Unveiling Persistent Brain Fingerprints: Exploring Stable Power Spectral Features with Variational Autoencoder-Based Models

Abstract

This thesis investigates utilizing power spectral density (PSD) extracted from magnetoencephalography (MEG) signals and variational autoencoders (VAE) based models to generate distinct brain fingerprints at different time points for subjects not included in the training phase. The promising accuracy results (over 0.83) in both linear and non-linear models reveal the potential of PSD for subject identification tasks. We took advantage of the VAE models’ flexibility by slightly adjusting the VAE loss function, penalizing the distance between latent representations (brain fingerprints) of the same subject at two different time points. This adjustment led to superior subject identification performance compared to standard VAE [31] and beta-VAE [34] models. Furthermore, this modified VAE model demonstrates notable superiority over the ln-BRRR [73] model in this task. While the latter has demonstrated favorable results [7], our model surpasses it by a considerable margin while demanding considerably less training time. Employing a model interpretability technique based on Shapley values, the study identifies sets of frequency bands and brain areas relevant by the model in generating differentiable fingerprints. These findings highlight the complex interplay between diverse brain regions and frequency bands in subject differentiation.