Jianmo Liu, Canwei Yi, Haowen Luo, Pengfei Yu, Zhaofa Liu, Hongfei Zhao, Huiming Liao, Xiao Zhang, Weijiang Han, Chumin Zhou, Yingping Yi
Nutritional, electrolyte, inflammatory, and insulin resistance indices are significantly associated with PSI, and their integration with clinical features produces a predictive model with good discriminative performance.
BACKGROUND: Post-stroke infection (PSI) is a major cause of poor prognosis in acute ischemic stroke (AIS). The independent and combined predictive value of nutritional, electrolyte, inflammatory, and insulin resistance indices for PSI requires further validation.
METHODS: We retrospectively analyzed 6,149 AIS patients from the Second Affiliated Hospital of Nanchang University (June 2017-April 2025). Multivariate logistic regression, restricted cubic spline (RCS), and ROC curves assessed the predictive value of nutritional (GNRI, PNI), electrolyte (serum sodium, potassium), inflammatory (NLR, PLR), and insulin resistance (TyG, TyG-BMI) indices. Hierarchical clustering selected clinical features; five ML algorithms (XGBoost, RF, LightGBM, Extra Trees, CatBoost) with 5-fold cross-validation built predictive models. SHapley Additive exPlanations (SHAP) identified key predictors.
RESULTS: NLR, PLR and TyG were elevated in PSI patients, while GNRI, PNI, and serum sodium were decreased (all p < 0.05). RCS identified malnutrition (GNRI ≤98.799, PNI ≤ 46.898), a U-shaped sodium-PSI relationship (inflection: 139.611 mmol/L), elevated NLR (>2.703), and elevated TyG as risk factors; After adjusting for confounding factors such as comorbidities, the statistical significance of potassium metabolic disturbances disappeared (p = 0.151), suggesting that their impact on PSI may be due to confounding by underlying diseases rather than an independent causal association. The non-linear potassium-PSI association (inflection: 3.529 mmol/L) was not significant in the fully adjusted model. TyG-BMI was not independently associated with PSI. The CatBoost model integrating all four biomarker categories with clinical features performed best (AUC = 0.750, 95% CI: 0.722-0.779). SHAP identified TyG (|SHAP| = 0.409), GNRI (0.358), NLR (0.293), invasive procedure (0.194), coronary heart disease (0.179), and atrial fibrillation (0.137) as top features.
CONCLUSION: Nutritional, electrolyte, inflammatory, and insulin resistance indices are significantly associated with PSI, and their integration with clinical features produces a predictive model with good discriminative performance.