Intelligent single trial analysis of

event-related dynamics

The rich temporal information conveyed by the EEG/MEG signals is generated by widespread network of overlapping, often interacting neural sources. Some processes are sequential while others have parallel subprocesses, very often at very different timescales. It is therefore no surprise that individual single trial responses are characterized by high variability, over and above the variability introduced by environmental and/or physiological sources unrelated to the experiment. The remaining variability appears as changes in amplitude and latency or even as distinct categories of responses. It is increasing to realize that such contributions are often related to the physiological processes one wishes to study, or the way they couple to the background brain activity. 

We have developed an efficient framework for mining the information from the single trial recordings and organizing it in an intelligible way. Robust techniques, ranging from spatial filtering operators to tomographic estimates of regional brain activity, are first employed to extract the time courses of single trial activity corresponding to the brain region under study. The extracted temporal patterns are then analyzed, handled and classified within our framework which combines pattern analysis principles with a graph theoretic representation that enables the direct visualization of the employed steps and facilitates the interaction with the user.

One key idea relies on classical feature extraction steps that are then analyzed in a multidimensional space, the feature space, the geometry of which can systematically be explored by means of the Minimal Spanning Tree (MST) graph. This graph is further used for the derivation of a low-dimensional (2D) representation of the sample of single trial patterns. Embedded in a GUI environment the so-called reduced (2D) feature space constitutes the interface between the user and the overall recorded single trial brain activity. 

Different motifs of regional single trial activity can easily be recognized and analyzed further. This leads to a complete characterization of the response variations reflecting the dynamical changes within the specific brain region. In addition, possible associations between this brain region and others can readily be pursued as response covariation over trials and detected based on the similarity between the corresponding, independently constructed MST graphs. A variety of pattern analytic (e.g. clustering algorithms) and signal processing (e.g. wavelets) tools are incorporated within this flexible framework offering alternative, possibly, complementary ways to gain insight into their data.

Related key findings ...

For more technical details about these methods see Laskaris et al., 2001, 2002, 2003.