Enhancing EEG Signal-Based Emotion Recognition with Synthetic Data: Diffusion Model Approach

Emotions are crucial in human life, influencing perceptions, relationships, behaviour, and choices. Emotion recognition using Electroencephalography (EEG) in the Brain-Computer Interface (BCI) domain presents significant challenges, particularly the need for extensive datasets. This study aims to generate synthetic EEG samples similar to real samples but distinct by augmenting noise to a conditional denoising diffusion probabilistic model, thus addressing the prevalent issue of data scarcity in EEG research. The proposed method is tested on the DEAP and SADT datasets, showcasing up to 5.6% improvement in classification accuracy when using synthetic data with DEAP and similar positive results with SADT. This is higher compared to the traditional Generative Adversarial Network (GAN) based and Denoising Diffusion Probabilistic Model (DDPM) based approaches. This study further evaluates the effectiveness of state-of-the-art classifiers on EEG data, employing both real and synthetic data with varying noise levels, and utilizes t-SNE and SHAP for detailed analysis and interpretability. The proposed diffusion-based approach for EEG data generation appears promising in refining the accuracy of emotion recognition systems and marks a notable contribution to EEG-based emotion recognition.

💡 Research Summary

This paper presents a novel approach to address the critical challenge of data scarcity in Electroencephalography (EEG)-based emotion recognition. The core innovation lies in utilizing a Conditional Denoising Diffusion Probabilistic Model (DDPM), enhanced with a simple yet effective noise augmentation technique, to generate high-quality synthetic raw EEG data.

The study begins by outlining the significance of emotion recognition and the advantages of EEG as a non-invasive, high-temporal-resolution tool, while also acknowledging its limitations, such as low spatial resolution and the difficulty of collecting large-scale datasets. It positions deep learning as a promising solution, the efficacy of which is often bottlenecked by data availability. The authors critique traditional data augmentation methods and Generative Adversarial Networks (GANs) for their potential artifacts and instability (e.g., mode collapse), respectively, and advocate for DDPMs as a more stable and powerful alternative for generating complex time-series data like EEG.

The proposed methodology involves training a conditional DDPM, adapted from image super-resolution models, to generate synthetic EEG signals conditioned on a noise-augmented version of the original signal. This “augmentation module” adds Gaussian noise to the real EEG sample before feeding it as a conditioning input to the model. This key design choice prevents the model from merely replicating the training data and encourages the generation of diverse yet realistic synthetic samples that capture meaningful variations. The process consists of a forward diffusion that gradually adds noise and a reverse denoising process that iteratively refines a pure noise sample into a clean EEG signal, guided by the conditioned input.

The framework is rigorously evaluated on two publicly available emotion recognition datasets: DEAP and SADT. The synthetic data generated by the proposed method is used to augment the training sets of various state-of-the-art classifiers, including EEGNet, TSception, LSTM, and SVM. Experimental results demonstrate a significant improvement in classification accuracy on held-out real test data. Notably, the method achieves up to a 5.6% accuracy gain on the DEAP dataset compared to using only real data, outperforming both standard DDPM and GAN-based generation approaches. Performance generally improved as the proportion of synthetic data in the training set increased, validating the utility of the generated samples. Further analysis using t-SNE visualizations confirms that the synthetic data occupies a similar manifold to the real data while maintaining diversity. Additionally, SHAP analysis provides interpretability, identifying the EEG channels and time segments most influential for the classification decisions.

In conclusion, this work successfully demonstrates that a noise-augmented conditional diffusion model can effectively generate realistic and diverse synthetic EEG data. This approach mitigates the data scarcity problem and leads to tangible improvements in emotion recognition accuracy, marking a significant contribution to the field of EEG-based affective computing and offering a robust alternative to existing generative methods.