Predicting hemorrhagic transformation in acute ischemic stroke using a machine learning model based on clinical and HR-MRI features

Abstract

Early prediction of hemorrhagic transformation (HT) in acute ischemic stroke (AIS) is critical for clinical decision-making. We aimed to develop and externally validate a highly interpretable machine learning model integrating high-resolution magnetic resonance imaging (HR-MRI) and clinical features to predict HT in patients with AIS. This multicenter retrospective study enrolled a primary cohort of 1400 patients with AIS from First Central Hospital of Baoding, randomly allocated into training and testing cohorts at a 7:3 ratio. An independent cohort of 500 patients from Nanjing Gaochun People's Hospital served as external validation. We extracted demographic, laboratory, and HR-MRI variables, including the hyperintense acute reperfusion marker (HARM) sign and plaque characteristics. Eight machine learning algorithms were trained and evaluated using the area under the receiver-operating characteristic curve (AUC). The SHapley Additive exPlanations (SHAP) method was utilized for model interpretability. The random forest (RF) model demonstrated the optimal predictive performance, achieving AUCs of 0.935, 0.898, and 0.863 in the training, testing, and external validation cohorts, respectively. Feature importance analysis revealed that the HARM sign, plaque enhancement grade, cerebral microbleeds, infarct core volume, baseline National Institutes of Health Stroke Scale score, and fasting blood glucose were the six most robust independent predictors of HT. The SHAP visualization provided transparent and personalized risk profiles for individual patients. The developed RF model, integrating multidimensional HR-MRI parameters and clinical variables, provides an accurate, robust, and interpretable tool for predicting HT, thereby facilitating personalized risk stratification and precision stroke management.