Triglyceride-glucose frailty index, metabolic-frailty phenotypes, and mortality in critically ill patients with acute kidney injury: Derivation, interpretation, and external validation.

BACKGROUND: Prognosis remains heterogeneous among critically ill patients with acute kidney injury (AKI). We evaluated the triglyceride-glucose frailty index (TyG-FI), a composite of metabolic burden and laboratory-based frailty, for mortality risk characterization, phenotype identification, prediction, and external validation. METHODS: We included 2230 adults with KDIGO-defined AKI from MIMIC-IV. Associations between TyG-FI and ICU, in-hospital, 28-day, 90-day, and 365-day mortality were assessed using multivariable logistic and Cox regression, restricted cubic splines, subgroup analyses, and sensitivity analyses. Consensus clustering identified metabolic-frailty phenotypes. Twelve prediction algorithms were evaluated using a 7:3 development-test split, with SHAP and LIME for interpretation. Double machine learning was used as an exploratory robustness analysis. External model validation was performed in 1831 eICU patients, and phenotype reproducibility was assessed using independent clustering and transport of locked MIMIC-IV centroids. RESULTS: Higher TyG-FI was independently associated with ICU mortality (OR, 1.524; 95% CI, 1.377-1.686) and 28-day mortality (HR, 1.234; 95% CI, 1.146-1.330), with consistent associations for in-hospital, 90-day, and 365-day mortality. The primary K = 2 solution identified lower-burden and high frailty-organ dysfunction phenotypes, with 28-day mortality rates of 14.4% and 31.5%. In the held-out test set, logistic regression and CatBoost achieved AUROCs of 0.764 and 0.769, respectively. In eICU, the corresponding AUROCs were 0.694 and 0.667. Independently derived eICU phenotypes had 28-day mortality rates of 11.0% and 25.5%, while transported phenotypes had rates of 10.7% and 26.2%, with 95.4% agreement. Double-machine-learning estimates remained positive across sensitivity analyses. CONCLUSIONS: TyG-FI was consistently associated with mortality and delineated a reproducible lower-burden versus high frailty-organ dysfunction phenotype structure. It may support multidimensional risk characterization in ICU-AKI, although its incremental predictive advantage over FI-Lab alone or additive formulations was limited.