Enhancing Brain Signal Generation Through a Hybrid Approach Integrating Reinforcement Learning and Diffusion Models

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Reliable Brain-Computer Interface (BCI) systems require extensive EEG datasets.
• Data collection is limited by subject fatigue and interindividual variability.
• Existing data augmentation methods struggle with EEG signal complexity.

Purpose of the Study:

• To propose a reinforcement learning-enhanced EEG diffusion (RLED) framework for adaptive data augmentation.
• To improve classification performance in endogenous EEG tasks like motor imagery and emotion recognition.
• To address challenges in EEG data collection for BCI development.

Main Methods:

• Developed a reinforcement learning (RL) mechanism to dynamically control the diffusion training process.
• Integrated RL to balance temporal, spectral, and class-related EEG features.
• Applied the RLED framework to four diverse EEG datasets.

Main Results:

• The RLED framework successfully generated high-quality synthetic EEG signals.
• Consistent improvements in classification performance were observed across datasets.
• Demonstrated the framework's effectiveness for EEG data augmentation and generalization.

Conclusions:

• The proposed RLED framework offers a promising solution for EEG data augmentation.
• RLED enhances the generalization capabilities of BCI systems.
• This approach can overcome limitations in collecting large-scale EEG training data for BCI applications.