Surveillance epidemiology and an open machine-learning benchmark of antimicrobial-resistant healthcare-associated infections in a Chinese district hospital

Abstract

Abstract Antimicrobial resistance (AMR) drives a growing share of healthcare-associated infection (HAI) mortality; however, data resources from Chinese hospitals outside university referral networks remain scarce. We conducted a retrospective analysis of 39 months of prospectively collected, hospital-wide HAI surveillance data (February 2023–April 2026) from a 1,100-bed district-level general hospital (accredited Class III Grade A) in Chongqing, China, comprising 1,240 episodes among 1,054 patients across 256,993 admissions and 1.14 million patient-days. The hospital-wide episode incidence was stable at 0.48 per 100 admissions and 0.82 per 1,000 patient-days, with no seasonal pattern or residual autocorrelation. Acinetobacter baumannii was the leading pathogen (19.4% of isolates), with carbapenem resistance reaching 92.0% hospital-wide and 98.5% in the intensive care unit, exceeding national benchmarks. Mechanical ventilation, bedridden status, central venous catheterization, and stroke were enriched among resistant infections (odds ratios 2.4–5.2; all Bonferroni-adjusted p ≤ 0.001). We then benchmarked three transparent machine-learning baselines on routinely collected, pre-infection variables using patient-grouped, 10 × 5-fold repeated cross-validation. AMR phenotype prediction achieved a mean area under the receiver operating characteristic curve of 0.74–0.75 (area under the precision–recall curve 0.55–0.56 against a prevalence of 0.31) for logistic regression and random forest, whereas infection-site classification was not clinically useful (accuracy 0.39–0.52 versus a 0.46 majority-class baseline). A previously reported AUROC of 0.91 was traced to two forms of leakage (patient overlap between partitions and inclusion of total length of stay) and is corrected here. The de-identified dataset, bilingual codebook, and complete reproducibility pipeline are deposited under permissive licences, providing an open, leakage-audited benchmark for AMR prediction research in an under-represented hospital tier.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗