State-space modeling and estimation for multivariate brain signals

Abstract

Brain signals are derived from underlying dynamic processes and interactions between populations of neurons in the brain. These signals are typically measured from distinct regions, in the forms of multivariate time series signals and exhibit non-stationarity. To analyze these multi-dimensional data with the latent dynamics, efficient statistical methods are needed. Conventional analyses of brain signals use stationary techniques and focus on analyzing a single dimensional signal, without taking into consideration the coherence between signals. Other conventional model is the discrete-state hidden Markov model (HMM) where the evolution of hidden states is characterized by a discrete Markov chain. These limitations can be overcome by modeling the signals using state-space model (SSM), that model the signals continuously and further estimate the interdependence between the brain signals. This thesis developed SSM based formulations for autoregressive models to estimate the underlying dynamics of brain activity based on measured signals from different regions. The hidden state and model estimations were performed by Kalman filter and maximum likelihood estimation based on the expectation maximization (EM) algorithm. Adaptive dynamic model time-varying autoregressive (TV-AR) was formulated into SSM, for the application of multi-channel electroencephalography (EEG) classification, where accuracy obtained was better than the conventional HMM. This research generalized the TV-AR to multivariate model to capture the dynamic integration of brain signals. Dynamic multivariate time-varying vector autoregressive (TV-VAR) model was used to investigate the dynamics of causal effects of one region has on another, which is known as effective connectivity. This model was applied to motor-imagery EEG and motortask functional magnetic resonance imaging (fMRI) data, where the results showed that the effective connectivity changes over time. These changing connectivity structures were found to reflect the behavior of underlying brain states. To detect the state-related change of brain activities based on effective connectivity, this thesis further developed a novel unified framework based on the switching vector autoregressive (SVAR) model. The framework was applied to simulation signals, epileptic EEG and motor-task fMRI. The results showed that the novel framework is able to simultaneously capture both slow and abrupt changes of effective connectivity according to the brain states. In conclusion, the developed SSM based approaches were effective for modeling the nonstationarity and connectivity in brain signals.