2025 , Volume 30, № 5, p.108-122

This study addresses evaluation of the effectiveness for different deep learning approaches for generating synthetic electroencephalography (EEG) signals, which have significant applications in Оценка эффективности различных подходов к генерации... 121 neurophysiological research and medical diagnostics. Three models were analyzed: a generative adversarial network (GAN), a variational autoencoder (VAE), and a diffusion model based on a long short-term memory (LSTM) architecture. These models were trained on EEG signals recorded from the F7 channel of 28 participants engaged in cognitive tasks, focusing on replicating the temporal and structural patterns of real EEG data. Methodologically, the performance of each model was assessed using mean absolute error (MAE), mean squared error (MSE), and Pearson correlation metrics. The findings indicate that the LSTM-based diffusion model outperformed both GANand VAE, achieving the lowest MAE (0.3343) and MSE (0.1760), as well as higher correlation with real signals, highlighting its capability to generate realistic synthetic EEG data. The VAE showed moderate performance, with MAE and MSE of 0.4642 and 0.7319, respectively, while GAN demonstrated limited effectiveness with the highest errors and nearly zero correlation with real data. This study underscores the originality and value of employing diffusion models for EEG signal generation, as they offer enhanced stability and accuracy in replicating complex neural patterns. The results suggest that LSTM-based diffusion models have strong potential for advancing synthetic data applications in neurophysiology while enabling reliable data generation when real data are limited.