AMR surveillance in a district hospital exposes the scale of carbapenem resistance
Antimicrobial resistance is no longer an abstract threat reserved for tertiary referral centers. In this 1,100-bed district-level general hospital in Chongqing, China, healthcare-associated infection surveillance captured 1,240 episodes among 1,054 patients across 256,993 admissions and 1.14 million patient-days over 39 months. The burden was steady rather than episodic: the hospital-wide incidence held at 0.48 per 100 admissions and 0.82 per 1,000 patient-days, with no seasonal pattern and no residual autocorrelation. In practical terms, the infections were persistent, embedded in everyday care, and not driven by a transient outbreak wave.
The dominant pathogen was Acinetobacter baumannii, responsible for 19.4% of isolates. Carbapenem resistance reached 92.0% across the hospital and 98.5% in the intensive care unit, levels that sit above national benchmarks and leave very little room for empiric optimism when this organism is suspected. A resistant organism with that degree of drug failure forces clinicians to treat each case as a narrow-window decision, especially in units where ventilatory support, invasive lines, and immobility are common.
Patient-level features lined up with resistant infection
Certain bedside characteristics clustered tightly around resistant episodes. Mechanical ventilation, bedridden status, central venous catheterization, and stroke were all enriched among resistant infections, with odds ratios ranging from 2.4 to 5.2 and Bonferroni-adjusted _p_ values at or below 0.001. These are not trivial associations. Each reflects a recognizable pathway into vulnerability: interrupted airway defenses, prolonged device exposure, impaired mobility, and neurologic injury that can compromise swallowing, secretion clearance, and functional reserve.
For infection control teams, the pattern is familiar but still consequential. Patients with these features tend to accumulate exposures that make colonization and progression to infection more likely, while the repeated use of broad-spectrum antibiotics intensifies selective pressure in the background. Resistant flora flourish where invasive support is prolonged and host defenses are already strained.
The prediction models worked only within a narrow band
The study then tested three transparent machine-learning baselines using routinely collected, pre-infection variables. Patient-grouped, 10 × 5-fold repeated cross-validation was used, which matters because it limits the kind of leakage that can inflate apparent performance. In AMR phenotype prediction, logistic regression and random forest each achieved a mean area under the receiver operating characteristic curve of 0.74 to 0.75, with area under the precision-recall curve of 0.55 to 0.56 against a prevalence of 0.31.
Those numbers are usable for screening research but not for bedside certainty. They suggest that ordinary admission and clinical descriptors carry enough signal to separate some resistant from susceptible infections, yet not enough to replace microbiology or justify overconfidence at the point of care. A model that reaches the mid-0.7s in AUROC can support risk stratification, but its clinical value depends on how it performs in the exact local population where decisions are made.
Infection-site classification performed much worse. Accuracy ranged from 0.39 to 0.52, compared with a 0.46 majority-class baseline, which means the models offered little practical gain over guessing the most common class. Site prediction is often harder than phenotype prediction because anatomical origin can be influenced by mixed clinical trajectories, overlapping procedures, and heterogeneous documentation, all of which dilute the signal available in routine data.
A previous high score collapsed once leakage was removed
One of the most consequential findings was methodological rather than microbiologic. 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. Once those problems were removed, the estimate fell into a far more realistic range.
This correction matters because leakage can make a model look decisive when it is really learning the answer key embedded in the dataset. Total length of stay is especially problematic in prediction settings because it often reflects events that unfold after infection has already occurred, turning a retrospective clue into a false pretender as a pre-infection predictor. The revised benchmark is therefore less glamorous and far more useful: it tells researchers what can actually be extracted from clean, routine data in a real hospital workflow.
An open dataset creates a benchmark where one was missing
The de-identified dataset, bilingual codebook, and reproducibility pipeline were deposited under permissive licences, creating an open benchmark for AMR prediction research in an under-represented hospital tier. That is important because most published model development still leans on large academic centers, where case mix, staffing, microbiology capacity, and referral patterns can differ sharply from district hospitals. A model trained only in a high-resource university network can fail quietly when moved into a setting with different patient acuity and different resistance ecology.
Open, leakage-audited data make it possible to compare methods on equal terms. They also force a more honest standard: if a model cannot survive patient-level separation and exclusion of post-outcome variables, it should not be presented as clinically ready. The value of this benchmark lies in its restraint. It does not promise that machine learning will replace surveillance or microbiology; it defines the boundary between plausible prediction and overfit illusion.
Hospital infection control still depends on surveillance before prediction
The surveillance side of the project remains central. Prospective, hospital-wide monitoring over nearly three years showed that the burden of antimicrobial-resistant healthcare-associated infection was stable but substantial, with a pathogen profile dominated by carbapenem-resistant Acinetobacter baumannii. In the intensive care unit, where invasive support concentrates the sickest patients, resistance reached its highest levels and the margin for empiric therapy narrowed further.
This is where prediction tools may eventually help: not by replacing culture results, but by prioritizing suspicion, informing early empiric choices, and directing stewardship resources toward the highest-risk patients. Still, the model performance here leaves the field with a realistic constraint. Routine variables can identify some risk, but they cannot yet carry the burden of clinical decision-making on their own, especially when resistance is this entrenched and the underlying hospital ecology is so difficult to shift.