A Transparent AI Model for Predicting Post-Thrombolytic Outcomes in Acute Ischemic Stroke Patients

OBJECTIVES To develop and validate an interpretable artificial intelligence (AI) framework to predict early neurological improvement and long-term functional outcomes after intravenous thrombolysis in acute ischemic stroke (AIS), identifying key prognostic biomarkers and implementing clinically actionable decision-support tools. METHODS We conducted a retrospective analysis of clinical data from AIS patients treated with intravenous recombinant tissue plasminogen activator (rtPA) between January 2023 and October 2024. Functional outcomes were prospectively followed, and risk factors for poor prognosis were thoroughly examined. Deep neural network (DNN) and Kolmogorov-Arnold Network (KAN) models were developed to predict 3-month neurological outcomes in AIS patients following thrombolysis. RESULTS The DNN model exhibited superior predictive performance for both early neurological improvement (ΔNIHSS, AUC = 0.985) and long-term functional outcomes (90-day mRS, AUC = 0.828) compared to five alternative machine learning algorithms. SHapley Additive exPlanations (SHAP) analysis highlighted consistent key predictors for both outcome, including uric acid (UA), platelet count (PLT), onset-to-treatment time (OTT), age, door-to-needle time (DNT), and admission NIHSS score(National Institute of Health stroke scale). Stratification analysis by Oxfordshire Community Stroke Project (OCSP) classification confirmed strong performance across stroke subtypes, with subgroup AUCs reaching 1.00 for posterior circulation infarcts (POCI) in ΔNIHSS prediction and 0.970 for lacunar infarcts (LI) in mRS prediction. KAN modeling further validated these predictors' reliability (mRS AUC = 0.97, ΔNIHSS AUC = 1.00) and provided precise mathematical formulations of relationships between clinical variables and outcomes (e.g., UA nonlinear function: f(x) = 0.03(0.441-x)³-0.137, R²=0.83). An online clinical decision support platform incorporating the top SHAP features was deployed to enable real-time prognostication for AIS patients undergoing thrombolysis. CONCLUSIONS Our clinically validated AI framework accurately predicts post-thrombolysis outcomes in AIS by identifying key prognostic biomarkers in an interpretable manner. The implemented online platform enables real-time risk stratification, facilitating personalized therapeutic decision-making and potentially improving patient outcomes.