Integrating lipometabolic and adiposity indices to enhance risk stratification for metabolic dysfunction-associated steatotic liver disease in type 2 diabetes: clinical utility and interplay between triglyceride-glucose body mass index and low-densit

BACKGROUND: Metabolic dysfunction-associated steatotic liver disease (MASLD) significantly exacerbates the prognosis of patients with type 2 diabetes mellitus (T2DM). We aimed to compare metabolic and adiposity-related surrogates for MASLD using data-driven feature selection and to validate a parsimonious risk model across distinct populations. METHODS: We included 449 T2DM patients from the WMU cohort (discovery) and 306 from the Japanese NAGALA cohort (external validation). A two-stage data-driven feature-selection framework (Boruta and LASSO) was implemented to identify a parsimonious two-variable signature (TyG-BMI and SGLT2i). Based on these results, TyG-BMI was prioritized for systematic evaluation. Association and dose-response relationships were assessed via multivariate logistic regression and restricted cubic splines. Multiplicative and additive interactions between TyG-BMI and LDL-C were further explored. Clinical utility was evaluated via AUC, NRI, IDI, and decision curve analysis. RESULTS: The Boruta algorithm ranked TyG-BMI as the feature with the highest importance score for MASLD classification. Subsequently, LASSO regression (utilizing the 1-standard-error criterion λ1se) identified a parsimonious two-variable signature comprising TyG-BMI and SGLT2i. In the WMU cohort, TyG-BMI exhibited a potent association with MASLD (T3 vs. T1: OR = 7.36, 95% CI 3.89-13.94) and a significant linear dose-response relationship (P for overall < 0.001). Incorporation of TyG-BMI into the baseline model improved discriminative performance (AUC increased from 0.7288 to 0.7920) and was associated with improved continuous reclassification (NRI: 0.6088, P < 0.001). DCA and calibration plots confirmed high clinical net benefit and accuracy. Furthermore, a significant synergistic interaction was observed between TyG-BMI and low-density lipoprotein cholesterol (LDL-C). CONCLUSIONS: TyG-BMI, selected through data-driven feature selection, may serve as a practical candidate predictor of MASLD in patients with T2DM. The observed interaction between TyG-BMI and LDL-C suggests that their joint assessment may further refine MASLD risk stratification. The derived parsimonious model offers a high-performing, non-invasive tool for early MASLD risk stratification across Asian populations.