Abstract WP020: Integrating Peripheral Inflammatory Markers to Predict Early Neurological Deterioration After Intravenous Thrombolysis: A Multicenter Cohort Study

Introduction: Intravenous thrombolysis is a standard therapy for acute ischemic stroke (AIS). Nevertheless, a subset of patients experience early neurological deterioration (END) after treatment, adversely affecting outcomes. Inflammation is believed to play a key role in END, yet systematic evidence on the predictive value of peripheral inflammatory markers remains limited. This study evaluated the ability of 14 routinely available hematologic markers—encompassing absolute cell counts and composite ratios—to predict END after thrombolysis, and examined their performance within a multivariable ensemble model. Methods: We analyzed a prospective, multicenter stroke registry that enrolled AIS patients treated with intravenous alteplase within 4.5 hours of onset across 18 stroke centers in Henan Province, China (January 2023–December 2023). END was defined as an increase of ≥4 points in NIHSS within 24 hours post-thrombolysis. Extracted inflammatory markers included white blood cells, neutrophils, lymphocytes, monocytes, eosinophils, basophils, and composite indices: NLR, LMR, PLR, SII, SIRI, AISI, MHR, and NAR. For each marker, receiver operating characteristic (ROC) curves and areas under the curve (AUCs) were computed. An XGBoost model integrating all markers was developed to assess combined predictive performance and calibration. Results: A total of 5412 eligible AIS patients were included; 709 (13.1%) developed END within 24 hours. Among single predictors, monocyte count (AUC = 0.702), MHR (AUC = 0.698), and AISI (AUC = 0.639) showed the best discrimination, whereas most other markers yielded AUCs within 0.55–0.71, indicating limited standalone performance. In contrast, the multivariable XGBoost model achieved strong discrimination with a cross-validated mean AUC of 0.891. Calibration curves demonstrated close agreement between predicted probabilities and observed risks, indicating good model fit and clinical applicability. Conclusions: Although individual inflammatory markers display modest predictive ability for END after thrombolysis, integrating multiple markers within a machine-learning framework markedly enhances risk prediction. These findings support incorporating peripheral inflammatory information into acute stroke care to enable earlier identification and risk stratification of END, thereby informing personalized interventions.