Composite indices of insulin resistance and sensitivity versus matched component-variable models for incident diabetes prediction

Background The triglyceride–glucose (TyG) index, single-point insulin sensitivity estimator (SPISE), and metabolic score for insulin resistance (METS-IR) are widely used non-insulin-based surrogate indices. Whether their fixed formulas predict diabetes better than models retaining the same component measurements remains uncertain. Methods We analyzed 112,070 participants from the RICH cohort and externally validated frozen RICH models in 15,453 participants from the NAGALA cohort. Cox models compared each index with a matched model containing the same component variables separately; all models included age, sex, systolic blood pressure, and alanine aminotransferase. Internal validation used 1,000 bootstrap resamples and leave-one-center-out validation. External paired AUC differences were estimated using 1,000 participant-level bootstrap resamples. Additional analyses evaluated the contribution of FPG and repeated the index–component comparisons with FPG retained separately in both models. Likelihood-ratio tests evaluated formula constraints. Results Higher TyG and METS-IR and lower SPISE were nonlinearly associated with incident diabetes. In RICH, component-variable models had higher optimism-corrected 3-year AUCs than the TyG, SPISE, and METS-IR models by 0.051 (95% CI, 0.037–0.065), 0.009 (0.004–0.016), and 0.059 (0.047–0.073). In NAGALA, these differences were 0.072 (0.040–0.106), −0.006 (−0.021 to 0.010), and 0.046 (0.014–0.079), respectively. These models underestimated 3-year diabetes risk (observed-to-expected ratios, 1.25–2.52). Entering FPG separately in both models narrowed differences. In NAGALA, no clear difference was observed for TyG or SPISE, whereas the METS-IR model had a higher AUC than its matched component-variable model by 0.013 (0.004–0.022). Conventional likelihood-ratio tests favored freely estimated formula-scale coefficients for all three indices (all Holm-adjusted P < 0.001), although SPISE inference was sensitive to center clustering. Freely estimating these coefficients improved AUC in both cohorts only for TyG. The comparison of TyG with FPG plus triglycerides showed consistent results in the expanded sample. Component-variable advantages for TyG and METS-IR persisted with splines and at 2.5 years. Conclusion Component-variable models showed higher discrimination than the matched TyG and METS-IR models, whereas the small difference for SPISE was not reproduced externally. These differences were reduced when FPG was retained separately. Freely modeling log-transformed FPG and triglycerides improved discrimination over TyG in both cohorts.