Machine Learning-Based Early Detection and Risk Stratification of Brain Stroke: Analytical Approach to Predictive Diagnosis and Preventive Remedies

Abstract

Stroke is a leading global cause of death and disability, with outcomes highly dependent on early detection and timely prevention. This study proposes a multimodal machine learning (ML) framework integrating Electronic Health Records (EHRs), radiomic features from CT/MRI, and wearable biosignals to predict and stratify stroke risk. The framework employs Gradient Boosting Machines for structured data, 3D-CNNs for imaging, and LSTMs for sequential signals, with predictions fused via a meta-learner and clustered into low, moderate, and high-risk groups using K-means. Trained on 10,247 patients and validated on 2,031, the ensemble achieved an AUROC of 0.91, surpassing the CHA₂DS₂-VASc score (0.76). SHAP analysis identified atrial fibrillation burden, blood-pressure variability, and carotid-plaque volume as key predictors, while Kaplan–Meier curves showed clear risk separation (4%, 12%, 28% incidence). Simulated prevention pathways indicated a potential 17% reduction in stroke events and $1.4M savings per 10,000 individuals screened. These findings underscore the potential of interpretable ML models in enabling personalized, precision-based preventive neurology