A district-level general hospital in Chongqing recorded stable rates of healthcare-associated infections over 39 months, but its leading pathogen showed carbapenem resistance far above national reference levels, according to a study published in Nature. The work also releases an open dataset and machine-learning benchmark for predicting resistant infections, after the authors found that an earlier, higher-performing model had been inflated by data leakage.
The hospital, a 1,100-bed Class III Grade A facility, logged 1,240 infection episodes among 1,054 patients between February 2023 and April 2026. Those episodes arose across 256,993 admissions and about 1.14 million patient-days, the paper states.
Why it matters
Antimicrobial resistance is responsible for a rising portion of deaths from infections acquired in hospitals, yet surveillance data from Chinese hospitals outside the university referral network are hard to come by. Most published work comes from large academic centres, leaving district-level institutions — which serve a large share of patients — poorly represented in the evidence base.
The Chongqing study addresses that gap with nearly three and a quarter years of prospectively collected, hospital-wide surveillance, and it goes further by publishing the underlying data and code so other researchers can test their models against it.
What the surveillance found
The hospital-wide burden of infection was steady. Nature reports an incidence of 0.48 episodes per 100 admissions and 0.82 per 1,000 patient-days, with no seasonal variation and no lingering correlation between successive time points.
Acinetobacter baumannii was the most frequently isolated organism, accounting for 19.4% of isolates. Resistance to carbapenems — a class of last-resort antibiotics — reached 92.0% across the hospital and 98.5% in the intensive care unit. The authors describe those figures as higher than national benchmarks.
Four factors were more common among patients with resistant infections: mechanical ventilation, being bedridden, central venous catheterisation and stroke. The associated odds ratios ranged from 2.4 to 5.2, and all remained significant after Bonferroni correction, with p values at or below 0.001.
A benchmark, and a correction
The research team then tested three transparent machine-learning baselines using only variables recorded before infection occurred, drawn from routine hospital data. Validation was repeated across folds and grouped by patient so that records from the same person could not appear in both training and test sets.
For predicting resistance phenotype, logistic regression and random forest models reached mean areas under the receiver operating characteristic curve of 0.74 to 0.75. The area under the precision–recall curve was 0.55 to 0.56 against a prevalence of 0.31. Classifying infection site, by contrast, did not perform usefully: accuracy of 0.39 to 0.52 compared with a majority-class baseline of 0.46.
The paper also revisits a previously reported area under the curve of 0.91. The authors traced that result to two sources of leakage — patients appearing in more than one data partition, and total length of stay being included as a predictor — and present the corrected figures in their place.
The de-identified dataset, a bilingual codebook and the full reproducibility pipeline have been deposited under permissive licences. The authors position the release as an open, leakage-audited benchmark for resistance prediction in a hospital tier that has had little representation in this kind of research.
The study was approved by the hospital's ethics committee, which waived individual consent because the analysis used retrospective, de-identified data gathered during routine infection control work. It was supported by the Chongqing Science and Health Joint Medical Research Project, grant 2026MSXM127, and appears in Scientific Reports under an open access licence.
Watch for whether other district hospitals adopt the same surveillance and reporting template, and whether independent groups use the released benchmark to test models of their own.
