TyG-GVI marks a sharper mortality signal in critically ill unstable angina
Critically ill patients with unstable angina who carried a higher triglyceride-glucose glycemic variability index, or TyG-GVI, faced a materially worse short-term prognosis in a retrospective study that paired conventional regression with machine learning. The signal was strongest in the intensive care unit, where higher TyG-GVI tracked with mortality even after extensive adjustment. In a population already marked by unstable coronary physiology, the finding points to a metabolic burden that extends beyond ischemic symptoms alone.
The cohort came from two hospitals and two very different data environments. Investigators built the primary model in 1,444 adults from the MIMIC-IV 3.1 database spanning 2008 to 2022, then tested the most promising approach in an independent retrospective cohort of 340 ICU patients from Quzhou Affiliated Hospital of Wenzhou Medical University. Across the development cohort, 108 ICU deaths occurred, representing 7.5%, and 190 in-hospital deaths occurred, representing 13.2%. Those event counts gave the analysis enough separation to explore whether a combined measure of dysglycemia and lipid-metabolic instability could identify the sickest patients early in the ICU course.
Metabolic instability tracked with early death rather than broad inpatient risk
The core exposure blended two biologically relevant dimensions: the triglyceride-glucose index, which is often used as a surrogate for insulin resistance, and glycemic variability, which captures short-term fluctuations rather than a single static glucose value. In critically ill coronary patients, that combination may be more informative than either marker alone because unstable myocardium is vulnerable to both fuel dysregulation and inflammatory stress. A patient with repeated glycemic swings may be experiencing a much harsher endocrine environment than another patient with the same average glucose but more stable control.
Compared with the low TyG-GVI group, the high TyG-GVI group had a higher risk of ICU death in the fully adjusted Cox model, with a hazard ratio of 2.29 and a 95% confidence interval from 1.43 to 3.66. The association with in-hospital mortality was weaker, with a hazard ratio of 1.54 and a 95% confidence interval from 1.11 to 2.15, although the latter did not remain statistically significant in every model specification. ICU mortality appears to be the cleaner endpoint here, likely because metabolic instability may exert its earliest effect during the most physiologically fragile phase of care.
The dose-response relationship was not linear. Restricted cubic spline modeling showed significant non-linearity for both ICU mortality and in-hospital mortality, and the data-derived nadir sat at a TyG-GVI of approximately 3.135. Below and above that point, risk did not move in a flat, proportional way; instead, the curve bent, suggesting a threshold-like relationship rather than a simple monotonic gradient. For bedside clinicians, this kind of pattern matters because it argues against treating the marker as a passive label and instead positions it as a potential reflection of unstable metabolic reserve.
Machine learning favored Random Forest and kept its edge in external validation
To translate the risk signal into a prediction tool, the investigators trained three classifiers: Random Forest, XGBoost, and LightGBM. Boruta-based confirmatory importance analysis and SHAP interpretation were used to keep the model interpretable rather than turning it into an opaque black box. Random Forest emerged with the numerically highest discrimination in the development cohort, where the area under the receiver operating characteristic curve reached 0.830.
The same algorithm held up in the independent external validation cohort, where the AUC rose to 0.874. Those values suggest the model was not merely fitting noise in the MIMIC-IV dataset. Consistent discrimination across two cohorts from different institutions strengthens the argument that TyG-GVI carries reproducible prognostic information in critically ill unstable angina, even though discrimination alone does not prove clinical utility.
SHAP-based interpretation also kept the model clinically legible. In practice, explainability matters because ICU risk tools must do more than generate a score; they must show which variables are driving the estimate and whether those drivers are biologically plausible. A combined marker of glucose volatility and triglyceride burden fits that requirement better than many machine learning outputs because it maps onto insulin resistance, stress hyperglycemia, and the inflammatory milieu that can destabilize coronary disease.
Why the combined index may fit unstable angina biology
Unstable angina is not just a symptom label. It reflects a coronary state in which plaque instability, endothelial dysfunction, and myocardial oxygen supply-demand mismatch can deteriorate rapidly. Insulin resistance and glycemic variability can amplify that process by worsening endothelial function, promoting oxidative stress, and feeding inflammatory pathways that may leave coronary lesions more vulnerable. Elevated triglycerides, when coupled with glucose instability, may identify a metabolic phenotype that tolerates critical illness poorly.
This biological framing helps explain why a composite index outperformed a single static glucose or lipid measure. A hospitalized patient can have one reassuring laboratory value and still experience wide metabolic swings over time. The combined TyG-GVI metric captures a moving target, which is often what critical care prognosis depends on.
The study design, however, remained retrospective. Both the development and validation cohorts were observational, and the external cohort was still drawn from a single hospital system. Formal calibration and clinical utility were not fully established, and the investigators themselves called for prospective validation before any clinical use. A predictive model can be discriminative and still misestimate absolute risk at the bedside.
What the current evidence supports, and what it does not
The study supports two concrete conclusions. First, high TyG-GVI was independently associated with ICU mortality in critically ill unstable angina. Second, an interpretable machine-learning model incorporating the index performed consistently in an external validation cohort. Those are meaningful signals for hypothesis generation and for future risk-stratification research.
The work does not justify immediate adoption as a routine clinical decision tool. Calibration was not formally assessed in the summary, and no decision-curve result was provided to show whether the model would improve management decisions over standard ICU assessment. Prospective validation will be needed to determine whether TyG-GVI improves real-world triage, monitoring intensity, or therapeutic targeting in unstable angina patients who are already critically ill.
For now, the study places a combined metabolic marker alongside the familiar coronary risk vocabulary of ischemia, instability, and physiologic stress. In the ICU, those dimensions rarely move independently, and the mortality signal attached to TyG-GVI suggests that glycemic volatility and triglyceride burden may be more than laboratory background noise.