This preprint reports a target trial emulation that estimated the per-protocol effect of different red blood cell (RBC) transfusion strategies on ICU-acquired infection among adults admitted to the intensive care unit with sepsis. The primary aim was to determine whether a restrictive transfusion strategy versus a liberal strategy, defined by hemoglobin thresholds, altered the risk of an ICU-acquired infection occurring at least 72 hours after ICU admission.
The investigators conducted a target trial emulation using the MIMIC-IV database (version 3.1). Eligible records included adult ICU stays meeting Sepsis criteria at ICU admission. The analysis followed a per-protocol framework implemented via cloning of patient trajectories and analytic techniques designed to estimate causal contrasts between predefined transfusion strategies.
Two transfusion strategies were emulated:
Clones of each eligible ICU stay were assigned to one of these strategies to emulate adherence to each policy over the ICU course.
The primary outcome was the first ICU-acquired infection occurring at least 72 hours after ICU admission during the ICU stay. The authors also examined 28-day cumulative incidences using an approach that jointly accounted for ICU discharge and ICU mortality as competing events; this provided strategy-specific risks and risk differences at 28 days.
Per-protocol effects were estimated using a clone–censor–weight approach together with a marginal structural model. This approach censors clones when they deviated from their assigned strategy and applies weights to account for informative censoring and measured time-varying confounding, aiming to approximate the causal effect of adhering to each transfusion policy.
As a complementary analysis, a parametric g-formula was used. That analysis jointly modeled ICU discharge and ICU mortality as competing events to derive strategy-specific 28-day cumulative incidences and risk differences, providing an alternative estimate that accounts for informative censoring through competing outcomes.
The analytic cohort comprised 4,013 eligible ICU stays. In the primary, marginal-structural-model–based analysis, the adjusted conditional odds ratio for ICU-acquired infection comparing liberal versus restrictive transfusion strategies was 0.954 (95% CI, 0.797 to 1.142), indicating little evidence of a difference between strategies.
In the complementary g-formula analysis, the 28-day risk difference for liberal versus restrictive strategy was −0.02 percentage points (95% CI, −0.15 to 0.11), consistent with the primary analysis and likewise suggesting no clinically meaningful difference in risk.
The authors report that findings were generally robust across prespecified subgroup and sensitivity analyses. The preprint does not provide further detail on the exact subgroups or the full set of sensitivity analyses in the abstract section; readers should consult the full manuscript or supplementary material for those specifics.
The work was reviewed by the Ethics Committee of Hiroshima University Hospital, which provided ethical approval (approval number E2026-0054). Data used were de-identified records from MIMIC-IV v3.1 available through PhysioNet to credentialed users; the authors state they are not permitted to redistribute raw data. The code and algorithms for cohort selection, variable construction, and analysis are available on the project GitHub repository cited in the manuscript.
The authors declared no competing interests. The manuscript is a preprint posted on medRxiv and has not undergone peer review; the authors and publisher caution that its findings should not be used to guide clinical practice until validated.
In this target trial emulation using a large critical-care electronic health record dataset, emulated transfusion policies defined by hemoglobin thresholds (restrictive Hgb ≤7.0 g/dL versus liberal Hgb >7.0 g/dL) produced no evidence of a clinically meaningful difference in the risk of ICU-acquired infection among adults with sepsis in the ICU. Point estimates from both the marginal structural model and the parametric g-formula were near the null and consistent with one another.
Readers should note that these results derive from an observational database using causal inference emulation methods and that the study is a preprint pending peer review. Details on subgroup definitions, sensitivity analyses, and full model specifications are reported in the full manuscript and supplementary materials accessible via the authors’ GitHub repository and the medRxiv posting.