Prediction of Long-Term Stroke Recurrence Using Machine Learning Models.

Area of Science:

• Biomedical Informatics
• Clinical Neurology
• Machine Learning in Healthcare

Background:

• Individual-level prediction of long-term recurrent ischemic stroke risk remains challenging.
• Current risk assessment methods have limitations in accuracy and personalization.

Purpose of the Study:

• To develop and evaluate machine-learning models for predicting ischemic stroke recurrence.
• To identify key clinical variables associated with stroke recurrence.
• To optimize model performance metrics for clinical implementation.

Main Methods:

• Utilized patient-level electronic health record data.
• Trained six interpretable machine-learning algorithms with four feature selection strategies.
• Developed 288 models for up to 5-year stroke recurrence prediction using various sampling strategies.

Main Results:

• Model performance (AUROC) was stable across 1-5 year prediction windows, with highest scores for 1-year predictions (0.79).
• Age, BMI, and laboratory values (HDL, HbA1c, creatinine) were identified as highly important features.
• Sampling strategies improved the balance between model specificity and sensitivity.

Conclusions:

• Machine learning algorithms can effectively predict long-term stroke recurrence.
• Laboratory-based variables are strongly associated with recurrence and can guide personalized interventions.
• Optimized models can serve as decision support tools for targeted stroke prevention strategies.