Combined clinical and DWI radiomics model for predicting 90-day functional outcomes after intravenous thrombolysis in acute ischemic stroke patients

Abstract

Objective To develop an interpretable machine learning model based on Diffusion-Weighted Imaging (DWI) radiomics and D-dimer for 90-day outcome prediction after intravenous thrombolysis in acute ischemic stroke (AIS), and to explore candidate upstream transcriptional programs of D-dimer using publicly available rat middle cerebral artery occlusion (MCAO) datasets as an exploratory, hypothesis-generating step. Methods This retrospective study included 115 AIS patients treated with intravenous thrombolysis (January 2024–September 2025). Patients were stratified by 90-day modified Rankin Scale (mRS) scores into favorable (0–2) and unfavorable (1–4) outcome groups. DWI radiomic features were extracted and selected to construct four machine learning models (Logistic Regression, Decision Tree, LDA, LightGBM). Independent clinical risk factors were identified to build a clinical model. The best machine learning model was combined with clinical factors to create a combined model (Combined-LR), interpreted via SHAP. Additionally, transcriptomic analysis of two rat MCAO datasets was conducted to investigate D-dimer-related molecular mechanisms. Results The Radiomics-LR model showed superior performance (training AUC 0.926, testing AUC 0.857). D-dimer was an independent risk factor. Combined-LR achieved the best performance (training AUC 0.939, testing AUC 0.862; optimism-corrected AUC 0.838 by bootstrap, 0.772 by cross-validation), significantly outperforming the clinical model (training P < 0.001, testing P = 0.036). Decision curve analysis indicated that the Combined-LR provided higher net clinical benefit across multiple risk threshold intervals, suggesting greater practical utility for clinical prognostication. Transcriptomic analysis revealed upregulation of complement/coagulation genes, including Serpine1 (PAI-1), C5ar1, and Itgam (CD11b). Conclusion We established an interpretable machine learning model integrating DWI radiomics and clinical variables for predicting 90-day post-thrombolysis outcomes in AIS. Transcriptomic analyses provided hypothesis-generating insights into D-dimer-related mechanisms. External multi-center validation is required.