An interpretable machine learning model integrating DWI radiomics and clinical variables (Combined-LR) achieved superior performance for predicting 90-day post-thrombolysis outcomes in acute ischemic stroke, with transcriptomic analyses providing hypothesis-generating insights into D-dimer-related mechanisms.
Key Findings
Results
The Radiomics-LR model demonstrated superior performance among the four machine learning models tested for predicting 90-day outcomes after intravenous thrombolysis.
Training AUC of 0.926 and testing AUC of 0.857 for the Radiomics-LR model.
Four machine learning models were evaluated: Logistic Regression, Decision Tree, LDA, and LightGBM.
DWI radiomic features were extracted and selected prior to model construction.
The study included 115 AIS patients treated with intravenous thrombolysis between January 2024 and September 2025.
Results
The Combined-LR model, integrating DWI radiomics with clinical factors including D-dimer, achieved the best overall predictive performance.
Combined-LR achieved training AUC of 0.939 and testing AUC of 0.862.
Optimism-corrected AUC was 0.838 by bootstrap and 0.772 by cross-validation.
The Combined-LR significantly outperformed the clinical-only model (training P < 0.001, testing P = 0.036).
The model was interpreted using SHAP (SHapley Additive exPlanations) for interpretability.
Results
D-dimer was identified as an independent clinical risk factor for 90-day functional outcomes after intravenous thrombolysis in AIS patients.
D-dimer was incorporated as an independent risk factor into the combined model.
Patients were stratified by 90-day modified Rankin Scale (mRS) scores into favorable (0–2) and unfavorable outcomes.
The study design was retrospective, including 115 patients from a single center.
D-dimer's role as an independent predictor motivated transcriptomic analysis of its upstream regulatory mechanisms.
Results
Decision curve analysis indicated that the Combined-LR model provided higher net clinical benefit across multiple risk threshold intervals compared to the clinical model.
Decision curve analysis was used to assess practical utility for clinical prognostication.
The Combined-LR showed greater net benefit across multiple risk threshold intervals.
This suggests greater practical utility for clinical prognostication relative to the clinical-only model.
The analysis was conducted alongside AUC comparisons to evaluate overall model performance.
Results
Transcriptomic analysis of rat MCAO datasets revealed upregulation of complement and coagulation pathway genes potentially related to D-dimer mechanisms.
Two publicly available rat middle cerebral artery occlusion (MCAO) transcriptomic datasets were analyzed.
Upregulated genes included Serpine1 (PAI-1), C5ar1, and Itgam (CD11b).
These genes are involved in complement and coagulation cascades.
The transcriptomic analysis was described as an 'exploratory, hypothesis-generating step' and not a confirmatory finding.
This analysis was conducted to investigate D-dimer-related molecular mechanisms in ischemic stroke.
Methods
The study enrolled 115 AIS patients treated with intravenous thrombolysis in a retrospective single-center design.
Study period was January 2024 to September 2025.
Outcome classification used the 90-day modified Rankin Scale (mRS): favorable (0–2) vs. unfavorable outcomes.
The retrospective design and single-center origin limit generalizability; the authors note that 'external multi-center validation is required.'
DWI (Diffusion-Weighted Imaging) was the imaging modality used for radiomic feature extraction.
What This Means
This research suggests that combining brain imaging data with blood test results can better predict how well stroke patients will recover three months after receiving clot-dissolving treatment. The researchers analyzed 115 patients who received intravenous thrombolysis (a common treatment for acute ischemic stroke) and built computer models that used features extracted from diffusion-weighted MRI scans alongside a blood marker called D-dimer. Their best-performing model (Combined-LR) was significantly more accurate at predicting whether patients would have good or poor functional outcomes at 90 days compared to a model using only clinical information, and the model's decisions could be explained using a technique called SHAP, making it more transparent for potential clinical use.
This research also suggests that D-dimer — a protein fragment released when blood clots break down — may play a meaningful role in stroke recovery, not just as a blood test marker but potentially through biological pathways involving inflammation and coagulation. An exploratory analysis of publicly available rat stroke data found that genes related to complement activation and blood clotting, including PAI-1 (Serpine1), C5ar1, and CD11b (Itgam), were upregulated, hinting at possible molecular mechanisms linking D-dimer levels to stroke outcomes. The authors are careful to note this is hypothesis-generating only and not a confirmed mechanism.
The practical implication of this work is that incorporating imaging-based computational features alongside routine blood tests could improve doctors' ability to identify which stroke patients are at higher risk for poor recovery after thrombolysis, potentially informing more tailored follow-up care. However, the study was conducted at a single center with a relatively small sample of 115 patients, and the authors emphasize that validation across multiple centers is necessary before such a model could be considered for broader clinical use.
Bai Y, Liu Q, Zhan R, Xu Q, Xia C, Qian Y. (2026). Combined clinical and DWI radiomics model for predicting 90-day functional outcomes after intravenous thrombolysis in acute ischemic stroke patients.. Frontiers in neurology. https://doi.org/10.3389/fneur.2026.1847290