Abstract
Subject-specific narrowband oscillatory rhythms in human electroencephalogram (EEG)
can be detected using tensor decomposition. In our previous studies, we explored the CANDE-
COMP/PARAFAC (CP) decomposition together with the more flexible Tucker model. However ,
we found that CP decomposition sometimes required a high number of latent components to rep-
resent the data latent structure accurately. On the other hand, the Tucker model appeared too
generalized, leading to very sparse solutions. Therefore, in this study, we focus on the PARA-
LIND model, which combines the more straightforward interpretability of CP decomposition
with the flexibility o f t he Tucker m odel. We d emonstrate i ts p erformance o n E EG recordings
from a patient following an ischemic stroke, comparing it to the previous two models.
Go back