Semi-supervised Learning for Dynamic Modeling of Brain Signals During Visual and Auditory Tests

摘要

Requirements of costly data labeling for data classification are relaxed with semi-supervised learning. This is particularly useful considering monitoring of a physiological process that continuously produces data and can be observed for a long time. We propose a new expectation-maximization (EM) procedure that implements semi-supervised learning and it is based on sequential independent component analysis modeling (SICAMM), that we have called EM-SICAMM. This procedure is applied for dynamic modeling of EEG signals measured from epileptic patients during visual and auditory neuropsychological tests. Those tests are done to evaluate the learning and memory cognitive function of the patients. Classification results demonstrate that EM-SICAMM outperforms, in terms of balanced error rate (BER) and kappa index, the following competitive methods: ICAMM, SICAMM, Gaussian mixture model (GMM), and hidden Markov model (HMM).