A machine learning model using the TyG/HDL ratio predicted sepsis-induced cardiomyopathy with an AUC of 0.736, outperforming TyG alone and significantly stratifying 30-day survival (p=0.002).
Does a machine-learning model based on TyG/HDL improve the diagnosis and 28-day prognostication of sepsis-induced cardiomyopathy in adult ICU patients with sepsis compared to TyG alone?
A machine-learning model incorporating the TyG/HDL ratio provides superior discrimination for early identification and prognostication of sepsis-induced cardiomyopathy compared to TyG alone.
Absolute Event Rate: 0% vs 0%
Introduction: Sepsis-induced cardiomyopathy (SICM) worsens short-term outcomes, yet there is no metabolism-based tool for its early recognition. Triglyceride-glucose index (TyG) and its ratio to high-density-lipoprotein cholesterol (TyG/HDL) are emerging surrogates of insulin-resistance and lipotoxicity, but their comparative predictive values for SCM remain undefined. We therefore aimed to develop and externally validate machine-learning models centred on TyG and TyG/HDL for the diagnosis and 28-day prognostication of SICM. Methods: We retrospectively assembled a development cohort of 2588 adult ICU patients with sepsis from MIMIC-III, MIMIC-IV and eICU databases and an external validation cohort of 828 septic patients from Xiangya Hospital (Changsha, China). After missing data handling, Boruta feature selection (500 iterations) retained relevant predictors. Five classifiers were benchmarked, random forest (RF) performed best and was used to build two independent models (TyG-RF and TyG/HDL-RF) to avoid collinearity. TyG/HDL-RF received nested 5-fold cross-validation, calibration, decision-curve and SHAP analyses. External testing is under way. Restricted cubic splines (RCS) and Kaplan–Meier (KM) curves assessed non-linear mortality associations. Results: Boruta selected 21 predictors; creatinine, SOFA score, haemoglobin, age, RBC count, TyG/HDL, CKD and anion gap were the eight most influential variables (mean |SHAP| > 0.05) for TyG/HDL-RF. In cross-validation the TyG/HDL-RF model achieved the highest mean ± SD AUC (0.736 ± 0.016), outperforming TyG-RF (0.714 ± 0.018) and all non-RF classifiers. Test-set calibration was excellent (intercept = 0.03; slope = 0.94). RCS curves showed U-shaped relationships between both TyG and TyG/HDL and 28-day mortality, with nadirs at TyG ≈ 9.0 and TyG/HDL ≈ 0.32. High versus low TyG/HDL strata displayed significantly different 30-day survival (log-rank p = 0.002), a separation not observed for TyG alone (p = 0.065). Conclusions: The metabolism-centred TyG/HDL-RF model shows superior discrimination, calibration and net benefit for early SICM identification versus TyG alone or conventional scores, with SHAP enhancing interpretability. Ongoing external validation in a large Chinese cohort will determine generalisability and support deployment as a free web-based calculator.
Dai et al. (Sun,) reported a other. A machine learning model using the TyG/HDL ratio predicted sepsis-induced cardiomyopathy with an AUC of 0.736, outperforming TyG alone and significantly stratifying 30-day survival (p=0.002).