Employing WGAN-GP for Synthesizing Biophysical Data: Generating Synthetic EEG for Concentration and Relaxation Level Prediction

This work introduces a novel approach utilizing Wasserstein Generative Adversarial Networks with Gradient Penalty (WGAN-GP) to generate synthetic EEG waves corresponding to mental states: concentration and relaxation. Our approach addresses the challenge of limited biophysical data by producing realistic synthetic brainwaves learned from real data, which is crucial for studies hindered by volunteer availability and privacy concerns. The classification accuracy improved from 92% using only the real dataset to 98.45% when combining real and synthetic data, demonstrating that our model’s synthetic data significantly enhances machine learning model training by expanding the dataset, thereby improving generalization and performance. The proposed WGAN-GP model also accurately generated EEG data, achieving optimum accuracy for synthetic concentration and 96.84% accuracy for synthetic relaxation data when classified by a Convolutional Neural Network (CNN). Techniques like discrete wavelet transform (DWT), downsampling, inverse DWT, and upsampling were used in this approach. These results underscore our model’s effectiveness in capturing key brainwave components-alpha, beta, theta, gamma, and delta. This method showcases significant potential to alleviate EEG data scarcity in biomedical applications through advanced generative modelling. Furthermore, this kind of application will significantly enhance human-robot interaction, enabling more responsive and adaptive systems in areas such as assistive technologies, mental health monitoring, and cognitive load assessment and other tasks where human engagement is needed.