Temporal shifts in prognostic factors for 90- and 180-day outcomes after stroke thrombolysis: A machine learning analysis.

INTRODUCTION: Prognostication at 90 and 180 days after thrombolysis for acute ischemic stroke (AIS) is critical, yet the temporal evolution of key predictors remains inadequately understood. The utility of machine learning for systematically comparing prognostic factors across these distinct time points remains to be fully established. METHOD: We retrospectively analyzed consecutive AIS patients undergoing intravenous thrombolysis from October 2020 to December 2024. Features were selected from pre-therapy baseline data via univariable analysis and LASSO regression to develop five machine learning models. Model performance and clinical utility were assessed by AUC-ROC and Decision Curve Analysis (DCA), respectively. The primary endpoint was a modified Rankin Scale > 2 at 90 and 180 days. RESULTS: A total of 432 patients were included. At 90 days, 81 patients (18.8%) had an unfavorable outcome. By the 180-day follow-up, 86 patients (19.9%) were lost to follow-up. Among the remaining 346 patients, 48 (13.9%) had an unfavorable outcome. On the holdout test set, Logistic Regression (AUC = 0.757) and Random Forest (AUC = 0.833) were the optimal models for 90- and 180-day outcomes, respectively. While baseline NIHSS score and age were dominant predictors for both endpoints, a notable temporal shift in biomarker significance emerged: admission fibrinogen was a key predictor at 90 days, but was supplanted by white blood cell count for the 180-day prognosis. CONCLUSION: Our study reveals a crucial temporal evolution in prognostic biomarkers after thrombolysis, shifting from fibrinogen at 90 days to white blood cell count at 180 days. This dynamic landscape of predictors, identified through machine learning, underscores the necessity of developing time-specific models to accurately forecast patient recovery.