Prediction of vancomycin-resistant phenotype in ICU-onset sterile-site Enterococcus-positive cultures: development and pseudo-temporal validation of a clinical prediction model using MIMIC-IV
In brief
An ICU model reached 0.84 AUROC predicting vancomycin-resistant enterococci
In a later-year validation group of 259 ICU admissions, the model distinguished vancomycin-resistant from susceptible enterococcal isolates with an AUROC of 0.835. Species indicators were important, and the model predicts resistance in the isolate, not whether it is causing infection; it also did not outperform treating everyone at lower decision thresholds. External validation is needed before clinical use.
- Journal
- Frontiers in cellular and infection microbiology (Q1)
- Published
- 2 October 2026
- Study design
- Retrospective cohort
- Evidence level
- Level 3, Low (CEBM 3b)
- Authors
- Wenhao Sun, Xiao-Lan Gu, Qingmei Ma, Jinxia Hou, Yanbin Chang
- PMID
- 42890140
- DOI
- 10.3389/fcimb.2026.1935284
Why clinicians should know about it
- Picked for Microbiology (medical) (top studies of the week, 18 October 2026): Model predicts vancomycin-resistant Enterococcus phenotype
Abstract
BACKGROUND: Vancomycin-resistant enterococci challenge intensive care. During the interval between organism identification and final phenotypic antimicrobial susceptibility testing, clinicians must balance delayed effective anti-VRE therapy against unnecessary anti-VRE exposure. We aimed to develop and pseudo-temporally validate a model for the vancomycin-resistance phenotype of the index isolate; the model does not determine whether the isolate represents clinically significant infection. METHODS: In this retrospective MIMIC-IV cohort, adult ICU admissions with a sterile-site Enterococcus-positive culture obtained more than 48 hours after ICU admission were assigned by de-identified anchor-year group to development (groups labelled 2008-2016; n = 552) or pseudo-temporal validation (2017-2022; n = 259). Predictors were restricted to information available by specimen collection. Elastic Net with bootstrap stability screening retained 19 of 141 candidates; four tree-based models were comparators. RESULTS: Among 811 admissions (507 resistant isolates; prevalence, 62.5%), Elastic Net achieved a validation AUROC of 0.835 (95% CI, 0.778-0.889) and calibration slope of 0.91 (0.71-1.18). Tree-model AUROCs were 0.842-0.855; non-significant pairwise DeLong tests were evidence insufficient to detect differences, not evidence of equivalence. Removing both species indicators reduced AUROC to 0.657 (DeLong P < 0.001). Development-OOF-fitted Platt scaling improved validation slopes for Random Forest, XGBoost, and CatBoost from 1.41, 2.41, and 1.46 to 0.94, 1.02, and 1.01. At the Youden index threshold, Elastic Net sensitivity was 0.884 and specificity was 0.690. In exploratory decision-curve analysis, its net benefit did not exceed treat-all at thresholds 0.10 or 0.20; the paired difference was 0.044 (95% CI, 0.014-0.071) at 0.30. CONCLUSIONS: The model predicts isolate resistance phenotype, not adjudicated infection, and is not a stand-alone treatment rule. These estimates are specific to a single-center cohort with high VRE prevalence. Independent external validation, local recalibration, clinical adjudication, and periodic updating are prerequisites for use.
Abstract as published, via PubMed.
For healthcare professionals. The summary is generated by AI from the published abstract, and the evidence level is assigned automatically from the study design on the Oxford CEBM hierarchy. Neither is medical advice. Read the full paper before changing practice.