Biomarkers plus urine output flag ICU AKI, dialysis risk

• Explainable XGBoost models (with SHAP — SHapley Additive exPlanations) modestly outperformed logistic regression for admission-day prediction of new-onset acute kidney injury (AKI) and renal replacement therapy (RRT): AKI AUC 0.76 vs 0.74; RRT AUC 0.92 vs 0.90 (p < 0.001).

• Top admission-day contributors were urine output, endostatin, baseline creatinine, lactate, albumin, and NGAL (neutrophil gelatinase-associated lipocalin); urine output and plasma biomarkers drove most incremental improvement.

• Operational implication: use admission urine-output signals plus endostatin/NGAL (where available) to focus pharmacist-led nephrotoxin stewardship and renal-dose reviews on patients flagged at higher AKI/RRT risk.

• AKI is common in ICU care and associated with worse outcomes: in this cohort 62% had AKI, 22% developed new-onset AKI within 48 hours, and 12% started RRT within 7 days — all linked to longer ICU stay and higher 30‑day mortality.

• Early recognition remains difficult because creatinine rises slowly and admission urine-output is often unreliable or missing, delaying nephroprotective actions.

• Identifying reliable admission-day predictors clarifies which data elements should drive risk flags, helping pharmacy prioritise renal-dose adjustments and targeted nephrotoxin reviews within clinical decision support.

• Retrospective multicentre cohort from the SWECRIT biobank of adult ICU admissions (2015–2018) at four Swedish ICUs; 4,732 admissions included — 2,603 analysed for new-onset AKI (excluded AKI present on admission) and 4,716 for RRT.

• Built XGBoost models combining admission clinical data and prospectively collected plasma biomarkers (including endostatin and NGAL) batch-analysed from the biobank; compared with logistic regression.

• Evaluated performance with repeated tenfold cross-validation and mean AUC (area under the receiver operating characteristic curve), interpreted predictors using SHAP, and used KDIGO criteria for new-onset AKI within 48 h and RRT initiation within 7 days as outcomes.

• XGBoost outperformed logistic regression: new-onset AKI mean AUC 0.76 (95% CI 0.70–0.81) vs 0.74 (95% CI 0.68–0.81); RRT mean AUC 0.92 (95% CI 0.89–0.95) vs 0.90 (95% CI 0.87–0.94); p < 0.001 for both comparisons.

• Top SHAP-ranked admission predictors: new-onset AKI — urine output, endostatin, baseline creatinine, lactate, albumin; RRT — creatinine, urine output, endostatin, NGAL, SAPS 3. SHAP dependence plots showed non-linear thresholds (urine-output inflection ≈1500 mL/day; endostatin risk rise ≈75 ng/mL).

• Cohort outcomes: overall AKI 62%, new-onset AKI 22%, RRT within 7 days 12%. XGBoost using the top-five predictors achieved AUC 0.75 (AKI) and 0.90 (RRT), with the main gains attributable to inclusion of urine output and plasma biomarkers plus XGBoost’s modelling of non-linear effects.

• Implement an ICU-admission AKI/RRT risk workqueue in the electronic health record (example: Epic) that surfaces per-patient drivers (urine output, creatinine, lactate, albumin, and endostatin/NGAL when available) to prioritise same‑day pharmacist nephrotoxin and renal‑dose reviews.

• Operationalise SHAP-informed triggers (for example, urine output < ~1500 mL/day and/or endostatin ≳ ~75 ng/mL combined with contextual labs) to route high-risk patients to ICU pharmacy and align staffing using a dashboarded queue.

• Treat model output as decision support — a radar rather than autopilot: urine output acts as an early clinical alarm and endostatin/NGAL can identify evolving injury before creatinine rises; pharmacists should use these signals to decide when to hold nephrotoxins or adjust dosing.

• Establish governance: local validation and pilot testing, playbooks for pharmacist‑led actions (with physician co-sign or communication pathways), monitoring of accuracy and false alerts, and periodic reviews of thresholds and data quality (baseline creatinine sourcing, urine‑output capture).

Strengths:

• Multicentre cohort with prospectively collected admission biomarkers combined with routine clinical data; SHAP provided interpretable, clinically relevant non‑linear thresholds.

• Rigorous modelling approach using repeated tenfold cross-validation and direct comparison to logistic regression; focus on admission‑only predictors aligns with early ICU decision points and potential pharmacy workflows.

Limitations:

• Retrospective, single‑region study without an independent external test set; generalisability and real‑time performance require external and prospective validation.

• Data and measurement constraints: baseline creatinine was imputed when missing, body weight/BMI were excluded, and admission urine output was often unreliable or estimated; creatinine may have biased RRT decisions. Absolute AUC improvements were modest.

Explainable XGBoost admission models (with SHAP) identify urine output plus plasma biomarkers (endostatin, NGAL) as actionable early signals for AKI/RRT risk and are suitable for targeted pilot deployment to triage patients and focus pharmacy interventions, contingent on local validation and biomarker availability.