Prediction of recurrent ischemic stroke using machine learning from real-world data.

Area of Science:

• Neurology
• Artificial Intelligence
• Data Science

Background:

• Recurrent ischemic stroke (RIS) affects approximately 33% of patients in Malaysia.
• Limited studies exist on using artificial intelligence (AI) for RIS prediction with real-world data.

Purpose of the Study:

• To develop and evaluate machine learning models (SVM, KNN, RUSBoost) for predicting RIS.
• To utilize real-world data from the Malaysian National Neurology Registry.

Main Methods:

• Retrospective study of 7,697 patients (2009-2016).
• Development and evaluation of SVM, KNN, and RUSBoost models.
• Application of Synthetic Minority Over-Sampling Technique (SMOTE) for imbalanced data.
• Ten-fold cross-validation for model assessment using accuracy, sensitivity, specificity, PPV, and AUC.

Main Results:

• RUSBoost demonstrated superior performance with an AUROC of 0.943 and 86.5% sensitivity.
• SHAP analysis identified age, race, glucose, hypertension, hyperlipidemia, and diabetes duration as significant risk factors.
• SMOTE improved RUSBoost's discrimination during training (AUROC=0.986).

Conclusions:

• Machine learning on real-world data is a promising tool for predicting RIS risk.
• RUSBoost is a reliable and generalizable model for clinical risk prediction.
• Integrating AI into clinical practice can improve early treatment decisions and preventive strategies for recurrent stroke.