[Paper Review] EEG Synthetic Data Generation Using Probabilistic Diffusion Models

Giulio Tosato, Cesare Maria Dalbagno+1 more|arXiv (Cornell University)|Mar 6, 2023

EEG and Brain-Computer Interfaces8 citations

TL;DR

The paper proposes generating synthetic EEG data with denoising diffusion probabilistic models trained on electrode-frequency distribution maps to augment training data for emotion classification, showing improved classifier accuracy with synthetic data.

ABSTRACT

Electroencephalography (EEG) plays a significant role in the Brain Computer Interface (BCI) domain, due to its non-invasive nature, low cost, and ease of use, making it a highly desirable option for widespread adoption by the general public. This technology is commonly used in conjunction with deep learning techniques, the success of which is largely dependent on the quality and quantity of data used for training. To address the challenge of obtaining sufficient EEG data from individual participants while minimizing user effort and maintaining accuracy, this study proposes an advanced methodology for data augmentation: generating synthetic EEG data using denoising diffusion probabilistic models. The synthetic data are generated from electrode-frequency distribution maps (EFDMs) of emotionally labeled EEG recordings. To assess the validity of the synthetic data generated, both a qualitative and a quantitative comparison with real EEG data were successfully conducted. This study opens up the possibility for an open extendash source accessible and versatile toolbox that can process and generate data in both time and frequency dimensions, regardless of the number of channels involved. Finally, the proposed methodology has potential implications for the broader field of neuroscience research by enabling the creation of large, publicly available synthetic EEG datasets without privacy concerns.

Motivation & Objective

• Motivate data scarcity in EEG-BCI and the need for high-quality synthetic data.
• Develop a diffusion-based method to generate EEG-like samples from EFDMs.
• Create an open-source toolbox capable of processing time-domain and frequency-domain EEG data.
• Assess whether synthetic data add novel information beyond the original dataset.
• Evaluate classifier performance with real data only vs. real plus synthetic data.

Proposed method

• Adapt OpenAI improved-diffusion model to generate 128-channel, 128x128 EFDM-derived images.
• Construct electrode-frequency distribution maps (EFDMs) from STFT of EEG data (up to 100 Hz).
• Train a classifier in PyTorch using CrossEntropyLoss on real data, then evaluate augmentation effects.
• Train diffusion model with diffusion_steps=1000 and linear noise schedule; image_size=128; batch_size=32.
• Evaluate whether diffusion-generated data replicate or augment real data by comparing classifier performance on unseen real data.
• Provide a GitHub-hosted toolbox and discuss potential future optimizations.

Figure 1: Progressive addition of Gaussian noise.

Experimental results

Research questions

• RQ1 Can diffusion-generated EEG samples provide information not present in the original training data?
• RQ2 Do synthetic samples improve classifier performance when used for data augmentation compared to real data alone?
• RQ3 Is the diffusion model capable of generating novel EEG-like samples rather than memorizing the training set?
• RQ4 How effective is EFDM-based data representation for diffusion-based EEG synthesis?
• RQ5 What are the practical implications and limitations of diffusion-based EEG data augmentation?

Key findings

• Classifier trained on real data achieved over 90% average accuracy on synthetic samples.
• Augmenting real data with diffusion-generated samples improves max average accuracy to 92.634% (40 epochs) and 92.984% (60 epochs).
• Hybrid training (real plus synthetic) consistently outperformed the model trained on real data alone.
• Evidence suggests synthetic data contain novel information beyond the original dataset, supporting augmentation use.
• Diffusion model training up to 60 epochs yields better performance than real-data-only training.
• Synthetic data can be shared without privacy concerns since they are not direct samples from individuals.

Figure 2: Progressive subtraction of Gaussian noise.