Association of the Cholesterol–High-Density Lipoprotein–Glucose Index with Large-Artery Atherosclerotic Versus Cardioembolic Stroke: A Single-Center Retrospective Study

Background/Objectives: We aimed to assess the association of the cholesterol–high-density lipoprotein–glucose index (CHG) with large-artery atherosclerotic stroke (LAA) versus cardioembolic stroke (CE), and to compare its discrimination and incremental value beyond a clinical model with those of the triglyceride–glucose index (TyG). Methods: This single-center retrospective study consecutively screened patients hospitalized with ischemic stroke at Beijing Anzhen Hospital, Capital Medical University, from September 2024 to August 2025 and included eligible patients with acute LAA or CE. CHG and TyG were calculated from the first examination results after admission. Missing BMI values were handled using 20 multiple imputations. Logistic regression, restricted cubic splines, and exploratory subgroup analyses were used to evaluate associations, with further adjustment for lipid-lowering and glucose-lowering medication use before the current stroke onset. Firth-penalized regression was also used to assess the robustness of the association estimates. Model performance was evaluated using 1000 paired outcome-stratified bootstrap resamples, comparing CHG, TyG, and joint modeling of the three component indicators and estimating differences in corrected AUC and their 95% confidence intervals. Results: The study included 921 patients, comprising 790 with LAA and 131 with CE. After multivariable adjustment, each 1-standard-deviation increase in CHG was associated with an odds ratio of 2.83 for LAA versus CE (95% confidence interval: 2.20–3.65, p < 0.001), with a nonlinear association (p for nonlinearity = 0.003). The OR in the Firth sensitivity analysis was 2.77 (95% confidence interval: 2.17–3.59). The single-index AUCs of CHG and TyG were 0.725 and 0.662, respectively (paired DeLong p < 0.001). Adding CHG increased the bootstrap-corrected AUC of the clinical model from 0.647 to 0.743 and reduced the corrected Brier score from 0.119 to 0.105. The difference in corrected AUC between the CHG-enhanced model and the TyG-enhanced model was 0.051 (95% confidence interval: 0.025–0.079), whereas its difference from the model jointly including the three component indicators remained uncertain. The direction of the main associations was consistent in sensitivity analyses, and no clear subgroup interactions were identified. Conclusions: In this cohort, higher CHG was positively associated with LAA versus CE after adjustment for measured covariates. Its enhanced model had higher discrimination than the TyG-enhanced model in internal validation, but its advantage over joint modeling of the component indicators remained uncertain. This was a single-center retrospective study with a substantial imbalance in group sizes, and residual confounding may remain. All performance estimates were derived from internal validation. The clinical utility of CHG and its applicability across centers have not been established and require independent external validation.