Single-trial analysis of electroencephalography and functional magnetic resonance imaging for brain decoding

Abstract

Neuroimaging techniques have facilitated investigations into the mechanisms and functions of the human brain. To analyze neuroimaging signals elicited by external stimuli or cognitive tasks, across-trial averaging is conventionally performed to demonstrate differences across conditions or groups of subjects. However, there is a need in modern neuroscience to study the variability across trials and to reveal the trial-to-trial dynamic information encoded in the brain. In this dissertation, we develop three single-trial analysis approaches for decoding human brain from electroencephalography (EEG) and functional magnetic resonance imaging (fMRI). First, to improve the signal-to-noise ratio (SNR) of evoked EEG, a novel spatial-temporal-spectral filter combining common spatial pattern based spatial filter and continuous wavelet transform based temporal-spectral filter was proposed and applied to a visual evoked potential (VEP) based brain-computer interface (BCI) system. The proposed filter was shown to provide significant improvements in terms of SNR of single-trial VEPs and accuracy of the BCI system, and proved to be robust in fewer channel montages. Using the BCI system in a real-time application, normal subjects were able to make 20 decisions with an accuracy of 90% within 1 min. These results suggest that the proposed filter can be a promising single-trial detection approach with applications in various fields of neuroscience and clinical neurophysiology. Second, we proposed a multivariate linear regression (MVLR) model to describe the trial-to-trial relationship between EEG/fMRI data and behavior responses. Since the number of independent variables (time-frequency points of EEG; voxels of fMRI) is markedly larger than the number of dependent variables (experimental trials) and the nearby independent variables are strongly correlated, we estimated the parameters of the model by partial least square (PLS) regression. As a real world application, we used the proposed MVLR model and PLS regression to study the neural mechanism of how pre-stimulus brain activities modulated subsequent pain perception. Pre-stimulus EEG temporal-spectral patterns and fMRI spatial patterns which were predictive to pain were identified and extracted. Further, we combined predictive information from pre-stimulus and post-stimulus EEG/fMRI to establish a neurophysiology based pain prediction tool, which provided significantly better performance than conventional pain prediction approaches. Third, since there are more independent variables than experimental trials in an MVLR model for fMRI studies, it is advantageous to reduce the number of independent variables by selecting an informative subset. Conventional dimension reduction techniques often over-simplified the complex relationship between fMRI data and class labels. We introduced a novel fMRI decoding approach based on a dimension reduction technique, namely sliced inverse regression (SIR). SIR can exploit class information for estimating dimension reduction directions regardless of linear or nonlinear relationship between data and labels. Simulation results showed that the proposed approach can detect dimension reduction directions and predict class labels more accurately than conventional techniques. We applied the proposed approach to predict the level of pain perception from laser-evoked fMRI data and achieved higher accuracy in pain prediction than conventional techniques. These results suggest that the SIR-based fMRI decoding approach is effective for decoding behavioral, perceptual or cognitive states from single-trial fMRI data.