GRaph-based analysis for stroke prediction (GRASP): A multi-modal model for identifying first ischemic stroke in...

Area of Science:

• Medical Informatics
• Cardiovascular Diseases
• Artificial Intelligence in Medicine

Background:

• Ischemic stroke causes millions of deaths globally, with current prediction models lacking accuracy.
• Existing models struggle to identify individuals at high risk who will experience a first ischemic stroke (FIS).

Purpose of the Study:

• To develop and evaluate GRaph-based Analysis for Stroke Prediction (GRASP), a novel multimodal approach for predicting FIS.
• To enhance the identification of individuals likely to experience FIS within at-risk populations sharing similar vascular risk profiles.

Main Methods:

• Utilized UK Biobank data from 1226 participants with vascular risk factors, including 317 who developed FIS.
• Employed a Graph Attention Network architecture integrating 136 variables (demographic, clinical, lifestyle, neuroimaging).
• Compared GRASP performance against baseline models and conventional machine learning classifiers.

Main Results:

• GRASP achieved an AUC-ROC of 0.82 and PR-AUC of 0.73, outperforming XGBoost (AUC-ROC 0.69, PR-AUC 0.60) and baseline models (AUC-ROC 0.55, PR-AUC 0.28).
• Demonstrated significant improvements in distinguishing at-risk individuals with future ischemic stroke (AR-FIS) from those without (AR-NIS).
• Optimal performance was observed in younger participants (≤55 years), with an AUC-ROC of 0.82.

Conclusions:

• GRASP significantly enhances FIS risk prediction in at-risk populations by integrating multimodal data within a graph-based framework.
• The model offers superior clinical risk stratification compared to traditional risk-factor-based models.
• GRASP effectively identifies individuals who will develop FIS, even those with similar vascular risk profiles.