Abstract Number: Esoc2026a1130 Development and Validation of a Multi-Modal Based Machine Learning Model for 90-Day Prognosis Outcome in Acute Ischemic Stroke Patients

Abstract Background and aims Intensive blood pressure management isn't beneficial for all acute ischemia patients after endovascular thrombectomy. This study aimed to develop a clinically useful 90 - day prognosis prediction model and identify key baseline characteristics of patients benefiting from such management. Methods Five machine learning models were used to predict patients with favourable (0 - 2) or unfavorable (3 - 6) mRS based on baseline characteristics, CT image and blood routine test. GridSearchCV and 5-CV validation were applied in the training set to optimize hyperparameters. The model performance was evaluated by discrimination using the AUC and decision curve analysis(DCA). The Shapley additive explanations was used to explain the relative importance and effect direction of each predictor. Results A total of 816 patients from the ENCHANTED2/MT randomized controlled trial were included to develop a prediction model. The Xgboost model exhibited optimal performance in predicting the 90-day prognosis of AIS patients, with an AUC-ROC of 0.807(Figure 1-3) . The DCA demonstrated that combining key features provided greater benefit than treating all or none across a broad range of threshold probabilities . Key predictive factors identified included admission NIHSS score, age, core infarct volume, and systolic blood pressure level(Figure 4) . Furthermore, patients with an NIHSS score 15, age 77 years, core infarct volume 15, and post-treatment systolic blood pressure 163 mmHg were found to achieve a good prognosis with intensive blood pressure reduction(Figure 5). Conclusions The prediction model based on combined clinical features established by the Xgboost model can reliably predict the prognosis of AIS Patints. Conflict of interest Name of author: nothing to disclose Figure 1 - belongs to Results Figure 2 - belongs to Results Figure 3 - belongs to Results Figure 4 - belongs to Results Figure 5 - belongs to Results