DiffEEG: A Self-Supervised Denoising Diffusion Model for Learning EEG Generic Representations

2026-07-13

AI Summary

This work addresses the challenges of scarce annotations and extreme class imbalance (seizure segments constituting less than 10%) in EEG-based epilepsy detection by proposing the first self-supervised foundation model based on denoising diffusion. The approach employs a 1D U-Net architecture augmented with multi-head self-attention for pretraining on large-scale unlabeled EEG data to learn generalizable neural representations. It further introduces an innovative policy gradient reinforcement learning fine-tuning mechanism that directly optimizes the clinically critical F1 score. Under strict patient-wise evaluation, the model achieves 59% F1 on a four-class seizure subtype classification task, 85% weighted F1 and 59% seizure recall on binary detection, and 97.6% segment-level accuracy, substantially reducing reliance on labeled data while enhancing sensitivity to rare seizure events.

Abstract

Deep learning for EEG-based seizure detection faces critical challenges: severe annotation scarcity and extreme class imbalance, where ictal events comprise less than 10\% of clinical recordings. We present DiffEEG, a 9.6M-parameter self-supervised foundation model that addresses both limitations through denoising diffusion pre-training and reinforcement learning (RL)-based fine-tuning. Pre-trained on 1.3M unlabeled segments from the Temple University Hospital Seizure Corpus (TUHSZ), DiffEEG learns generic neural representations via a 1D U-Net with multi-head self-attention. For downstream adaptation, a reinforced decision layer employs policy gradient optimization to directly maximize F1-score, prioritizing sensitivity to rare seizure events over overall accuracy. Under strict patient-wise evaluation (279 patients, Leave-One-Fold-Out), DiffEEG achieves 61\% accuracy and 59\% F1 for 4-class seizure subtyping, and 81\% accuracy with 85\% weighted F1 for binary detection, maintaining clinically viable seizure recall (59\%) despite extreme imbalance (6.7\% prevalence). Segment-level evaluation establishes an upper bound of 97.6\% accuracy, confirming strong architectural capacity. DiffEEG demonstrates that diffusion-based pre-training combined with metric-aware reinforcement learning enables clinically deployable seizure monitoring with minimal labeled data requirements.

Problem

• EEG
• seizure detection
• annotation scarcity
• class imbalance
• self-supervised learning

Innovation

• denoising diffusion
• self-supervised learning
• reinforcement learning
• EEG representation learning
• seizure detection