Related Experiment Video
Updated: Sep 2, 2026
Assessment of Blood-brain Barrier Permeability by Intravenous Infusion of FITC-labeled Albumin in a Mouse Model of Neurodegenerative Disease
Published on: November 8, 2017
Dynamic albumin-to-RDW trajectories for time-dependent risk stratification in acute brain injury
Juan Wang1,2,3, Man-Man Xu1,2,3, Zheng Peng2,3
1Department of Neurosurgery, Nanjing Drum Tower Hospital, Affiliated Hospital of Medical School, Nanjing University, Nanjing, China.
Background:
Acute brain injury (ABI) requires early, interpretable risk stratification to support ICU risk assessment. We evaluated whether longitudinal albumin-to-red cell distribution width ratio (ARR) trajectories provide time-dependent prognostic information and support prediction with multicenter validation.
Methods:
In this multicenter, retrospective cohort study (NSICU, MIMIC-IV, eICU), latent class growth modeling identified ARR trajectories. Nonproportional hazards were addressed with time-stratified Cox models (0-7, 7-14, >14 days) and restricted mean survival time (RMST) at 7, 14, and 28 days. Multiple machine-learning classifiers were benchmarked to develop an interpretable prediction model with internal and external validation. Performance assessment included discrimination, calibration, decision-curve analysis, and SHAP-based interpretability.
Results:
Among 8,270 ICU patients (NSICU n = 5,093; MIMIC-IV n = 744; eICU n = 2,433), four ARR trajectories were identified. Relative to Gradual Decline, recovery-type trajectories were associated with lower early mortality risk (0-7 days: HRs 0.29 and 0.14; 7-14 days: HRs 0.43 and 0.50), whereas the Sustained High-Risk trajectory showed excess hazard primarily beyond 14 days (HR 1.42). These time-dependent patterns were broadly supported by RMST and sensitivity analyses. The final ExtraTrees model achieved AUROC 0.869 and Brier score 0.066 on the internal hold-out set; external validation yielded AUROC 0.731, PR AUC 0.429, and Brier score 0.161, indicating attenuated but informative external performance with modest decision-curve benefit. A web-based research calculator provides individualized risk estimates and SHAP explanations.
Conclusion:
Standardized ARR trajectories provide dynamic, phase-specific prognostic information in ABI and complement static assessments. When integrated into an interpretable machine-learning model, these trajectories may support adjunctive risk estimation as patient status evolves. Prospective evaluation and site-level recalibration are warranted.
