A Self-Interpretable Deep Learning Model for Seizure Prediction Using a Multi-Scale Prototypical Part Network.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Epileptic seizure prediction (ESP) is vital for patient quality of life.
• Deep learning models show promise but lack clinical trust due to their "black-box" nature.
• Existing interpretability methods are post hoc, not inherent to the model.

Purpose of the Study:

• To develop a self-interpretable deep learning model for patient-specific epileptic seizure prediction.
• To provide transparent reasoning and decision basis for ESP.
• To enhance clinical applicability of deep learning in epilepsy management.

Main Methods:

• Proposed Multi-Scale Prototypical Part Network (MSPPNet) for patient-specific ESP.
• MSPPNet measures input similarity to learned prototypes for transparent predictions.
• Utilized multi-scale prototypes to capture diverse EEG signal features.

Main Results:

• Achieved state-of-the-art performance on public epileptic EEG datasets (CHB-MIT and Kaggle).
• Demonstrated high sensitivity (93.8% on CHB-MIT, 88.6% on Kaggle) with low false prediction rates.
• Provided self-interpretable evidence alongside predictions, enhancing transparency.

Conclusions:

• MSPPNet is the first self-interpretable deep learning model for epileptic seizure prediction.
• The model offers transparent reasoning, addressing clinical mistrust in deep learning.
• This approach holds significant potential for advancing clinical ESP applications.