Continuous glycemic phenotype derived from latent class mixed models predicts 28-day mortality after traumatic brain injury: a two-cohort study

Abstract Acute stress hyperglycemia is common after traumatic brain injury (TBI) and is associated with adverse outcomes. However, the prognostic value of longitudinal glucose trajectories during the early ICU stay remains poorly characterized, and whether post-TBI glycemic dynamics are better captured as discrete categories or as a continuous metabolic spectrum is unknown. We aimed to characterize 72-h glucose trajectory patterns as a continuous phenotypic spectrum and evaluate their independent association with 28-day mortality in a retrospective two-cohort study using MIMIC-IV (discovery) and eICU-CRD (validation). Adult TBI patients with ≥ 3 glucose measurements within 72 h of ICU admission and ICU stay ≥ 1 day were included. Latent class mixed models (LCMM) characterized glucose trajectories. The primary exposure was the continuous posterior probability of belonging to the persistent hyperglycemia class (prob2), scaled per interquartile range (IQR) increase. Adjusted odds ratios (aOR) were estimated using Firth's penalized logistic regression. Incremental prognostic value was quantified by continuous net reclassification improvement (NRI) and integrated discrimination improvement (IDI) using the hard classification models for comparability with existing literature. LCMM identified a spectrum anchored by two trajectory patterns: “rapidly declining glucose” and “persistent stress hyperglycemia.” Each IQR increase in prob2 was associated with a 59% higher risk of 28-day mortality in MIMIC-IV (aOR 1.59, 95% CI 1.271–1.999, p < 0.001) and a 58% higher risk in eICU-CRD (aOR 1.58, 95% CI 1.326–1.876, p < 0.001). The association was robust to adjustment for diabetes, baseline glucose, insulin use, glycemic variability, and early deaths, and was stronger in non-diabetic patients (MIMIC-IV: aOR 1.70; eICU-CRD: aOR 1.63) than in those with diabetes (interaction p = 0.478 and p = 0.118). The risk signal was detectable within 24 h of monitoring and was not modified by injury severity. Trajectory information significantly improved risk reclassification (continuous NRI 0.52 for MIMIC-IV and 0.43 for eICU-CRD), particularly refining risk estimates in intermediate-risk patients, while improvement in overall discrimination was modest (IDI 0.01 and 0.02). The continuous posterior probability of a persistent hyperglycemic glucose trajectory during the first 72 ICU hours independently predicts 28-day mortality in TBI, with robust external replication. Viewing the glycemic trajectory as a continuous metabolic spectrum—rather than a discrete classifier—captures a graded stress hyperglycemia signal that adds incremental prognostic value beyond conventional severity scores and static glycemic metrics, and may facilitate individualized risk stratification.