Abstract
This study aims to predict hemorrhagic stroke outcomes, including 90-day prognosis and in-hospital mortality, using machine learning models and SHapley Additive exPlanations (SHAP) analysis. Data were collected from a national Stroke Registry from January 2014 to July 2022. Various predictive factors were considered, such as stroke severity at presentation, patient demographics, laboratory results, admission location, and other clinical features. Random forest, logistic regression, XGboost, support vector machines, and decision trees were trained and evaluated. SHAP analysis was conducted to identify key predictors. The RF model demonstrated superior performance in predicting prognosis, while LR was more effective in predicting in-hospital mortality. The National Institute of Health Stroke Score (NIHSS) and admission location were key predictors. Despite its limitations, this research underscores the importance of advancing stroke registries and emphasizes the necessity for comprehensive external validation of predictive models. Furthermore, it demonstrates the importance of initial stroke severity in influencing patient outcomes and highlights the significance of admission to stroke units in reducing poor outcomes. This may help shape interventions to enhance stroke center capacities and influence strategic policies. This study contributes towards developing more precise predictive models for hemorrhagic stroke outcomes, potentially impacting clinical practice and optimizing resource allocation significantly.