INTRODUCTION:
Health systems often rely on fixed resistance thresholds to adjust first-line therapy, but routine data are noisy and unevenly sampled across provinces, creating brittle decisions¹. This can trigger false “change-guideline” alerts in small-n settings and miss genuine emerging drifts². We present a general, uncertainty-aware decision method that converts routine antibiograms into stable, threshold-aligned policy signals, implemented in a lightweight, containerised analytics workflow³.

METHODS:
For each organism–drug–specimen and province–year, we fit a Beta–Binomial empirical-Bayes model⁴,⁵,⁷ to obtain calibrated resistance estimates and Pr(θ>τ) relative to syndrome-specific concern thresholds (τ). To suppress one-year anomalies, we require Pr(θ>τ) ≥ 0.80 for K=2 consecutive years and a positive 3-year slope in the EB mean⁶. We prioritise actions by excess failures per 1,000 tests, computed as 1000·max(θ̂−τ, 0). The pipeline executes end-to-end from CSV input in minutes and ships as a Docker/Compose environment for full reproducibility³.

RESULTS:
Applied to South Africa’s public-sector AMR surveillance (NICD/NHLS, 2021–2024)⁸, the method produces a compact, rank-ordered list of province–pair changes where signals are both statistically credible and persistent, while de-emphasising volatile, data-sparse cells. Outputs include Empirical-Bayes summaries, stability flags, and an actionable change list for stewardship committees.

CONCLUSIONS:
This framework operationalises technological transformation for antimicrobial stewardship. The containerised workflow provides a transparent, reproducible, and scalable method to generate stable, evidence-based policy signals from routine surveillance. By combining empirical-Bayes estimation with persistence logic, it supports predictable and equitable updates to empiric therapy⁹. The approach aligns with sustainable development goals for Health and Wellbeing, promoting Good Governance through data-driven institutional decision-making¹⁰.