Markov Switching Model for Quick Detection of Event Related Desynchronization in EEG

Giuseppe Lisi, Diletta Rivela, Asuka Takai, Jun Morimoto

Quick detection of motor intentions is critical in order to minimize the time required to activate a neuroprosthesis. We propose a Markov Switching Model (MSM) to achieve quick detection of an event related desynchronization (ERD) elicited by motor imagery (MI) and recorded by electroencephalography (EEG). Conventional brain computer interfaces (BCI) rely on sliding window classifiers in order to perform online continuous classification of the rest vs. MI classes. Based on this approach, the detection of abrupt changes in the sensorimotor power suffers from an intrinsic delay caused by the necessity of computing an estimate of variance across several tenths of a second. Here we propose to avoid explicitly computing the EEG signal variance, and estimate the ERD state directly from the voltage information, in order to reduce the detection latency. This is achieved by using a model suitable in situations characterized by abrupt changes of state, the MSM. In our implementation, the model takes the form of a Gaussian observation model whose variance is governed by two latent discrete states with Markovian dynamics. Its objective is to estimate the brain state (i.e., rest vs. ERD) given the EEG voltage, spatially filtered by common spatial pattern (CSP), as observation. The two variances associated with the two latent states are calibrated using the variance of the CSP projection during rest and MI, respectively. The transition matrix of the latent states is optimized by the "quickest detection" strategy that minimizes a cost function of detection latency and false positive rate. Data collected by a dry EEG system from 50 healthy subjects, was used to assess performance and compare the MSM with several logistic regression classifiers of different sliding window lengths. As a result, the MSM achieves a significantly better tradeoff between latency, false positive and true positive rates. The proposed model could be used to achieve a more reactive and stable control of a neuroprosthesis. This is a desirable property in BCI-based neurorehabilitation, where proprioceptive feedback is provided based on the patient's brain signal. Indeed, it is hypothesized that simultaneous contingent association between brain signals and proprioceptive feedback induces superior associative learning.