Robustness of ML-Based Seizure Prediction Using Noisy EEG Data From Limited Channels.

Area of Science:

• Neuroscience
• Biomedical Engineering
• Artificial Intelligence

Background:

• Epilepsy affects over 50 million globally, with many experiencing uncontrollable seizures.
• Deep learning (DL) models show high accuracy on clinical EEG data for seizure prediction but struggle in real-world applications.
• Consumer-grade wearable EEG headsets present challenges due to lower data quality and variable channel counts.

Purpose of the Study:

• To evaluate the robustness of DL seizure prediction models trained on clinical EEG data when applied to consumer-grade wearable EEG data.
• To assess the impact of reduced channel count, streaming data, and data imbalance on model performance.

Main Methods:

• Utilized the SPERTL deep learning model for seizure prediction.
• Tested SPERTL's performance on simulated consumer-grade EEG data, varying channel numbers, data stream characteristics, and class imbalance.
• Compared results against a baseline SPERTL model configured for channel independence.

Main Results:

• The SPERTL model demonstrated resilience to noisy and streaming EEG data, with minimal degradation in Area Under the Curve (AUC) (2-3%).
• Performance significantly degraded with reduced channel count (up to 16% AUC reduction) and increased data imbalance (up to 32% AUC reduction).
• Baseline AUC for the channel-independent SPERTL model was 98.56%.

Conclusions:

• DL models exhibit surprising robustness to real-world data variations like noise and streaming, but channel reduction and data imbalance pose significant challenges for seizure prediction accuracy.
• Findings highlight the need for further research into robust DL architectures and data preprocessing techniques for reliable seizure prediction using consumer-grade wearable EEG devices.