Machine Learning-Based Early Risk Stratification for Sepsis-Related Troponin-Defined Myocardial Injury Using Routine Metabolic Indicators.

BACKGROUND: Early recognition of sepsis-related myocardial injury during sepsis remains difficult, partly because harmonized echocardiographic phenotyping is often unavailable in retrospective critical-care datasets. Routine metabolic indicators, including the triglyceride-glucose (TyG) index and the triglyceride-to-HDL cholesterol (TG/HDL) ratio, may support early risk stratification, but their value within a multicenter machine-learning framework has not been well defined. METHODS: In this retrospective multicenter study, we included 2,588 adult patients with sepsis from the MIMIC-IV and eICU Collaborative Research Databases as the development cohort and 504 patients from the Xiangya ICU cohort as the external validation cohort. The primary endpoint was an operational sepsis-related troponin-defined myocardial injury phenotype rather than echocardiography-confirmed septic cardiomyopathy. The original primary models used 14 Boruta-selected clinical predictors, with TyG and TG/HDL added according to the model setting. Five machine-learning classifiers were compared using a stratified 70/30 split with five-fold cross-validation. SHAP was used to explain model behavior and feature contribution. RESULTS: RF was retained as the primary interpretable model because it provided stable performance and straightforward model explanation, although boosting models achieved numerically higher discrimination in some settings. The combined TyG plus TG/HDL RF model achieved an AUC of 0.725 in the internal test set. In the Xiangya ICU cohort, the external validation AUC was 0.619 (95% CI 0.568-0.669) and the Brier score was 0.256. External calibration showed underprediction, and decision-curve findings did not support immediate clinical use. Joint stratification identified a subgroup with concomitant elevation of both metabolic indices and a higher risk of the operational troponin-defined phenotype (OR 1.59, 95% CI 1.22-2.07). CONCLUSIONS: An interpretable random forest model incorporating routine admission laboratory data, including metabolic indices, may assist exploratory early risk stratification for sepsis-related troponin-defined myocardial injury. The model should not be interpreted as a diagnostic tool for echocardiography-confirmed septic cardiomyopathy, and prospective validation is required before clinical implementation.