Multi-Scale Spatio-Temporal Attention Network for Epileptic Seizure Prediction

Area of Science:

• Neurology
• Biomedical Engineering
• Machine Learning

Background:

• Epileptic seizure prediction from electroencephalogram (EEG) data is crucial for epilepsy management.
• Existing methods often struggle with complex, noisy EEG data, leading to low prediction accuracy due to single-scale feature extraction.

Purpose of the Study:

• To develop an advanced deep learning model for improved epileptic seizure prediction.
• To address limitations in feature extraction from noisy EEG data.

Main Methods:

• Proposed a Multi-scale Spatio-temporal Attention Network (MSAN) incorporating a backbone module, spatial pyramid module, and multi-scale sequential aggregation module.
• Utilized Long Short-Term Memory (LSTM) blocks for temporal feature aggregation.
• Implemented a dual-loss function to mitigate class imbalance.

Main Results:

• Achieved 96.27% average sensitivity and 0.00/h false prediction rate on the CHB-MIT dataset.
• Attained 93.57% average sensitivity and 0.044/h false prediction rate on the Kaggle dataset.
• Outperformed 10 state-of-the-art seizure prediction models.

Conclusions:

• The MSAN model demonstrates superior performance in epileptic seizure prediction.
• Multi-scale spatio-temporal feature learning effectively enhances prediction accuracy.
• The proposed method offers a promising advancement for clinical epilepsy diagnosis and treatment.