Interpreting deep learning models for epileptic seizure detection on EEG signals

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Deep Learning (DL) models are state-of-the-art for medical decision support but lack clinical trust due to poor interpretability.
• Interpretable AI is crucial for integrating advanced computational tools into clinical practice, especially for time-series data like EEG.

Purpose of the Study:

• To develop an interpretable Deep Learning model for online epileptic seizure detection using EEG signals.
• To align DL model behavior with expert medical knowledge for enhanced clinical adoption.
• To investigate feature interpretability and model behavior in the context of seizure detection.

Main Methods:

• Developed a DL model using EEG signals, incorporating domain knowledge in signal processing, network architecture, and output post-processing.
• Analyzed frequency patterns in the first convolutional layer and their relation to standard EEG bands (delta, theta, alpha, beta, gamma).
• Utilized the DeepLIFT method to identify signal waveforms contributing most to seizure prediction.

Main Results:

• Kernel size in the first layer impacts feature interpretability and model sensitivity, with similar final performance after post-processing.
• Signal amplitude was identified as the primary feature for seizure prediction, suggesting a need for larger datasets to learn complex frequency patterns.
• The methodology generalized well across the patient population, achieving an F1-score of 0.873 and detecting 90% of seizures.

Conclusions:

• Integrating expert knowledge into DL model development enhances interpretability and clinical relevance for seizure detection.
• The study demonstrates a viable approach to building trusted AI tools for neurological disorder diagnosis.
• Future work should focus on larger datasets to capture more intricate frequency patterns for improved seizure detection accuracy.