Ward Colonisation Pressure Scores Add Little Power to Predict Hospital-Onset Infections

A large secondary analysis published in BMC Infectious Diseases has found that ward colonisation-pressure scores and antibiotic exposure histories, two variables long suspected of carrying signal about hospital-onset pathogen acquisition, added only marginal discriminatory power when tested against matched controls. The research team, led by Xiaoqing Gan and Qingjiang Lyu, took an unusually rigorous approach to the underlying data. Before any modelling, they audited the source code, exposure definitions, repeated records and matching identifiers in the released matched sample.

After data-quality exclusions, the final dataset comprised 54,934 records representing 34,444 patients and 16,650 matched case-control sets. This careful cleaning matters because matched case-control designs, while useful for comparing cases with similar controls, do not by themselves establish absolute clinical risk, and errors in exposure reconstruction can silently distort results. The authors evaluated four logistic regression models and one random forest classifier using five-fold cross-validation that deliberately kept connected patients and entire matched sets together, preventing information leakage across the training and testing splits.

Uncertainty was estimated by resampling these components. The baseline model using patient variables alone achieved an area under the receiver operating characteristic curve of 0.542, with a 95 percent confidence interval of 0.537 to 0.549, barely above the 0.5 expected from chance. Adding antibiotic courses and colonisation-pressure scores raised the area under the curve to 0.563, and the random forest reached 0.568.

Within-set concordance for the full logistic model was 0.544, meaning the model correctly identified which member of a matched pair was the case only slightly more often than a coin flip. The exploratory association analyses told a similarly restrained story. Using conditional logistic regression with component-robust confidence intervals, the researchers found that the pooled conditional odds ratio for drug-susceptible Enterobacterales colonisation pressure was 0.993 per score unit, with a confidence interval of 0.982 to 1.004, a range that includes the null value of 1.

Several antibiotic associations remained inverse, and pathogen-specific pressure associations differed markedly between the two matching branches examined in design sensitivity analyses. The results stayed weak even when colonisation pressure was omitted entirely, reduced to a simple presence indicator, or represented by training-fold ranks, indicating that no particular operationalisation of the score rescued its performance. One notable technical contribution of the paper is its numerical auditing of the colonisation-pressure implementation.

The authors found that the archived 90-day colonisation-pressure calculation was supported by their checks, lending credibility to the released scores. However, raw exposure reconstruction remained unavailable, meaning the team could not independently verify how antibiotic exposure histories had been generated from the source records. This limitation is important, because the authors are careful to state that their conclusions apply to the released scores as distributed, not to any conceivable version of these measures.

The study's conclusions are deliberately cautious. The authors write that the variables added limited discrimination in the matched sample and that the results do not establish absolute infection risk, causal protection, or usefulness for clinical triage, including low-burden interventions. This last point addresses a tempting but unsupported inference: an inverse antibiotic association in matched data cannot be read as evidence that antibiotics protect against infection, since matching and confounding structures can produce such patterns without any causal mechanism.

For infection prevention teams, the message is that ward colonisation-pressure scores, at least as currently implemented and released, are not ready to serve as triage tools for hospital-onset target-pathogen infections. The authors note that prospective clinical prediction would require fully verified exposure generation and evaluation in an appropriate population cohort, rather than a matched case-control sample. Until such cohorts exist and the exposure pipelines are independently validated, the modest area-under-the-curve values reported here, hovering between 0.54 and 0.57, stand as a reminder that plausible ward-level and medication variables can look compelling in theory yet deliver nearly chance-level discrimination in carefully validated matched data.

Subject of Research: Hospital-onset pathogen prediction using ward colonisation pressure and antibiotic exposure

Article Title: Ward colonisation-pressure scores and antibiotic exposure for distinguishing hospital-onset target-pathogen cases from matched controls: a secondary analysis with grouped internal validation and exploratory conditional associations

Article References: Gan, X., & Lyu, Q. (2026). Ward colonisation-pressure scores and antibiotic exposure for distinguishing hospital-onset target-pathogen cases from matched controls: a secondary analysis with grouped internal validation and exploratory conditional associations. BMC Infectious Diseases. https://doi.org/10.1186/s12879-026-14701-9

Image Credits: AI Generated

DOI: 10.1186/s12879-026-14701-9

Keywords: colonisation pressure, hospital-onset infection, antibiotic exposure, case-control study, machine learning, infection prevention, drug-resistant pathogens, predictive modelling, BMC Infectious Diseases, Mass General Brigham, Ward, colonisation-pressure