Stress hyperglycemia identifies a high-mortality phenotype within lower triglyceride-glucose levels in critically ill adults: a MIMIC-IV landmark cohort study

Abstract Background. The triglyceride-glucose (TyG) index is widely applied to stratify metabolic risk in critically ill patients, but it indexes chronic insulin resistance and may not capture the acute glycemic stress that dominates short-term outcome. We asked whether, among patients in the lower half of the TyG distribution, the stress-hyperglycemia ratio (SHR) identifies a distinct high-risk phenotype. Methods. We studied adults in MIMIC-IV at their first ICU stay in whom TyG and SHR were computable from clinically obtained triglycerides, glucose and HbA1c within ± 24 h of admission (N = 3,877 after a 24-hour landmark; N = 3,855 in adjusted models). The primary outcome was 28-day all-cause mortality. Patients were cross-classified by median TyG and median SHR; discordance was also modeled continuously (percentile-rank difference; SHR residualized on TyG, linear and spline). Because SHR contains glucose in its numerator, relative glycemia was additionally examined with a decoupled parameterization entering log glucose and log HbA1c-estimated average glucose (eAG) separately, with collinearity diagnostics. Prespecified analyses included additive and multiplicative interaction, inverse-probability-of-being-measured weighting, and competing-risk analysis of incident acute kidney injury (AKI). Results. Among the 1,938 patients in the lower half of the TyG distribution, 28-day mortality was 18.0% with high SHR versus 11.2% with low SHR (adjusted hazard ratio [HR] 1.62, 95% CI 1.26–2.07). This group contained 131 deaths — 23.5% of all 28-day deaths — and its mortality exceeded that of all patients in the upper half of the TyG distribution (18.0% vs 15.0%) at similar SAPS-II (32.1 vs 31.9). It was largely non-diabetic (89.1% without diagnosed diabetes), with normal HbA1c (5.5%) and admission glucose that would not trigger conventional thresholds (median 134 mg/dL; 60% below 140 mg/dL). The gradient tracked SHR rather than TyG (per-SD HR 1.18 vs 1.03 [ns]). In the decoupled model, higher log glucose was associated with higher mortality (1.33, 1.19–1.48) and higher log eAG with lower mortality (0.84, 0.73–0.96), supporting a contribution from glycemia relative to the HbA1c-estimated baseline; collinearity was modest (variance inflation 2.5–3.1; condition number 5.4). Results did not depend on the median split (continuous discordance per SD 1.24, 1.13–1.36) and were unchanged after excluding anemia and chronic kidney disease (1.68, 1.25–2.26). An unmeasured confounder would need a risk ratio of 2.13 to explain the association away. TyG×SHR interaction was null on both scales; incident AKI did not differ across groups; the apparent C-index was 0.766 for the clinical base model and 0.772 after adding the four-category TyG–SHR phenotype. Conclusions. Within the lower half of the TyG distribution there exists a high-mortality phenotype defined by relative rather than absolute hyperglycemia, which is not apparent from TyG rank, HbA1c or a conventional glucose threshold alone. These are prognostic observations from a single database that identify a group for earlier risk recognition and trial enrichment rather than a treatment target.