In an observational investigation evaluating dynamic biochemical patterns during critical illness, researchers identified four reproducible early serum chloride trajectories among adults hospitalized with acute brain injury across two independent intensive care databases, yet discovered that the mortality risk linked to those trajectories varied widely between datasets. While static measures such as admission values or peak thresholds have historically dominated risk prediction studies in brain trauma and intracranial bleeding, dynamic group-based trajectory models attempt to capture both the direction and velocity of physiological changes. In this investigation, a four-class trajectory model developed blindly with respect to outcomes in one database successfully categorized patients in a secondary database with exceptional diagnostic fidelity. However, the strong adjusted association between high-rising chloride concentrations and post-landmark mortality observed in the initial hospital cohort attenuated markedly and became statistically inconclusive in the external validation dataset, highlighting that structural reproducibility does not ensure transportable clinical prognostication.
The study cohort comprised adult patients admitted with traumatic brain injury, spontaneous intracerebral hemorrhage, or nontraumatic subarachnoid hemorrhage who were identified across two major open-access clinical resources: the single-center Medical Information Mart for Intensive Care IV (MIMIC-IV version 3.1) and the multicenter eICU Collaborative Research Database (eICU-CRD version 2.0). To avoid immortal-time bias and ensure that laboratory exposures were fully measured before tracking patient outcomes, the researchers implemented a strict 72-hour landmark framework. Patients were required to be at least 18 years of age, survive and remain physically present in the intensive care unit at 72 hours, possess an ascertainable hospital discharge disposition, and demonstrate adequate laboratory monitoring density across the first three days. The coverage rule demanded valid chloride measurements in at least four out of six 12-hour windows, including at least one observation during the initial 24 hours and at least one during the final 24 hours of the landmark period. Out of 3,239 landmark-eligible patients in MIMIC-IV, 1,817 met this coverage criterion and experienced 382 post-landmark in-hospital deaths, whereas 1,154 of 2,922 eligible patients across 106 hospitals in eICU-CRD met inclusion, among whom 197 post-landmark deaths occurred.
To model early longitudinal chloride variations without biasing the patterns toward patient survival, the researchers applied finite Gaussian mixture group-based trajectory modeling within the MIMIC-IV cohort entirely blinded to clinical outcomes. Laboratory measurements were standardized to millimoles per liter, and median values within each 12-hour window were used to temper the influence of extreme draws or irregular sampling. An outcome-blind selection evaluated two through four trajectory classes with linear and quadratic polynomial time functions across 50 random starting allocations. The four-class quadratic specification achieved a Bayesian information criterion of 51,197.73 and an entropy of 0.853, comfortably meeting all prespecified criteria for convergence, odds of correct classification, and class size. Candidate five-class models produced slightly lower information criteria but were rejected because their smallest latent classes encompassed only 3.74 percent and 3.6 percent of patients, violating the predefined 5 percent minimum threshold. The locked four-class architecture, encompassing its polynomial coefficients, common residual standard deviation, and class priors, was then transported unchanged into eICU-CRD to test fixed-parameter external fit.
The resulting trajectory phenotypes delineated distinct longitudinal courses over the initial 72 hours of intensive care: a low-stable trajectory designated C1, a mid-stable trajectory designated C2 that served as the reference standard, a high-rising trajectory designated C3, and a very-high rise and plateau pattern designated C4. In the MIMIC-IV derivation cohort, predicted chloride concentrations at the midpoints of the first and sixth windows spanned from 94.36 to 96.75 mmol/L for C1, 103.23 to 103.99 mmol/L for C2, 106.56 to 112.16 mmol/L for C3, and 110.35 to 120.88 mmol/L for C4. When classified by maximum posterior probability, 138 MIMIC-IV patients (7.6 percent) fell into C1, 822 (45.2 percent) into C2, 669 (36.8 percent) into C3, and 188 (10.3 percent) into C4, with average posterior probabilities between 0.891 and 0.950. Applying the fixed model parameters to eICU-CRD assigned 93 patients (8.1 percent) to C1, 380 (32.9 percent) to C2, 491 (42.5 percent) to C3, and 190 (16.5 percent) to C4. External diagnostic metrics confirmed robust structural transfer, yielding an entropy of 0.858, an average posterior probability range of 0.902 to 0.946, a class-by-window root mean square error of 0.93 mmol/L, and a 93.3 percent agreement rate compared with an unconstrained local model refit.
Patient characteristics across the trajectory groups revealed marked baseline imbalances and systematic differences between the two underlying hospital populations. In both databases, serum sodium concentrations closely tracked chloride shifts, with first-day median sodium rising from 133.0 mmol/L in MIMIC-IV C1 to 145.2 mmol/L in C4, and from 131.0 mmol/L to 144.0 mmol/L across the corresponding classes in eICU-CRD. The need for invasive mechanical ventilation during the initial 24 hours rose steeply with higher chloride trajectories, advancing from 38.4 percent in C1 to 76.6 percent in C4 in MIMIC-IV, and from 33.3 percent to 72.6 percent in eICU-CRD. Neurological illness severity also differed between groups and databases; while median first-day Glasgow Coma Scale scores remained at 15 across all four classes in the single-center MIMIC-IV cohort, first-day scores in eICU-CRD dropped from a median of 14 in C1 and 13 in C2 down to 9 in C3 and 7 in C4. Furthermore, the multicenter eICU-CRD population presented with a notably smaller proportion of intracerebral hemorrhage cases and higher rates of traumatic brain injury than the MIMIC-IV population, establishing substantial clinical case-mix divergence.
To evaluate post-landmark in-hospital mortality while accounting for competing risks, the investigators fitted multivariable Fine-Gray proportional subdistribution hazards models and complementary cause-specific Cox proportional hazards models across 50 multiply imputed datasets per database. Live hospital discharge was treated as a competing event because patients who left the hospital alive were no longer at risk for in-hospital death. To explicitly incorporate class-assignment ambiguity, a trajectory membership was randomly drawn for each patient from their four posterior probabilities within each imputed dataset. Fully adjusted regression models controlled for patient age, sex, specific acute brain injury type, first-day Glasgow Coma Scale score, pre-existing chronic kidney disease, diabetes, baseline serum creatinine, first-day sodium, first-day bicarbonate, lowest mean arterial pressure, invasive mechanical ventilation, vasoactive therapy, and admission sepsis status, ensuring that the estimated associations were conditional on early clinical severity and overlapping physiological parameters.
In the primary MIMIC-IV survival analysis, early chloride trajectory classification exhibited a highly significant global relationship with post-landmark in-hospital mortality, producing a three-degree-of-freedom Wald test p-value of less than 0.001. Compared against the mid-stable reference group C2, patients assigned to the very-high rise and plateau trajectory C4 experienced an adjusted subdistribution hazard ratio of 2.96, with a 95 percent confidence interval spanning from 1.95 to 4.48, reflecting a nearly threefold elevation in cumulative mortality incidence. The low-stable trajectory C1 and high-rising trajectory C3 also showed statistically significant elevations in mortality risk relative to C2, yielding adjusted subdistribution hazard ratios of 1.87 (95 percent confidence interval, 1.18 to 2.95; p = 0.008) and 1.49 (95 percent confidence interval, 1.11 to 2.01; p = 0.009), respectively. The corresponding cause-specific hazard ratio for C4 versus C2 reached 2.54 (95 percent confidence interval, 1.70 to 3.79; p < 0.001), corroborating a substantially elevated instantaneous death rate among patients remaining hospitalized and event-free.
When identical multivariable models were evaluated in the transported eICU-CRD population, the prognostic association largely dissipated, yielding an overall global p-value of 0.166 across trajectory classes. Relative to the mid-stable trajectory C2, the fully adjusted subdistribution hazard ratio for the very-high rise and plateau trajectory C4 fell to 1.21, with a wide 95 percent confidence interval of 0.71 to 2.07 (p = 0.476). The corresponding adjusted estimates for the remaining classes were 0.77 (95 percent confidence interval, 0.35 to 1.70; p = 0.522) for C1 and 0.77 (95 percent confidence interval, 0.50 to 1.18; p = 0.229) for C3, while the cause-specific hazard ratio for C4 versus C2 rested at 1.00 (95 percent confidence interval, 0.61 to 1.66). A formal interaction contrast comparing the C4-versus-C2 subdistribution hazard ratios directly between the two critical care databases produced a ratio of 2.43 (95 percent confidence interval, 1.24 to 4.79; p = 0.010), statistically verifying profound between-database heterogeneity in the transported prognostic coefficient.
The study also examined incident creatinine-defined acute kidney injury between 72 and 168 hours after intensive care admission as an exploratory secondary endpoint. Patients with pre-existing end-stage renal disease, baseline dialysis, or acute kidney injury during the initial 72-hour exposure window were systematically excluded, leaving risk cohorts of 1,421 patients with 211 kidney injury events in MIMIC-IV and 785 patients with 114 events in eICU-CRD. Trajectory classification demonstrated no clear association with subsequent acute kidney injury in either database, yielding global p-values of 0.921 in MIMIC-IV and 0.343 in eICU-CRD. The adjusted subdistribution hazard ratio for C4 versus C2 was 1.21 (95 percent confidence interval, 0.68 to 2.15) in MIMIC-IV and 1.56 (95 percent confidence interval, 0.79 to 3.07) in eICU-CRD. The authors cautioned that this secondary analysis evaluated subsequent kidney injury only among early survivors free of initial renal dysfunction and relied strictly on serum creatinine criteria without incorporating urine output, meaning these observational estimates cannot be interpreted as proof of renal safety.
Sensitivity analyses reinforced the divergence in risk prediction between the two cohorts while confirming that the underlying exposure dynamics differed fundamentally from simple summary statistics. In MIMIC-IV, trajectory classification contributed statistically significant predictive information beyond simple measures of mean chloride and rate of change (p = 0.0009), whereas class assignment added no incremental predictive value beyond mean and slope in eICU-CRD (p = 0.173). Hospital-stratified models and cluster bootstraps within eICU-CRD yielded a subdistribution hazard ratio of 0.98 (95 percent confidence interval, 0.56 to 1.72), demonstrating that multicenter hospital clustering did not explain the attenuated prognostic association. The researchers concluded that while longitudinal trajectory modeling can reliably classify dynamic chloride patterns across independent healthcare environments, clinicians must not treat statistical trajectories as universal biological entities or interpret observational associations as justification for rigid clinical chloride targets or alterations in neurocritical resuscitation practices.
Subject of Research: Transportability of serum chloride trajectory phenotypes and post-landmark outcomes in acute brain injury
Article Title: Early serum chloride trajectory phenotypes and prognostic transportability across two critical care databases in acute brain injury
Article References: Liao, L., Li, T., Ma, J., & Sun, H. (2026). Early serum chloride trajectory phenotypes and prognostic transportability across two critical care databases in acute brain injury. Frontiers in Neurology, 17, Article 1976092. https://doi.org/10.3389/fneur.2026.1976092
Image Credits: AI Generated
DOI: 10.3389/fneur.2026.1976092
Keywords: acute brain injury, serum chloride, group-based trajectory modeling, prognostic transportability, critical care, MIMIC-IV, eICU-CRD, traumatic brain injury, intracerebral hemorrhage, subarachnoid hemorrhage, in-hospital mortality, competing risks, Fine-Gray model, acute kidney injury, landmark analysis, hyperchloremia, neurocritical care
News Source: Cassandra Pierce. (November 12, 2026). Early Chloride Trajectory Phenotypes Yield Inconsistent Mortality Associations Across Critical Care Databases. Scienmag.



