TyG-CVAI for identifying hyperuricemia: geographical external validation, model updating, and explainable machine learning

Background TyG-CVAI can be calculated from routine metabolic and anthropometric measurements, but evidence for same-visit identification of prevalent hyperuricemia across regions, its incremental value beyond conventional variables, and its advantage over the constituent variables themselves remains limited. Objective To assess the geographical transportability of the triglyceride-glucose Chinese visceral adiposity index (TyG-CVAI) for identifying prevalent hyperuricemia and to determine whether model updating or more complex machine-learning methods materially improve externally validated performance. Methods We analyzed prospectively collected, single-visit health-examination and questionnaire data from 2,119 adults and performed a cross-sectional analysis. The Shandong center was used for model development (n=1,288) and the Jiangsu center for geographical external validation (n=831). Associations were estimated with multivariable logistic regression and robust Poisson regression, with restricted cubic splines for dose-response assessment. Single-index discrimination, transferred thresholds, calibration, decision-curve analysis, five prespecified logistic model formulations, recalibration, reverse validation, and dataset shift were evaluated. Machine-learning models were tuned by nested resampling in Shandong and evaluated once in Jiangsu; serum urate and all outcome-derived variables were excluded from every prediction step. Results Hyperuricemia was present in 433 participants (20.43%). A 1-SD increase in TyG-CVAI was associated with higher odds of prevalent hyperuricemia in Shandong (OR, 1.74; 95% CI, 1.39-2.18), Jiangsu (OR, 1.97; 95% CI, 1.50-2.59), and the pooled cohort (OR, 1.84; 95% CI, 1.55-2.19), with no center interaction (P = 0.984). The pooled prevalence ratio per 100-unit increase was 1.12 (95% CI, 1.09-1.16). TyG-CVAI AUCs were 0.74 in Shandong and 0.72 in Jiangsu. Applying the rounded Shandong cutoff of 593 in Jiangsu yielded sensitivity 0.75 and specificity 0.52. The external AUC increased from 0.71 for the noncomponent clinical model to 0.74 after adding TyG-CVAI, but a components-only model achieved 0.74 and the clinical-plus-components model 0.76. Penalized logistic regression had the highest external machine-learning AUC point estimate (0.74), but no tree model differed significantly after multiplicity adjustment. The prespecified sex interaction was not significant after Holm correction. Conclusions TyG-CVAI was consistently associated with prevalent hyperuricemia across two centers and showed moderate discrimination. It provided a small increment over a clinical model that deliberately excluded its components, but no clear advantage over models using the constituent variables directly. TyG-CVAI may help prioritize direct serum-urate testing when its components are already available, but it should not be used as a stand-alone diagnostic or prognostic test.