A stroke prediction framework using explainable ensemble learning.

Area of Science:

• Computational Neuroscience
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Stroke is a critical neurological event caused by interrupted blood flow to the brain, leading to brain cell death.
• Early stroke symptom recognition is vital for prevention and promoting healthy lifestyles, but current diagnostic tests like FAST have limitations.
• Existing stroke prediction methods require enhancement for improved accuracy and reliability.

Purpose of the Study:

• To develop and assess multiple machine learning (ML) models for a robust stroke risk prediction framework.
• To establish a superior stroke prediction model using a stacking-based ensemble method.
• To identify key risk factors contributing to stroke prediction through model interpretability.

Main Methods:

• Development and evaluation of multiple machine learning (ML) models.
• Implementation of a stacking-based ensemble technique to combine the intelligence of the top three ML models.
• Utilized Shapley's Additive Explanations (SHAP) for analyzing black-box ML model predictions.

Main Results:

• The proposed stacking-based ensemble model demonstrated superior performance on a public stroke prediction dataset, with only one misclassification.
• Achieved exceptional performance metrics: 99.99% accuracy, precision, and F1-score; 100% recall; and perfect ROC, MCC, and Kappa scores of 1.0.
• SHAP analysis identified age, Body Mass Index (BMI), and glucose levels as the most significant risk factors for stroke.

Conclusions:

• The stacking-based ensemble ML model offers a highly accurate and robust framework for stroke risk prediction.
• This advanced ML approach surpasses current state-of-the-art methods in stroke prediction accuracy.
• Identifying key risk factors like age, BMI, and glucose levels can guide targeted preventive strategies.