EEG-Based Seizure Prediction via Model Uncertainty Learning.

Area of Science:

• * Neuroscience and Biomedical Engineering
• * Artificial Intelligence in Healthcare

Background:

• * Deep neural networks (DNNs) excel at feature extraction for electroencephalogram (EEG) based seizure prediction.
• * Existing DNN models lack the ability to quantify prediction uncertainty, limiting their clinical credibility.
• * Patient-specific seizure prediction remains a challenge due to model unreliability.

Purpose of the Study:

• * To develop a novel end-to-end patient-specific seizure prediction framework using model uncertainty learning.
• * To enhance the reliability and credibility of DNN-based seizure prediction models.
• * To introduce a lightweight convolutional neural network (CNN) architecture with uncertainty quantification.

Main Methods:

• * Proposed a reparameterized EEG-based lightweight CNN (RepNet) using depthwise separable convolutions for multi-scale feature extraction.
• * Developed a modified Monte Carlo dropout (MMCD) strategy to simulate dropout sampling and leverage temporal information in EEG signals.
• * Integrated RepNet with MMCD (RepNet-MMCD) for uncertainty-aware seizure prediction.

Main Results:

• * RepNet-MMCD achieved high performance on two public datasets: Sensitivity (93.1%, 81.6%), False Positive Rate (FPR) (0.033/h, 0.056/h), and Area Under Curve (AUC) (0.950, 0.903).
• * The RepNet architecture significantly reduces computational cost by converting multi-scale convolutions into a single layer post-training.
• * The MMCD strategy demonstrated performance improvements when applied to other baseline seizure prediction methods.

Conclusions:

• * The proposed RepNet-MMCD framework offers a reliable and computationally efficient approach for patient-specific seizure prediction.
• * Incorporating model uncertainty learning significantly enhances the credibility of DNN-based seizure prediction.
• * The MMCD strategy is a versatile technique that can improve the performance of various seizure prediction models.