Mining information from event-related recordings

In this article we describe a signal-processing framework for mining information from event-related recordings. Pattern-analytic tools are combined with graph-theoretic techniques and signal understanding methodologies in a user-friendly environment with the scope of learning, parameterization, and representation of the ST data manifold. Through the first part, we provide a general outline of our methodological approach while trying to demonstrate all the different stages, where DM tools can be applied. In the second part, we provide a more detailed demonstration, give a synopsis of the obtained results and take the opportunity to underline the merits of the adopted algorithmic procedures. To enable the full justification of our framework, instead of just including a technical demonstration of some of the incorporated DM and KDD tools, we address the problem of response variability: an issue of great neuroscientific importance and the subject of continuous debate. The major question in all the previous studies was the validity of "signal plus noise" model, i.e., whether a stereotyped evoked response is linearly superimposed on the ongoing brain activity after every stimulus presentation, a prerequisite for the validity of ensemble-averaging. Using data from a simple visual experiment targeting at the early neuromagnetic response known as N70m, we try to bridge the gap between the "conservative-party" that suggests heavy averaging as the only way to study the brain's response and the "neurodynamics-party" that claims the averaged-signal has very little to say about how the real-time processing of an input from a sensory pathway is actually performed in the cortex.