Interpretable machine learning of non-traditional lipid indices for diagnostic classification of CHD in patients with comorbid MASLD and T2DM: a multicenter study

Background: Patients with comorbid metabolic dysfunction-associated steatotic liver disease (MASLD) and type 2 diabetes mellitus (T2DM) have a significantly heightened risk for coronary heart disease (CHD). Conventional lipid profiles often underestimate residual cardiovascular risk. This study identifies valuable non-traditional lipid indicators and develops an interpretable machine learning framework for CHD identification in this population. Methods: This multicenter retrospective study analyzed 1,823 patients with MASLD and T2DM. Following 1:1 propensity score matching, 630 participants were used for association analysis, whereas the complete unmatched Cohort I (n = 1,665) was used for machine learning model development, with an independent cohort of 158 patients for external validation. Logistic regression and restricted cubic spline (RCS) models evaluated associations between CHD risk and eight non-traditional lipid indices. Six machine learning algorithms were compared using cascaded feature selection, with model transparency provided by Shapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME). Results: < 0.001). The final Stacking ensemble model incorporated 10 variables. This model showed competitive and relatively balanced performance, with AUCs of 0.750 and 0.756 and Brier scores of 0.156 and 0.191 in the internal test set and the independent external validation cohort, respectively. SHAP and LIME analyses further indicated that CRI-II, eGFR, age, and LCI were the major drivers of model classification. Conclusion: Non-traditional lipid indices, particularly CRI-II, showed strong associations with CHD in patients with concomitant MASLD and T2DM. Integrating these indicators into an interpretable machine learning framework provides a robust and transparent tool for CHD identification, potentially facilitating early clinical evaluation and decision-making.