Prediction model for early neurological deterioration in large artery atherosclerotic stroke

Background To develop and validate a predictive model for early neurological deterioration (END) in acute ischemic stroke due to large artery atherosclerosis. Methods We included 433 patients admitted to our hospital between August 2023 and January 2025, randomly split into training (325) and internal validation (108) cohorts (3:1). END was defined as an NIHSS increase of ≥2 points within 7 days. Univariate analysis and LASSO regression selected variables in the training cohort; multivariate logistic regression was used to build the predictive model, which was visualized as a nomogram. Model performance was assessed using ROC curves, calibration curves, and decision curve analysis (DCA). Results END occurred in 27.1% of the training cohort and 26.9% of the internal validation cohort. Seven variables were identified: neutrophil count, platelet count, lymphocyte count, fasting plasma glucose, total cholesterol, homocysteine, and D-dimer. Six were independent predictors ( p < 0.05); fasting plasma glucose showed a trend toward association ( p = 0.084) and was retained based on LASSO selection. The training cohort AUC was 0.791 (95% CI: 0.730–0.851), with 72.7% sensitivity, 78.5% specificity, and 76.9% accuracy; the internal validation cohort AUC was 0.778 (95% CI: 0.675–0.881), with 75.9% sensitivity, 70.9% specificity, and 72.2% accuracy. Calibration curves showed acceptable fit (mean absolute errors 0.023 and 0.048). Hosmer-Lemeshow tests were all p > 0.05. DCA suggested that the model may offer some net benefit across a relatively wide range of threshold probabilities. Conclusion The predictive model developed in this study (incorporating neutrophil count, platelet count, lymphocyte count, fasting plasma glucose, total cholesterol, homocysteine, and D-dimer) may serve as a straightforward initial screening tool for early neurological deterioration in patients with acute ischemic stroke due to large artery atherosclerosis. Six of these variables were independent predictors; fasting plasma glucose showed a trend toward association ( p = 0.084) but was retained in the model based on LASSO selection. The model may facilitate early identification of high-risk patients and provide some reference for individualized treatment decisions.