Related Experiment Video
Updated: Jan 9, 2026
A Fibrin-Enriched and tPA-Sensitive Photothrombotic Stroke Model
Published on: June 4, 2021
Factors Influencing Stroke Severity Based on Collateral Circulation, Clinical Markers and Machine Learning
1Department of Applied Statistics, National Taichung University of Science and Technology, Taichung 404336, Taiwan.
Insights
Stroke severity is significantly influenced by collateral circulation and stroke laterality. Machine learning models, particularly tree-based ensembles, accurately predict stroke severity using clinical and imaging data, aiding personalized patient care.
Area of Science:
- Neurology
- Medical Imaging
- Machine Learning
Background:
- Stroke is a leading cause of disability, with severity varying based on multiple factors.
- Understanding stroke severity determinants is crucial for patient outcomes and treatment planning.
- Collateral circulation plays a significant, yet often understudied, role in stroke severity.
Purpose of the Study:
- To identify and analyze key variables influencing stroke severity.
- To investigate the specific role of collateral circulation in determining stroke severity.
- To evaluate the predictive performance of machine learning models for stroke severity.
Main Methods:
- Analysis of clinical (SBP, FPG, BUN), imaging (ipsilateral collateral flow, unilateral-bilateral stroke), and biochemical data.
- Application of statistical tests (chi-square, Mann-Whitney U) for group comparisons.
- Utilized SMOTE for class imbalance, followed by cross-validation of Logistic Regression, Random Forest, XGBoost, and SVM models.
Main Results:
- Reduced or absent ipsilateral collateral flow and unilateral-bilateral stroke were strongly linked to increased severity (p < 0.001).
- Systolic blood pressure (SBP) and fasting plasma glucose (FPG) showed significant associations with stroke severity.
- Random Forest and XGBoost models, trained on SMOTE-balanced data, demonstrated high predictive accuracy (83.3% and 80.2% respectively).
Conclusions:
- Collateral status and stroke laterality are primary determinants of stroke severity.
- SBP and FPG provide additional prognostic value, while BUN is borderline significant.
- Tree-based ensemble models trained with SMOTE offer reliable stroke severity prediction for risk stratification and personalized care planning.
Abstract:
Background/Objectives: Stroke is a serious neurological disorder that significantly affects patients' quality of life and overall health. The severity of a stroke can vary widely and is influenced by multiple factors, such as clinical presentation, diagnostic findings, and the site of onset. This study aimed to identify and analyze key variables that contribute to stroke severity, with a particular focus on the role of collateral circulation. Methods: This study analyzed clinical, imaging, and biochemical variables-ipsilateral collateral flow on MRA, MRI unilateral-bilateral stroke, systolic blood pressure (SBP), fasting plasma glucose (FPG), and blood urea nitrogen (BUN). Group differences used chi-square and Mann-Whitney U tests. Class imbalance was addressed with SMOTE; Logistic Regression, Random Forest, XGBoost, and SVM were cross-validated, reporting accuracy, precision, recall, and F1 with 95% CIs. Results: Reduced or absent ipsilateral collateral flow and unilateral-bilateral stroke were strongly associated with greater severity (p < 0.001). SBP was significant (p = 0.034), FPG was significant (p = 0.023), and BUN was borderline (p = 0.059). SMOTE improved prediction: Random Forest achieved accuracy 83.3% (CI: 79.1-87.6) and F1 84.0% (CI: 79.1-88.9); XGBoost reached accuracy 80.2% (CI: 71.5-89.0) and F1 81.4% (CI: 73.8-89.0). Logistic Regression improved to F1 70.8% (CI: 55.4-86.2), whereas SVM declined to accuracy 52.2% (CI: 37.5-67.0). Conclusions: Collateral status and unilateral-bilateral stroke are key determinants of severity; SBP and FPG add prognostic value, with BUN borderline. Tree-based ensembles trained on SMOTE-balanced data provide the most reliable predictions for risk stratification. These findings suggest that future work may focus on integrating such predictive models into Clinical Decision Support Systems (CDSSs) to enhance early risk identification, strengthen CDSSs, and enable more personalized care planning for stroke patients.
