Trajectory-based inflammatory phenotyping in acute brain injury: Explainable mortality prediction and mediation analysis
Year: 2026
Type: article
Study design: Observational Study
Abstract: BACKGROUND: Acute brain injury (ABI) is clinically heterogeneous in the ICU, and immune-inflammatory responses evolve rapidly; single-time-point indices may miss temporally patterned risk signals. METHODS: Across NSICU, MIMIC-IV, and eICU, we modeled longitudinal phenotypes of the systemic inflammation response index (SIRI), aggregate index of systemic inflammation (AISI), and systemic immune-inflammation index (SII) using latent class growth modeling. Trajectory phenotypes were incorporated into an interpretable prediction pipeline, with model derivation and internal testing in NSICU and external validation in the pooled eICU-MIMIC-IV cohort. RESULTS: The cohort included 9,697 ABI patients (NSICU 5,332; eICU 3,695; MIMIC-IV 670). For each index, three phenotypes were identified: Low-Stable (LS), Mid-Dynamic (MD), and High-Plateau (HP). Relative to LS, weighted analyses showed higher mortality odds for SIRI MD (OR 1.43, 95% CI 1.20-1.70) and HP (OR 2.47, 2.07-2.95), and for HP phenotypes of AISI (OR 1.81, 1.52-2.14) and SII (OR 1.76, 1.48-2.09). Static baseline/mean tertiles attenuated after adjustment, whereas LS/MD/HP patterns preserved graded risk. In exploratory mediation analysis, the unadjusted proportion mediated through SIRI HP in the hyperglycemia-mortality association was 24.4%. SIRI phenotype membership was retained in the final model; logistic regression achieved AUROC 0.846/0.874/0.777 and AUPRC 0.415/0.459/0.374 across cross-validation, internal testing, and external validation. CONCLUSIONS: Longitudinal inflammatory phenotyping, especially SIRI trajectories, supports reproducible ABI risk stratification and interpretable prediction across ICU databases.
Language: English
Field: Medicine
Domain: Health Sciences
Open Access status: hybrid
Estimated APC paid: $3,150
OpenAlex ID: W7224872334