Answers are grounded in this paper and cite their sources.
Metabolic dysregulation and glycemic instability are common during critical illness and may influence postoperative outcomes in patients with colorectal cancer. We investigated the prognostic value of the triglyceride-glucose glycemic variability index (TyG-GVI) and developed explainable machine learning models for ICU mortality prediction. This retrospective cohort study included 328 postoperative patients with colorectal cancer at Quzhou People’s Hospital who remained alive and under ICU observation at the prespecified 24-hour landmark after ICU admission. Patients who died or were discharged from the ICU before the landmark were excluded, and subsequent ICU and in-hospital mortality were evaluated using Cox and logistic regression, restricted cubic splines, Kaplan-Meier analysis, and subgroup analyses. Random Forest, XGBoost, and LightGBM models were developed using routinely available clinical variables and interpreted using SHAP. Among the 328 included patients, 27 (8.2%) died in the ICU and 46 (14.0%) died during hospitalization. Higher TyG-GVI was consistently associated with increased ICU and in-hospital mortality across multiple regression models. In fully adjusted Cox models, the hazard ratios were 4.40 (95% CI 1.60–12.1) for ICU mortality and 2.11 (95% CI 1.12–3.95) for in-hospital mortality. Restricted cubic spline analyses showed significant overall associations for both outcomes ( P < 0.05), without significant nonlinearity (P for nonlinearity = 0.400 and 1.000, respectively). Kaplan-Meier analysis showed significantly poorer survival in the higher TyG-GVI group for both outcomes ( P < 0.001). In the exploratory internal test set, which contained 8 ICU deaths, Random Forest had the highest observed AUC (0.827; 95% CI, 0.644–0.934), followed by LightGBM (0.806; 95% CI, 0.619–0.921) and XGBoost (0.762; 95% CI, 0.567–0.898). SHAP analysis identified TyG-GVI as the highest-ranked feature by mean absolute SHAP value within the fitted Random Forest model. Among postoperative patients with colorectal cancer who remained alive and under ICU observation at 24 h after ICU admission, higher TyG-GVI was associated with subsequent ICU and in-hospital mortality. In exploratory machine-learning analyses, TyG-GVI showed the highest relative feature importance within the fitted Random Forest model; however, its incremental predictive value beyond conventional clinical variables remains undetermined. Given the limited number of events and the absence of external validation, these findings should be considered hypothesis-generating and require confirmation in larger multicenter cohorts.
No takes yet. Share an insight, caveat, or question.
Lei et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: