Heterogeneous data fusion for brain psychology applications

Abstract

This thesis aims to apply Empirical Mode Decomposition (EMD), Multiscale Entropy (MSE), and collaborative adaptive filters for the monitoring of different brain consciousness states. Both block based and online approaches are investigated, and a possible extension to the monitoring and identification of Electromyograph (EMG) states is provided. Firstly, EMD is employed as a multiscale time-frequency data driven tool to decompose a signal into a number of band-limited oscillatory components; its data driven nature makes EMD an ideal candidate for the analysis of nonlinear and non-stationary data. This methodology is further extended to process multichannel real world data, by making use of recent theoretical advances in complex and multivariate EMD. It is shown that this can be used to robustly measure higher order features in multichannel recordings to robustly indicate ‘QBD’. In the next stage, analysis is performed in an information theory setting on multiple scales in time, using MSE. This enables an insight into the complexity of real world recordings. The results of the MSE analysis and the corresponding statistical analysis show a clear difference in MSE between the patients in different brain consciousness states. Finally, an online method for the assessment of the underlying signal nature is studied. This method is based on a collaborative adaptive filtering approach, and is shown to be able to approximately quantify the degree of signal nonlinearity, sparsity, and non-circularity relative to the constituent subfilters. To further illustrate the usefulness of the proposed data driven multiscale signal processing methodology, the final case study considers a human-robot interface based on a multichannel EMG analysis. A preliminary analysis shows that the same methodology as that applied to the analysis of brain cognitive states gives robust and accurate results. The analysis, simulations, and the scope of applications presented suggest great potential of the proposed multiscale data processing framework for feature extraction in multichannel data analysis. Directions for future work include further development of real-time feature map approaches and their use across brain-computer and brain-machine interface applications.