Objectives: To develop and validate a hierarchical Bayesian decision tool that combines regional antimicrobial susceptibility surveillance data with sparse local hospital antibiograms to guide empirical antibiotic selection in rural Australian hospitals. Methods: A Beta-Binomial conjugate model combined ATLAS Oceania susceptibility data (5,613 isolates, 5 pathogens, 2018-2024) and AGAR 2023 data as regional priors, with drug-specific variance scaling. A decision-theoretic utility layer ranked regimens integrating posterior susceptibility (sigmoid threshold function), evidence-based Clostridioides difficile risk, spectrum pressure, cost, and allergies. Validation comprised Monte Carlo simulation (10,000 scenarios), decision curve analysis, and Brier score decomposition. Adjustments addressed CLSI/EUCAST breakpoint discrepancies, community-onset susceptibility correction, and co-resistance correlation. Results: Bayesian posteriors achieved 95.7% credible interval coverage (target: 93-97%) and were more accurate than local-only estimates in 97.7% of scenarios (mean absolute error 0.018 vs 0.039; Brier skill 87.5% at n=10). Each prior was equivalent to approximately 50 local isolates. Recommendations were robust across ATLAS downweighting factors (calibration 95.6-95.8%) and utility weight sets (93.9% concordance). Expected value of perfect information was <1% across all syndromes. Conclusions: Hierarchical Bayesian borrowing from regional surveillance halves susceptibility estimation error for rural hospitals with typical quarterly volumes. The decision-theoretic framework appropriately prioritises stewardship-aligned choices when susceptibility differences are above clinical thresholds. The tool is available as an open-source prototype requiring prospective validation before clinical deployment.
Hayden Farquhar (2026) studied this question.