Prognostic value of triglyceride-glucose glycemic variability index in critically ill unstable angina patients via machine learning with independent hospital validation
The triglyceride-glucose (TyG) index is associated with outcomes in critical illness, but the prognostic value of combining it with glycemic variability, the TyG-glycemic variability index (TyG-GVI), has not been evaluated in critically ill patients with unstable angina. In this study we examined the association between TyG-GVI and ICU and in-hospital mortality and developed an interpretable machine-learning model for mortality risk. We retrospectively analyzed 1,444 adult patients with unstable angina from the MIMIC-IV 3.1 database (2008–2022) as the development cohort, dichotomized at the median TyG‑GVI. Associations with ICU and in-hospital mortality were estimated using Cox proportional-hazards and logistic regression, and the dose-response relationship was modelled with restricted cubic splines. Random Forest, XGBoost and LightGBM classifiers were developed with Boruta confirmatory importance analysis and SHAP-based interpretation, and the optimal model was externally validated in an independent retrospective cohort of 340 ICU patients with unstable angina from the Quzhou Affiliated Hospital of Wenzhou Medical University. A total of 108 ICU deaths (7.5%) and 190 in-hospital deaths (13.2%) occurred. Compared with the low TyG-GVI group, high TyG-GVI was associated with ICU mortality in the fully adjusted Cox model (HR 2.29, 95% CI 1.43–3.66) and with in-hospital mortality (HR 1.54, 95% CI 1.11–2.15), although the latter did not reach statistical significance in every model specification. The dose-response relationship was non-linear (P for non-linearity < 0.01 for ICU mortality and < 0.05 for in-hospital mortality), with a data-derived nadir at a TyG-GVI of approximately 3.135. Random Forest achieved the numerically highest discrimination in the development cohort (AUC 0.830) and in the external validation cohort (AUC 0.874). High TyG-GVI was independently associated with ICU mortality and was also associated with in-hospital mortality in critically ill patients with unstable angina. An interpretable machine-learning model incorporating TyG-GVI showed consistent discrimination performance in the development and external validation cohorts and has been deployed as a web-based research prototype. Prospective validation and formal assessment of calibration and clinical utility are required before clinical use.
Authors
- Zhiguang Zheng
- Ke‐Jie He (ORCID: https://orcid.org/0000-0001-7406-6254)
- Yanqiu Jiang
- Junwei Qiu
- Shengshan Xu
- Yunru Qian
Institutions
- Nanjing University of Chinese Medicine (CN)
- Wenzhou Medical University (CN)
- Quzhou City People's Hospital (CN)
- Jiangmen Central Hospital (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1038/s41598-026-79521-y
- Primary Topic
- Hyperglycemia and glycemic control in critically ill and hospitalized patients
- Type
- article
- Field-Weighted Citation Impact
- 0.00