62) and strong geographic dependence (spatial correlation=0.71; 95% CrI: 0.63–0.79). Posterior relative risk was highest in Ankpa (RR=2.34; 95% CrI: 1.89–2.81), Dekina (2.08; 1.71–2.49), Bassa (1.96; 1.60–2.36), and Igalamela-Odolu (1.84; 1.51–2.22), identifying these LGAs as major residual zero-dose hotspots. Bayesian policy simulation predicted improvement in programme coverage from 69.4% (95% CrI: 65.2–73.5) under the status quo to 82.6% with improved cold-chain systems, 84.8% with combined outreach and HRH strengthening, 93.7% with an integrated intervention package, and 95.2% when the integrated package was combined with Health Equity Index targeting. The posterior probability of achieving at least 90% coverage increased from 0.42 under the status quo to 0.998 under the integrated HEI-targeted strategy. Residual MNCHW gaps were heterogeneous, spatially structured, and strongly associated with contextual, environmental, accessibility, and health-system determinants. The findings demonstrate that software-enabled integration of R-based Bayesian machine learning, SHAP explainability, multilevel causal mediation, INLA-BYM2 spatial modeling, and QGIS/ArcGIS geospatial translation can convert routine programme data into actionable implementation intelligence. The Hybrid Bayesian Translational Epidemiology Framework therefore provides a reproducible pathway from evidence synthesis → prediction → causal explanation → spatial prioritization → precision mop-up, with potential to strengthen health equity, primary healthcare performance, resource allocation, and progress toward Universal Health Coverage.
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: bronze
OpenAlex ID: W7208817837
