Antibiotic trajectories after cardiopulmonary bypass and in-hospital mortality across critical care databases

Technical Summary: Antibiotic Trajectories After Cardiopulmonary Bypass and In-Hospital Mortality

Problem Statement
Antibiotic administration following cardiopulmonary bypass (CPB) is governed by competing clinical imperatives: preventing procedure-related infections while avoiding unnecessary antimicrobial exposure. Current retrospective database analyses often rely on a single binary variable to define antibiotic exposure (e.g., "any use" vs. "no use"). This approach fails to distinguish clinically distinct patterns, such as short perioperative prophylaxis, late escalation due to suspected infection, or sustained broad-spectrum therapy. Furthermore, multi-database studies face significant measurement challenges, including heterogeneous medication recording systems, varying routes of administration, and non-interchangeable severity scores across different critical care databases. These limitations obscure reproducible clinical signals regarding how post-CPB antibiotic management patterns correlate with patient outcomes.

Methodology
The authors conducted a retrospective multi-database cohort study utilizing data from three critical care databases: MIMIC-IV, INSPIRE, and SICdb. The study focused on 14,966 adult patients undergoing their first CPB event, excluding those with ECMO due to sparse sample sizes for trajectory modeling.

• Data Harmonization: Antibiotic records from day 0 to day 7 post-CPB were extracted and harmonized by generic name, route, and antibacterial spectrum. Local topical records were excluded, while inhaled antibiotics were retained.
• Trajectory Construction: A rule-based framework classified patients into distinct exposure trajectories based on the timing and spectrum of antibiotics:
  • Short prophylaxis (reference group).
  • Late escalation.
  • Persistent broad-spectrum therapy.
  • Early broad-spectrum de-escalation.
  • Anti-MRSA therapy without broad-spectrum exposure.
  • Persistent narrow-spectrum therapy.
  • Limited/no antibiotic exposure.
• Statistical Modeling: The primary outcome was in-hospital mortality. Multivariable logistic regression was employed to estimate associations between trajectories and mortality. Crucially, the model adjusted for age, sex, data source, and a source-standardized severity score. Because raw severity scores (SOFA, ASA, SAPS3) were not directly comparable across databases, the authors calculated a within-source, within-modality z-score to preserve local severity rankings without forcing a common metric.
• Robustness Checks: Sensitivity analyses included excluding deaths occurring within 48 and 72 hours of the index event to assess reverse causation and time-alignment bias. "Leave-one-database-out" analyses were performed to evaluate the stability of findings across sources. E-values were calculated to quantify the robustness of associations against unmeasured confounding.

Key Results

• Cohort Characteristics: Among 14,966 patients, 450 (3.0%) died before hospital discharge. The most common trajectory was anti-MRSA therapy without broad-spectrum exposure (33.8%), followed by short prophylaxis (23.3%) and persistent narrow-spectrum therapy (19.5%).
• Mortality Associations: Compared to short prophylaxis, two trajectories showed the strongest adjusted associations with in-hospital mortality:
  • Persistent broad-spectrum therapy: Adjusted Odds Ratio (aOR) 11.26 (95% CI 7.34–17.28).
  • Late escalation: aOR 10.37 (95% CI 6.99–15.38).
• Model Performance: Adding the source-standardized severity score to the model improved apparent discrimination (AUC) from 0.822 to 0.849 and reduced the Akaike Information Criterion (AIC) from 3266.69 to 3117.38.
• Sensitivity and Stability:
  • The associations for persistent broad-spectrum therapy and late escalation remained robust and even strengthened after excluding deaths within 48 and 72 hours.
  • In contrast, the association for early broad-spectrum de-escalation attenuated significantly after excluding early deaths (aOR dropped from 4.89 to 1.00 after 72-hour exclusion), suggesting this signal may be driven by early fatal illness patterns (reverse causation) rather than a distinct trajectory risk.
  • Leave-one-database-out analyses confirmed that the signals for persistent broad-spectrum therapy and late escalation were reproducible across MIMIC-IV, INSPIRE, and SICdb, although the magnitude of association varied by source.

Significance and Claims
The paper posits that its primary contribution is not the claim that antibiotics directly cause mortality, but rather the development of a data structure and trajectory framework capable of separating short prophylaxis from sustained or delayed broad-spectrum treatment across heterogeneous clinical databases.

The authors conclude that persistent broad-spectrum therapy and late antibiotic escalation during the first postoperative week serve as reproducible, clinically interpretable markers of increased in-hospital mortality risk. These trajectories capture postoperative states (e.g., ongoing infection suspicion, prolonged instability) that binary exposure measures miss. However, the authors explicitly state that these findings should be interpreted as prognostic and hypothesis-generating rather than causal. They emphasize that antibiotic trajectories are shaped by evolving infection suspicion, illness severity, and complications, and thus represent "confounding by indication." The study advocates for trajectory-based exposure definitions as a descriptive risk stratification tool for multi-database critical care research, while calling for future work to incorporate microbiology and time-varying severity to disentangle treatment response from clinical deterioration.