Toward Generating Physiologically Plausible Artificial Electroencephalography Data: Effects of Convolution-based Upsampling Methods on Data Quality

Background: Electroencephalography (EEG) is key for clinical and cognitive research, yet limited EEG data availability restricts deep learning (DL) applications. Generative Adversarial Networks (GAN) based on Convolutional Neural Networks (CNN) can augment EEG datasets. However, the upsampling methods used in CNN-based GANs frequently introduce artifacts into the generated EEG data. Transposed convolutions are known to generate systematic high-frequency noise, and their alternative, interpolation followed by convolutions, often leads to degraded amplitudes. The commonly employed evaluation frameworks do not allow a detailed comparison of the data generated by different models and upsampling methods. Therefore, this study (1) compares the different upsampling methods used in CNN-based GANs, and (2) introduces a novel evaluation framework for generated EEG data. Methods: This study employed three models to evaluate different upsampling methods: using only transposed convolutions, only interpolation-and-convolution, and alternating between the two. The models were trained on EEG data collected during a mind-wandering task. We compared EEG-specific linear and nonlinear features in the time, frequency, and spatial domains extracted from each model's generated EEG data to the respective target feature distributions using bootstrapped KS tests and Wasserstein Distance. Results: The mixed upsampling method generated the best results on linear features in the time and frequency domain. In the spatial domain, transposed convolutions marginally outperformed the other methods. The interpolation-and-convolutions method generally performed best on nonlinear features, most likely due to the high-frequency noise introduced by transposed convolutions used in the other two methods. Conclusion: Our EEG-specific evaluation framework revealed distinct strengths and weaknesses in the upsampling methods used in CNN-based GANs. These findings suggest that the choice of upsampling method should depend on the intended use case of the generated data, and they demonstrate the advantages of our framework over commonly used evaluation approaches.