Explainable End-to-End Seizure Prediction via Stationary Wavelet Transform-Driven Dynamic Multiscale Fuzzy Clustering

Area of Science:

• Neuroscience and Biomedical Engineering
• Artificial Intelligence in Healthcare

Background:

• Epileptic seizure prediction is crucial for patient quality of life.
• Existing methods struggle with inter-subject EEG variability and complex spatiotemporal dynamics, limiting feature discriminability and model explainability.
• The "black-box" nature of deep learning models hinders clinical adoption.

Purpose of the Study:

• To develop an explainable epileptic seizure prediction framework addressing EEG variability and model transparency.
• To integrate advanced signal processing with transparent clinical decision-making for improved seizure forecasting.
• To enhance both the discriminability of EEG features and the explainability of prediction models.

Main Methods:

• Proposed a novel Stationary Wavelet Transform (SWT)-driven Dynamic Multiscale Fuzzy Clustering (SD-MFC) framework.
• Employed SWT for spectral-temporal decomposition and a geometric attention mechanism for cross-channel dependency modeling.
• Developed a Riemannian manifold-based fuzzy clustering algorithm and hierarchical feature fusion using multiscale convolutional kernels; incorporated contrastive learning for robustness.

Main Results:

• The SD-MFC framework demonstrated superior predictive performance on both intracranial and extracranial EEG datasets.
• Achieved a low False Positive Rate (FPR), indicating high reliability for clinical use.
• Proposed explainability methods (joint feature visualization, feature ablation) bridge the gap between deep learning and clinical diagnostics.

Conclusions:

• The SD-MFC framework offers a feasible and effective solution for clinical application of EEG-based seizure prediction.
• The model's enhanced explainability facilitates trust and adoption in clinical settings.
• This approach addresses key limitations in current seizure prediction technology, paving the way for improved patient care.