Space-Time Biosignal Processing Interference Mitigation, Feature Extraction, Source Localization and Brain Connectivity Analysis

Biosignal processing reveals information from measurements of physiological processes, and helps people understand the underlying mechanisms of these processes make medical diagnoses or evaluate therapies. However, conventional techniques often focus only on certain aspects of the signals and not all information contained in the signals is fully explored. This is especially true nowadays given recent advances in technology that allow more sensors to be deployed for signal acquisition at unprecedented resolutions in both space and time. While this increases the information content about the process at hand, it also creates challenges in how these multi-sensor recordings should be processed.;Signal processing performed in both the spatial and temporal domains has received considerable attention for many years in many different applications, since it allows for the rejection of interference, enhancement of signals of interest, detection of the appearance of signals, estimation of model parameters, and so on. There are many advantages to considering space-time signal processing in biomedical applications. To explore the capability of space-time techniques in biosignal processing, several challenges are explored in this dissertation, including interference mitigation, feature extraction, source localization and reconstruction, and brain connectivity analysis associated with the processing of the measurements made with EEG, MEG, or direct electrode insertion.;Some of the advantages of the techniques proposed in this dissertation are summarized below. Due to the elimination of the assumption of temporal stationarity, the proposed deterministic alternative to the standard prewhitening approach for interference suppression in EEG source localization is significantly more robust than prewhitening. The matched subspace algorithm for extracting discriminant features from multi-sensor measurements of extracellular APs is suitable for unsupervised AP sorting applications, and it outperforms existing popular feature extraction approaches. For source localization and reconstruction, both parametric and imaging methods are proposed for dealing with the localization of highly correlated sources, and these techniques are also able to mitigate the influence of the residual-source interference and the intrinsic bias. Finally, several algorithms for parameter estimation in dynamic causal modeling, calculation of the Cramer-Rao performance bounds for these estimates, and comparison of the accuracy of the algorithms against the theoretical performance limits under a variety of circumstances are discussed.