Head-to-head comparison and cross-population replication of 30 lipid-related composite indices for cardiovascular risk: a study in CHARLS and NHANES

Background Dozens of composite indices built from lipids, glucose and anthropometry are proposed each year as cardiovascular risk markers, almost always in single cohorts. How many replicate across populations and endpoint types is unknown. We compared 30 lipid-related composite indices head-to-head in two national cohorts on two continents. Methods The discovery cohort comprised 8,226 Chinese adults aged ≥ 45 years from the CHARLS 2011 blood subsample, free of cardiovascular disease (CVD) at baseline, with incident physician-diagnosed CVD ascertained through 2018. The replication cohort comprised 13,653 US adults aged ≥ 45 years from the NHANES 1999–2018 fasting subsample with mortality linkage; the outcome was CVD mortality. Thirty indices were compared using z-standardised Cox models with a pre-specified component-aware covariate set and false discovery rate (FDR) control. Replication required directional consistency and FDR significance in both cohorts. Between-cohort heterogeneity, dose-response, incremental value, calibration, decision-curve analysis, subgroup interactions and a same-outcome analysis of all-cause mortality were prespecified. Results Only 6 of 30 indices replicated across both cohorts: estimated glucose disposal rate (eGDR; protective; CHARLS HR per SD 0.758 [95% CI 0.712–0.806]; NHANES 0.786 [0.715–0.864]) and five adiposity-linked indices, including TyG-WC (1.121 and 1.165), with no detectable between-cohort heterogeneity for five of the six (I 2 = 0%). The TyG index itself, the atherogenic index of plasma and most inflammation-lipid ratios did not replicate. eGDR showed monotone dose-response in both cohorts (quartile 4 vs 1: 0.49 and 0.63; P-trend < 0.001) and modestly improved reclassification of 10-year CVD mortality (NRI 0.297, 95% CI 0.220–0.375), with well-calibrated predictions but negligible decision-curve net benefit. Associations were strongest before age 60 in NHANES (P-interaction ≤ 0.003). For all-cause mortality, only eGDR replicated, indicating that replication is strongest for incident disease. Conclusions Only one in five widely used lipid-related composite indices replicated across a Chinese incidence cohort and a US mortality cohort. eGDR and central-adiposity-coupled glycolipid indices were the consistent performers; single-cohort claims for other indices warrant caution. Because the cohorts differ in endpoint type, the findings are best read as cross-population, cross-outcome triangulation rather than formal external validation.