Exploring the prognostic impact of triglyceride-glucose index in critically ill patients with first-ever stroke: insights from traditional methods and machine learning-based mortality prediction.

BACKGROUND

The incidence and mortality of first-ever strokes have risen sharply, especially in the intensive care unit (ICU). Emerging surrogate for insulin resistance, triglyceride-glucose index (TyG), has been linked to stroke prognosis. We aims to explore the relationships between TyG with ICU all-cause mortality and other prognosis, and to develop machine learning (ML) models in predicting ICU all-cause mortality in the first-ever strokes.

METHODS

We included first-ever stroke patients from the eICU Collaborative Research Database in 2014-2015 as the primary analysis cohort (then divided into training and internal validation cohorts) and from local hospital's ICUs as the external validation cohort. Multivariate Cox proportional hazards models and restricted cubic spline analyses were used to evaluate the association between TyG and ICU/hospital all-cause mortality. Linear regression and correlation analyses were performed to examine the relationships between TyG with length of ICU/hospital stay and Glasgow Coma Score.

RESULTS

The primary analysis cohort included 3173 first-ever strokes (median age 68.0 [55.0-68.0] years; 63.0% male), while the external validation cohort included 201 first-ever strokes (median age 71.0 [63.0-77.0] years; 62.3% male). Multivariate Cox proportional hazards models revealed that the high TyG group (TyG ≥ 9.265) was associated with higher ICU (HR 1.92, 95% CI 1.38-2.66) and hospital (HR 1.69, 95% CI 1.32-2.16) all-cause mortality, compared with low TyG group (TyG < 9.265). TyG was also correlated with ICU length of stay (r = 0.077), hospital length of stay (r = 0.042), and Glasgow Coma Score (r = -0.132). TyG and other six features were used to construct ML models. The random forest model performed best in internal validation with AUC (0.900) and G-mean (0.443), and in external validation with AUC (0.776) and G-mean (0.399).

CONCLUSION

TyG (optimal cut-off: 9.265) was identified as an independent risk factor for ICU and hospital all-cause mortality in first-ever strokes. The ML model incorporating TyG demonstrated strong predictive performance. This emphasises the importance of insulin resistance (with TyG as a surrogate measure) in the prognostic assessment and early risk stratification of first-time stroke patients.