EGAD: an Empirical Bayes-Driven Gittins allocation framework for imported infectious disease surveillance under resource constraints

Abstract Background Rapid global mobility expansion complicates imported infectious disease surveillance, impeding detection of asymptomatic carriers in high-volume passenger flows. Existing standard operating procedure (SOP)-based, briefing-driven, and baseline reinforcement learning (RL) models are frequently confronted with risk non-stationarity and information sparsity under constrained resources, resulting in poor detection yield, lagged risk response, and initial inference volatility. We propose an Empirical Bayes-driven Gittins Adaptive Dynamic Framework (EGAD) to address these limitations. Methods EGAD models global departure regions as dynamic decision arms, integrating three core mechanisms: (1) an enhanced Gittins index kernel for resource-constrained environments; (2) an Empirical Bayes fusion module incorporating moment-matching priors with real-time testing feedback and international epidemiological intelligence to address rare events; and (3) a dual-discount structure to capture rapid risk non-stationarity. We validate EGAD via dynamic simulation and real-world surveillance records from Sichuan Province, China (June–November 2025), with inverse probability weighting (IPW) for counterfactual evaluation. Results In simulation, EGAD achieved an overall capture rate of 89.5 % (95 % CI: 88.0–91.3 %)—a 135.8 % improvement over the SOP baseline (38.0 %; 95 % CI: 31.2–45.5 %). The asymptomatic capture rate reached 48.4 % (95 % CI: 43.5–55.5 %), exceeding the baseline RL model (28.9 %; $$P < 0.001$$ ). EGAD provided a median early-warning lead of 2.5 days, anticipating 87.5 % of simulated outbreaks. In empirical IPW counterfactual analyses over 156 days, IPW projections suggest EGAD would achieve 634.5 expected detections (95 % CI: 524.1–745.9), a 68.3 % increase relative to the SOP (377.1; 95 % CI: 312.3–445.4). Projected annually, this corresponds to approximately 600 additional cases identified beyond SOP-based screening. In the empirical setting, EGAD risk scores led briefing reports by a median of 1.33 days. Conclusions EGAD addresses the challenges of risk non-stationarity, information sparsity, and constrained resources through forward-looking Gittins correction, Empirical Bayes prior fusion with spatial spillover, and dual-discount adaptation, yielding improvements in detection efficiency, early-warning lead time, and robustness under resource constraints. The framework offers a viable paradigm for adaptive port surveillance; prospective multi-site validation is needed to establish generalisability to other resource-constrained screening contexts, including clinical targeted screening for rare diseases under limited testing capacity.