Development and internal validation of an interpretable machine learning model incorporating LVEF and SII for Predicting complex coronary lesions in patients with type 2 diabetes mellitus and coronary heart disease

Background Noninvasive preoperative risk stratification for complex coronary lesions remains difficult for people living with type 2 diabetes (T2DM) and coronary heart disease (CHD). While left ventricular ejection fraction (LVEF) and the systemic immune-inflammation index (SII) have each been associated with such lesions, few studies have systematically developed and validated a machine learning (ML) model that integrates both markers in T2DM and CHD. Methods This retrospective analysis included 846 patients presenting with concurrent T2DM and CHD, divided into training (593) and validation (253) groups following a 7:3 split. Seven classifiers were developed using LASSO-selected predictors followed by multivariable logistic screening. A separately seeded internal sensitivity analysis evaluated paired discrimination, calibration, probability-scale SHAP, nested selection, and fixed LVEF reference breakpoints at 40% and 50%. All sensitivity analyses were conditional on the completed data matrix. Results Among the 846 patients, 475 (56.1%) were classified as having complex lesions. LASSO regression and multivariable logistic regression identified 7 key predictors: LVEF, log-SII, Triglyceride-Glucose index (TyG), high-sensitivity C-reactive protein (hs-CRP), log-transformed N-terminal pro-B-type natriuretic peptide (log-NT-proBNP), aspartate aminotransferase (AST), and hypertension. The initial gradient boosting machine (GBM) validation AUC was 0.743 (95% CI: 0.683–0.804). In the supplementary split, GBM achieved an AUC of 0.757 and a Brier score of 0.199, without significant AUC differences versus logistic regression ( P = 0.644) or AdaBoost ( P = 0.280). LVEF, hs-CRP, and log-SII had the largest mean absolute SHAP contributions. Fully adjusted LVEF remained inversely associated with complexity (OR per percentage point, 0.92; 95% CI: 0.89–0.95), whereas the log-SII association attenuated (OR, 1.35; 95% CI: 0.91–2.00). Conclusions A GBM-based model using routine clinical and laboratory data achieved moderate performance in detecting patients with T2DM and CHD at risk for complex coronary lesions. Superiority over simpler models and a clinically actionable LVEF threshold were not established. The single-center design and absence of external validation limit clinical application. Clinical trial number Not applicable.