GenEEG: Improving epileptic EEG detection through patient-adaptive latent diffusion and continual learning.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Neuroscience

Background:

• Automated seizure detection faces challenges like limited clinical EEG data, class imbalance, patient variability, and catastrophic forgetting.
• These issues hinder the effectiveness of machine learning models for epilepsy monitoring and prediction.

Purpose of the Study:

• To introduce GenEEG, a continual learning framework for generating adaptive synthetic EEG data to improve class-imbalanced seizure detection.
• To address data scarcity and enhance the generalizability of epilepsy seizure detection systems.

Main Methods:

• Developed a dual-conditioned variational autoencoder (VAE) and latent diffusion model (LDM) for controlled synthetic EEG generation.
• Implemented a continual learning leave-one-patient-out (CL-LOPO) protocol with fold-specific normalization.
• Utilized a hybrid approach combining Elastic Weight Consolidation and experience replay to mitigate catastrophic forgetting.

Main Results:

• GenEEG achieved macro F1-scores of 0.84 (adult) and 0.82 (pediatric) on Siena Scalp EEG and CHB-MIT datasets.
• GenEEG-augmented classifiers showed a 15 percentage point improvement over traditional oversampling baselines (F1: 0.84 vs. 0.69).
• Maintained ictal sensitivity above 75% across diverse populations and demonstrated memory efficiency.

Conclusions:

• GenEEG offers a reproducible and clinically relevant solution for data scarcity in seizure detection.
• The framework effectively generates synthetic EEG data that adapts to patients, improving classification metrics.
• While excelling in low-frequency capture, high-frequency fidelity requires further architectural improvements.