TyG Index and Machine Learning for the Discrimination of NSTEMI among Patients with Acute Heart Failure: A Retrospective Cross-Sectional Study

Background: Non-ST-segment elevation myocardial infarction (NSTEMI) can occur together with acute heart failure (AHF). Because their clinical presentations overlap, it is often difficult to identify NSTEMI in patients with AHF. The triglyceride-glucose (TyG) index has been linked to cardiovascular risk, but whether it helps identify concomitant NSTEMI in patients with AHF is still unclear. We examined the association between the TyG index and NSTEMI, its predictive value, and whether it adds incremental information to risk discrimination. Methods: We retrospectively reviewed 1,766 patients hospitalized for AHF. Among them, 684 were diagnosed with NSTEMI by coronary angiography. We used logistic regression and prespecified subgroup analyses to examine the association between the TyG index and NSTEMI. Boruta feature selection was used to identify relevant predictive features. We then developed six machine learning (ML) algorithms that incorporated the TyG index. Model performance was assessed by discriminative ability, calibration, clinical net benefit, and reclassification. SHapley Additive exPlanations (SHAP) analysis was used to quantify how much each feature contributed to model predictions. Results: A higher TyG index was associated with higher odds of concomitant NSTEMI (odds ratio [OR], 2.03, 95% CI, 1.67–2.46; P < 0.001). This association was consistent across subgroups defined by diabetes status, but it varied according to lipid-lowering medication use ( P for interaction = 0.028). Boruta feature selection identified N-terminal pro-B-type natriuretic peptide (NT-proBNP), cardiac troponin T (cTnT), the TyG index, triglycerides, creatinine, and fasting blood glucose as important features. Among the algorithms tested, the enhanced random forest model had the highest area under the receiver operating characteristic curve (AUROC) (0.8558 vs. 0.8488) and a slightly lower Brier score (0.1473 vs. 0.1488). At the selected classification threshold, sensitivity was 83.9% and specificity was 78.1%. Decision curve analysis showed similar clinical utility for the standard and enhanced random forest models. SHAP analysis showed that NT-proBNP and cTnT had the greatest influence on model predictions, and the TyG index also contributed to NSTEMI classification. Conclusions: In patients with AHF, a higher TyG index was independently associated with concomitant NSTEMI. The TyG index added incremental information to multivariable prediction models, although the improvement was modest and differed across algorithms. The TyG index may therefore be better used as a supplementary marker than as a standalone diagnostic tool.