Trend-Based Risk Score Flags High-Risk ICU Bloodstream Infections

Predictive Model DevelopmentPredictive External Validation

Authors:Shijun Zhou; Xilei Cai; Xiujuan Yang, et al.

Journal:Frontiers in Cellular and Infection Microbiology

DOI:10.3389/fcimb.2025.1636886

By turning routine ICU labs and their trajectories into a 10‑variable XGBoost risk score, this study offers pharmacists a way to proactively flag ICU bloodstream infection patients at highest 28‑day mortality risk, sharpening antimicrobial optimization and renal dosing priorities—though external performance signals a clear need for local validation and recalibration.

November 14, 2025

Quick Take

A 10-variable XGBoost prognostic model using routinely collected ICU data and dynamic rate-of-change laboratory features predicted 28-day mortality after ICU-acquired bloodstream infection (ICU-BSI) across 5,399 patients, with AUROC 0.81 internally (Anhui Medical University, AMU) and 0.71 on external validation (MIMIC-IV).

For pharmacists, the score can function as a prioritization cue to focus antimicrobial optimization, renal dosing and monitoring, and intensified surveillance on highest-risk ICU-BSI patients while prompting systematic review and possible de-escalation for lower-risk patients to preserve stewardship resources.

Why it Matters

ICU-acquired bloodstream infections are a common, high-mortality healthcare-associated problem in intensive care units with substantial clinical and economic burden; early prognostication can inform timely interventions.

Conventional tools are typically static or delayed by culture turnaround and therefore miss rapid trajectories in platelets, creatinine, bilirubin, procalcitonin, and ventilation exposure that matter for bedside pharmacotherapy decisions.

A prognostic model built from time-aligned, routinely collected ICU data directly addresses pharmacy operational needs for decision support—helping target stewardship efforts and scarce clinical resources toward ICU-BSI patients most likely to benefit.

What They Did

Retrospective cohort of adult ICU patients with blood cultures drawn ≥48 hours after ICU admission from two datasets: AMU (China, n=1,903; split 80/20 for training/internal validation) and MIMIC‑IV (USA, n=3,496) for external validation.

Developed an XGBoost model using 33 routinely available variables (demographics, Charlson Comorbidity Index, labs, treatments) plus time-aligned rate-of-change laboratory features computed relative to the blood-culture timestamp (t0); then simplified the model to 10 variables using SHapley Additive exPlanations (SHAP) importance rankings.

Hyperparameters were tuned with grid search and cross-validation; missing data for variables with <50% missingness were handled with Multiple Imputation by Chained Equations (MICE, m=50) and pooled using Rubin’s rules; variables with >50% missingness were excluded.

Primary outcome was 28-day mortality from culture time; evaluated internal (AMU test set) and external (MIMIC‑IV) performance, assessed subgroup performance by pathogen type, examined calibration and decision-curve net benefit, and published a local web app plus code on GitHub to support reuse.

What They Found

The simplified 10-variable XGBoost model achieved AUROC 0.93 in the training set, 0.81 internally (95% CI 0.76–0.85) and 0.71 externally on MIMIC‑IV (95% CI 0.70–0.73) across 5,399 ICU-BSI patients; cohort 28-day mortality was ≈25%.

By pathogen subgroup, internal (AMU) AUROCs were 0.91 for Gram‑positive and 0.90 for Gram‑negative infections; both declined to ~0.72 on external validation (MIMIC‑IV).

SHAP-ranked top predictors were antibiotic duration, platelet rate-of-change (plateletv1), serum creatinine (SCr) and its rate-of-change (SCrv1), duration of invasive mechanical ventilation (IMV), WBC rate-of-change (WBCv1), Charlson Comorbidity Index (CCI), total bilirubin rate-of-change, procalcitonin (PCT), and hemoglobin rate-of-change.

Notable SHAP-derived thresholds: SCr >110 µmol/L and IMV >24 hours were associated with sharply increased predicted risk; observed associations included lower observed mortality with longer antibiotic courses, recognizing this reflects correlation rather than proven causation.

Decision-curve analysis and calibration: the model showed positive net benefit internally but tended to overestimate absolute risk on MIMIC‑IV (net benefit mainly at lower thresholds); single-variable AUROCs were modest, indicating that the multivariable, dynamic model added discriminative value.

The principal methodological driver of improved discrimination was incorporation of time-aligned rate-of-change laboratory features rather than single timepoint values alone.

Takeaways

Expect future ICU-BSI tools to emphasize trajectories—falling platelets, rising creatinine or bilirubin, persistent high PCT—rather than single laboratory values, which aligns with pharmacists’ practice of trend-based assessment.

View a prognostic score like this as a smart prioritization cue to identify patients who merit earlier antimicrobial optimization, renal dosing adjustments, and closer monitoring, rather than as a standalone directive to change therapy.

Because discrimination fell and risk was overestimated on external data, any implementation needs local validation and recalibration; predictions should augment—not replace—pharmacist judgment and bedside assessment.

Be cautious interpreting associations such as longer antibiotic duration appearing protective: these reflect illness course and survivorship, not causal benefit. Stewardship decisions should continue to follow guideline-directed durations and individualized clinical review.

Strengths and Limitations

Strengths:

Large multicenter ICU-BSI cohort, rigorous XGBoost tuning with grid search/cross-validation, multiple imputation (MICE) for missing data, and TRIPOD‑AI–aligned reporting support internal validity.

Inclusion of dynamic rate-of-change laboratory features, SHAP-based simplification to 10 widely available variables, and an openly shared web tool and code enhance interpretability and potential for implementation and local testing.

Limitations:

External validation in MIMIC‑IV showed reduced AUROC and systematic risk overestimation, indicating domain shift and the need for local recalibration before clinical deployment.

Retrospective design with simulated dynamic features, inclusion of some variables available only after the prediction time (which can introduce look-ahead bias), and focus specifically on ICU-BSI constrain real-time performance and generalizability to other populations.

Bottom Line

Promising but not plug-and-play: this dynamic, XGBoost-based ICU‑BSI prognostic score highlights the value of laboratory trajectories and may help pharmacists prioritize care, but it requires local recalibration, prospective testing, and cautious exclusion or censoring of post‑prediction treatment variables before operational use.