Early serum chloride trajectory phenotypes and prognostic transportability across two critical care databases in acute brain injury

Background Studies of chloride after acute brain injury have largely relied on static measurements, and the cross-database transportability of dynamic phenotypes remains uncertain. Methods We conducted a retrospective 72-h landmark study of adults with traumatic brain injury, intracerebral hemorrhage, or nontraumatic subarachnoid hemorrhage in MIMIC-IV v3.1 and eICU-CRD v2.0. A four-class quadratic group-based trajectory model (GBTM) was derived without reference to outcomes in MIMIC-IV and transported to eICU-CRD with all parameters fixed. Fine-Gray models, with live discharge treated as a competing event, evaluated post-landmark in-hospital mortality; incident acute kidney injury (AKI) during 72–168 h was exploratory. Fifty database-specific multiply imputed datasets incorporated posterior-probability class draws. Results We included 1,817 MIMIC-IV patients (382 deaths) and 1,154 eICU-CRD patients (197 deaths). The trajectories were low-stable (C1), mid-stable (C2; reference), high-rising (C3), and very-high rise/plateau (C4). In MIMIC-IV, mortality associations differed across trajectories (global p < 0.001); adjusted subdistribution hazard ratios versus C2 were 1.87 (95% CI 1.18–2.95) for C1, 1.49 (1.11–2.01) for C3, and 2.96 (1.95–4.48) for C4. The corresponding eICU-CRD estimates were 0.77 (0.35–1.70), 0.77 (0.50–1.18), and 1.21 (0.71–2.07), with global p = 0.166. No clear association between trajectory class and incident AKI was observed in either database. Conclusion Early chloride trajectory phenotypes were identifiable in both databases, whereas their prognostic associations varied between databases. These observational findings support structural phenotype transportability while underscoring the need for external evaluation before prognostic interpretation.