Improving analysis of brain electrical signals

This project is about the development and application of advanced information processing methods for the analysis of functional brain signals. The general aim will be the proper description of the spatial and temporal characteristics of functional brain signals and their functional coupling during the course of temporally extended cognitive processing. Event related potentials (ERPs) measurable in the electro encephalogram (EEG) during cognitive activities are used to monitor physiological correlates of human cognition. Most ERP studies focus on "pulse"-evoked short duration cognitive processes which guarantee the temporal stability of signal components that allows to simply average across a set of similar ERPs to improve the signal-to-noise ratio. However, temporally extended cognitive processes, which are much more interesting from a psychophysiological point of view, show a much higher variability in their temporal behaviour. We will develop new ways of analysis which take this into consideration. To properly describe spatial and temporal characteristics of cognitive EEG we will extend our earlier approach which is build on the hypothesis that only subsequences of fixed length can be expected to be similar across trials in long-lasting cognitive ERPs. We will extend our method of discovering and then averaging across such subsequences by allowing subsequences to additionally show considerable variation on the time axis, i.e. compression and expansion. This will be achieved by adapting the methodology of Hidden Markov Models (HMM), which have been applied successfully in speech processing and biological sequence analysis. Our project will pursue two additional closely related objectives dealing with the functional coupling of cortical areas: the description of the EEG signals in terms of their complexity and the causal analysis of the relation between signals from different locations on the scalp. This will be achieved by computing continuous measures of complexity and causation for each EEG trial. These new measures will also be based on information theory and yield time series of complexity and causation which will again be analysed using the newly developed HMMs. All achievements together will advance cognitive neuroscience by progressing from the question of "Where" to a functional order of "Where and When" cognition happens in the brain.

Cognitive Neuroscience developed on the basis of Event Related Brain Potentials (ERP) throughout the last forty years. Due to the averaging technique necessary for signal to noise ratio enhancement predominantly early cognitive processing stages have been analyzed and considered. Prerequisites for averaging, i.e. low latency jitter and sufficient waveform stability, may not be sufficiently fulfilled with brain electrical signal that accompany later stages of temporally extended cognition; this might partly generate so called `slow cortical potentials` (SCP) showing also low spatial resolution. Therefore we concentrated on work dealing with the discovery of reoccurring spatio-temporal patterns, first giving a survey of work related to our approach. Thereafter we were mainly concerned with the improvement of our already existing method for discovery of common patterns by means of K- means Clustering. Switching to the probabilistic method of Gaussian Mixture Models (GMM), in particular to a GMM with integrated noise component which had originally been developed for speech analysis to estimate signals hidden in background noise, we focused on de-noising single trial ERPs. Additionally we already applied Principal Component Analysis (PCA) as well as Independent Component Analysis (ICA) to the same data. These procedures worked sufficiently good with artificially generated ERP test data. With real ERP data results with GMM were somewhat satisfactory. However, the application of PCA did not really work out - presumably due to correlated noise across channels. While analysis of ERPs and SCPs ultimately results in just high resolution functional neuro-anatomy, investigations of Gamma Band Synchronization, a timely topic, would add, if successful, a new qualitative aspect to cognitive neuroscience, i.e. tagging temporally and spatially functional coupling of structures . However, all our attempts made, i.e. looking simply for increases of Gamma power or searching for zero phase differences in narrow band filtered signals observed during various cognitions, did not yield satisfactory results. As the most promising procedure for monitoring brain electrical potentials that accompany temporally extended cognition, so far, turned out the method of Independent Component Analysis (ICA). ICA is able to decompose a set of single-trial ERPs into spatially fixed but temporally independent components, components that correspond to various artifacts, stimulus- and response-locked as well as unlocked activity. This kind of decomposition enables artifact elimination, noise reduction and single-trial-wise further analysis of isolated component, e.g. identification of generating sources.