Yksilöntunnistus pidemmän aikavälin EEG-datasta bayesilaisella matalan asteen regressiomallinnuksella

In human neuroimaging data, individual variability is often interpreted as noise. However, such variability can be leveraged, for example, to explain differences in cognitive functioning, to enable personalized diagnostic predictions, or to identify other brain recordings belonging to the same individual. Electroencephalography (EEG) data are well-suited for individual identification due to their high temporal resolution and ability to capture complex neural dynamics. However, EEG signals are often noisy and high-dimensional. The latent-noise Bayesian reduced rank regression (ln-BRRR) model is able to create distinctive lower-dimensional representations, often referred to as cortical fingerprints, from complex EEG data, as it combines dimensionality reduction with effective noise-handling strategies. This thesis investigates the performance of EEG-based cortical ln-BRRR fingerprints for individual identification and examines the impact of age on fingerprints and identification performance. The fingerprints were created based on two metrics derived from the EEG data: power spectral density and phase locking value (PLV). The identification accuracy of a total of 173 subjects was estimated using a cosine distance measure and two different settings. In within-session identification, fingerprints from two recordings acquired during the same session were compared, whereas in between-session identification, the temporal stability of ln-BRRR fingerprints was evaluated by comparing fingerprints from two recordings acquired five years apart. Furthermore, the performance of ln-BRRR was compared with two alternative models commonly used for dimensionality reduction: principal component analysis (PCA) and linear discriminant analysis (LDA). Within session, ln-BRRR fingerprints achieved a maximum identification accuracy of 86.3 \% using PLV. They outperformed PCA fingerprints and performed similarly to full-connectome PLV data, while LDA fingerprints achieved the highest overall performance. In between-session identification, while the performance of all methods decreased, the ln-BRRR fingerprints performed substantially better than full-connectome PLV data and also outperformed PCA and LDA fingerprints. Age was found to have only a weak effect on ln-BRRR fingerprints and no direct influence on subject identifiability. The results indicate good temporal stability of ln-BRRR fingerprints.