Scholarlink

Stroke is a major cause of long-term disability. Early and accurate diagnosis of stroke severity can improve patient outcomes. EEG is a non-invasive way to analyze brain activity changes during stroke, but interpreting complex EEG data remains challenging. This study develops an explainable multi-task learning approach for EEG-based stroke analysis. Feedforward neural networks extract hierarchical spatial-temporal EEG features. An ensemble of models with different architectures and electrode combinations is trained for robustness. A multi-task framework jointly models stroke diagnosis, type, location, and severity prediction with high accuracies of 98.11%, 100%, 100%, and 100% respectively. Explainability methods like LIME and SHAP are used to visualize influential EEG channels and time windows for each prediction. The models are evaluated on a public stroke EEG dataset and achieve state-of-the-art performance on multi-label classification and severity regression. Explanation methods provide clinically interpretable insights into key EEG patterns underlying decision-making. This presents an effective and transparent framework for multi-faceted EEG-based stroke analysis to support clinical decisions.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Ensemble Multi-task Learning Approach for Explainable EEG-Based Stroke Prediction

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Salma Nbili,
  • Samar Bouazizi,
  • Hela Ltifi

摘要

Stroke is a major cause of long-term disability. Early and accurate diagnosis of stroke severity can improve patient outcomes. EEG is a non-invasive way to analyze brain activity changes during stroke, but interpreting complex EEG data remains challenging. This study develops an explainable multi-task learning approach for EEG-based stroke analysis. Feedforward neural networks extract hierarchical spatial-temporal EEG features. An ensemble of models with different architectures and electrode combinations is trained for robustness. A multi-task framework jointly models stroke diagnosis, type, location, and severity prediction with high accuracies of 98.11%, 100%, 100%, and 100% respectively. Explainability methods like LIME and SHAP are used to visualize influential EEG channels and time windows for each prediction. The models are evaluated on a public stroke EEG dataset and achieve state-of-the-art performance on multi-label classification and severity regression. Explanation methods provide clinically interpretable insights into key EEG patterns underlying decision-making. This presents an effective and transparent framework for multi-faceted EEG-based stroke analysis to support clinical decisions.