Dynamic TyG trajectories cumulative TyG burden are associated with in-hospital mortality in acute brain injury: a multicenter interpretable machine-learning analysis

TL;DR

This study shows that tracking triglyceride-glucose levels over time in ICU patients with brain injuries can predict mortality better than single measurements, and these metrics can be used in a machine-learning model.

Contribution

The study introduces dynamic TyG trajectory analysis and cumulative TyG burden as novel time-sensitive predictors of mortality in acute brain injury patients.

Findings

Three distinct TyG trajectories were identified, with two showing higher mortality after day 7.

Cumulative TyG burden (TBM8p7) independently predicted in-hospital mortality.

TyG metrics added prognostic value beyond standard clinical indicators in a machine-learning model.

Abstract

Dynamic metabolic changes may influence outcomes after acute brain injury (ABI), but most ICU studies use only a single triglyceride–glucose (TyG) value. We examined whether ICU TyG trajectories and a cumulative TyG burden provide time-sensitive prognostic information and can be embedded in an interpretable mortality model.

Adults with ABI from three ICU databases (NSICU, MIMIC-IV, eICU) were retrospectively analyzed. TyG trajectories were derived from serial ICU measurements, cumulative exposure was summarized as prespecified threshold-based mean area under the curve (TBM), and in-hospital mortality was evaluated with 7-day time-stratified Cox models. A machine-learning model including TyG trajectory, TBM, and routinely available clinical variables was trained in NSICU and validated in the pooled external cohort.

Among 4,760 admissions, three trajectories were identified—low–slightly increasing (LSI), moderate–increasing (MI), and persistently high (PH). Mortality did not differ across trajectories during days 0–7, but after day 7 both MI (HR 1.48, 95% CI 1.18–1.86; P < 0.001) and PH (HR 1.51, 95% CI 1.17–1.93; P = 0.001) showed higher in-hospital mortality than LSI. TBM showed a parallel positive association; TBM8p7 remained significant in fully adjusted models (HR 1.42, 95% CI 1.18–1.70; P < 0.001). ExtraTrees was selected for its consistent internal and external validation performance, and model interpretability analyses placed TyG trajectory and TBM8p7 among the next most important predictors alongside SOFA score and vasopressor use.

In ICU-treated ABI, TyG is better modeled as a time-aware exposure: trajectory differences become prognostically relevant only after the first week, whereas cumulative TBM8p7 shows a graded, independent association with mortality. Both metrics add risk information beyond conventional severity indicators and can be integrated into an interpretable, externally tested model.

ICU ABI patients from a derivation cohort (NSICU, n = 3,819) and a pooled external cohort (MIMIC-IV + eICU, n = 941) had early serial TyG measured. LCGM identified three reproducible TyG trajectories (LSI, MI, PH); in a 7-day landmark analysis, mortality curves began to diverge after day 7. A continuous burden metric, TBM8p7 (threshold-based mean supra-threshold area above 8.7 across observed intervals), was also derived. These dynamic metabolic features were incorporated into the prediction model, which was deployed as a web/mobile tool with SHAP-based individual explanations.
Flowchart outlining the study design for predicting risk in ICU patients with ABI, detailing cohort selection, exclusions, modeling steps from data preprocessing to hyperparameter tuning, evaluation metrics, and clinical application via a web-based prediction tool.

ICU ABI patients from a derivation cohort (NSICU, n = 3,819) and a pooled external cohort (MIMIC-IV + eICU, n = 941) had early serial TyG measured. LCGM identified three reproducible TyG trajectories (LSI, MI, PH); in a 7-day landmark analysis, mortality curves began to diverge after day 7. A continuous burden metric, TBM8p7 (threshold-based mean supra-threshold area above 8.7 across observed intervals), was also derived. These dynamic metabolic features were incorporated into the prediction model, which was deployed as a web/mobile tool with SHAP-based individual explanations.

Figures