A Novel Explainable Attention-Based Meta-Learning Framework for Imbalanced Brain Stroke Prediction

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Machine Learning for Predictive Diagnostics

Background:

• Medical datasets for brain stroke prediction often exhibit class imbalance, hindering conventional machine learning model performance.
• Accurate stroke prediction is crucial for timely diagnosis and effective patient management.
• Interpretability of AI models in healthcare is essential for building trust and clinical adoption.

Purpose of the Study:

• To develop and evaluate a novel meta-learning framework for enhanced brain stroke prediction.
• To address the challenge of imbalanced datasets in medical prediction tasks.
• To integrate explainable artificial intelligence (XAI) for transparent model insights.

Main Methods:

• A hybrid resampling strategy using SMOTE and SMOTEENN to manage class imbalance.
• Dynamic feature selection to minimize data noise and improve model focus.
• A meta-learning ensemble combining Random Forest and LightGBM, refined by a deep learning meta-classifier.
• Integration of SHAP (Shapley Additive Explanations) for feature interpretability.

Main Results:

• The proposed framework demonstrated superior performance across three datasets (DF-1, DF-2, DF-3) compared to state-of-the-art methods.
• Achieved high accuracy (e.g., 0.992189 on DF-1) and F1-Scores (e.g., 0.992579 on DF-1).
• Significantly improved detection of minority-class instances without compromising overall performance.

Conclusions:

• The meta-learning framework offers a robust and effective solution for brain stroke prediction, particularly with imbalanced data.
• The integration of XAI enhances model transparency and trustworthiness.
• This approach provides a foundation for applying advanced AI techniques to other imbalanced medical prediction challenges.