Early hemoglobin trajectories and short-term mortality in ICU patients with traumatic brain injury: development in MIMIC-IV and external assessment in eICU-CRD

Background Hemoglobin (Hb) changes dynamically after traumatic brain injury (TBI), but the prognostic relevance of distinct longitudinal patterns and their added value beyond early Hb remain uncertain. We characterized early Hb trajectories in ICU patients with TBI and examined associations with mortality after a 120-h landmark. Methods MIMIC-IV was used for model development and multicenter eICU-CRD for external assessment. Daily median Hb values across five 24-h windows were modeled using a latent class linear mixed model. The three-class MIMIC-IV model was frozen before transport to eICU. Cox models evaluated mortality from 120 h through ICU day 28, with sensitivity analyses addressing selection, landmark attrition, simpler Hb summaries, classification uncertainty, PRBC exposure, missingness, and nonproportional hazards. Results In MIMIC-IV, 2,466 patients entered trajectory analysis and 2,409 entered post-landmark analysis. Three trajectories were identified: relatively high and stable, rapidly declining, and persistently low Hb. In fully adjusted models, rapidly declining (HR 1.74, 95% CI 1.18–2.57) and persistently low Hb (HR 1.57, 95% CI 1.19–2.08) were associated with higher mortality. In eICU, 2,097 patients were classified and 1,557 entered outcome analysis; persistently low Hb remained associated with mortality (HR 1.58, 95% CI 1.03–2.43), whereas the rapidly declining association was less precise (HR 1.24, 95% CI 0.80–1.92). Adding trajectory class to the MIMIC-IV clinical model increased the C-index from 0.786 to 0.792. After adjustment for first-day Hb, the rapidly declining trajectory remained associated with mortality (HR 1.77, 95% CI 1.20–2.62), whereas the persistently low association attenuated (HR 1.27, 95% CI 0.83–1.93). Independent three-class refitting in eICU did not recover the same latent-class structure (ARI 0.169). Conclusion Early Hb trajectories after TBI showed distinct prognostic associations. Persistently low Hb had the most consistent transported association but was attenuated after first-day Hb adjustment; rapid decline retained an association in MIMIC-IV but had limited external support. Incremental discrimination was modest. These patterns should be interpreted as treatment-embedded clinical courses rather than fixed biological phenotypes or evidence for trajectory-specific transfusion thresholds.