Research Square · Nov 16, 2026

Adjustment choices shape apparent transportability of prognostic indices: the role of cohort-specific confounders across MIMIC-IV, eICU-CRD and CHARLS

By Lin R, Lu Y, Ye C et al.

Background. Insulin-resistance (IR) surrogate indices, including the triglyceride-glucose (TyG) index, are increasingly proposed as prognostic markers. Their number is expanding rapidly, yet it is unclear whether the prognostic information they carry is transportable between care settings. We examined how omitting cohort-specific confounders from the common adjustment set alters the apparent transportability of four IR surrogate indices across care settings. Methods. Multi-database cohort study in three populations: MIMIC-IV (n = 1934; 28-day ICU mortality), eICU-CRD (n = 17262; in-hospital mortality), and CHARLS 2011 (n = 9324; mortality to 2018; 2015 wave for fasting sensitivity). ICU analyses used a 24-hour landmark. TyG, TyHGB, NHHR and TCBI used identical formulas (Table S1). Cohort-specific Cox models with three adjustment levels were fitted; per-SD hazard ratios were pooled by inverse-variance meta-analysis with I² and 95% prediction intervals. Results. TyG had a positive association estimate in every cohort (HR 1.074–1.083) and TyHGB in all three (1.065–1.128; pooled 1.078, 95% CI 1.040–1.117). Primary cohort-specific adjustment yielded I² = 0% for all indices; common adjustment suggested I² of 70.2% (TyG) and 84.5% (TyHGB) (95% CIs 0–99.2% and 38.7–99.6%; overlapping with the 0% under cohort-specific adjustment). NHHR and TCBI showed I² = 0% under both strategies. Prediction intervals for TyG (0.828–1.403) and TyHGB (0.854–1.361) included unity. ΔAUC after adding individual indices ranged from − 0.0043 to 0.0053; inverse-probability weighting did not support reading unweighted ICU estimates as upper bounds. Conclusions. Under per-SD scaling, cohort-specific adjustment yielded I² = 0% for all four indices, whereas common adjustment suggested I² of 70.2% (TyG) and 84.5% (TyHGB), though 95% CIs were wide and overlapping under k = 3. Reintroducing cohort-specific confounders (SOFA, APACHE IV) returned I² to 0%, consistent with residual confounding from omitting severity measures rather than a general adjustment-choice effect. Pooled hazard ratios combine different mortality horizons and do not constitute clinical transfer. Assessment of candidate indices should distinguish measurement calibration, conditional association, fixed-rule performance and incremental prediction for a defined population and horizon.

doi.org/10.21203/rs.3.rs-11308952/v1Open accessBiomarkers

AgingResearch.News