Explainable End-to-End Seizure Prediction via Dynamic Multiscale Cross-Band Fusion Filter Network.

Area of Science:

• Medical Artificial Intelligence (AI)
• Neuroscience
• Signal Processing

Background:

• Epileptic seizure prediction using electroencephalogram (EEG) signals is crucial for patient quality of life.
• Existing AI models struggle with insufficient feature representation and limited decision explainability.
• There is a need for advanced models that offer both high accuracy and clinical interpretability.

Purpose of the Study:

• To propose a novel dynamic multiscale cross-band fusion filter network (MCFNet) for end-to-end epileptic seizure prediction.
• To address limitations in feature representation and explainability in current EEG-based prediction models.
• To develop a feasible scheme for clinical application of AI in seizure prediction.

Main Methods:

• EEG signals decomposed into multiscale components with a cross-band fusion attention mechanism for signal fusion.
• A synchronous spectral filtering network with static and dynamic modules captures periodic components and cross-channel dependencies.
• Introduced joint feature visualization and feature ablation analysis for model explainability.

Main Results:

• MCFNet achieved high performance on the CHB-MIT dataset: 97.13% sensitivity and 97.22% specificity.
• The model demonstrated a low false positive rate (FPR) of 0.0326/h.
• Experimental results confirmed superior predictive performance and low FPR, indicating clinical feasibility.

Conclusions:

• MCFNet offers a significant advancement in AI-driven epileptic seizure prediction.
• The model's enhanced feature representation and explainability address key challenges in the field.
• MCFNet presents a viable solution for real-world clinical application of EEG-based seizure prediction.