Modelos de predição linear para análise de sinais eletroencefalográficos (EEG) e de matrizes multieletrodo (MEA)

Abstract

This work establishes models of neurophysiological signals, which are composed of spontaneous activity measurements taken by means of multielectrode arrays (MEAs) applied on in vitro cell cultures; as well as of neurological signals based on electroencephalography. These models suppose that MEAs are employed as neuroprostheses applied for detection and forecast of epileptic seizures, based on EEG signals or on invasive measurements which are taken in a cellular level. From this point of view, the signal processing tools must fulfil a problematic trade-off involving low computational complexity and real-time operation. Such requirements lead to the choice of auto-regressive adaptive-linear filtering and high-order statistics (HOE) as the techniques to be used in order to cope with, respectively, non-stationary signals and nonlinear systems. Linear prediction of both signals is quite efficient, particularly in the case of MEA signals, for which the model is stable and accurate. On the other hand, the convergence times for EEG signals are lower then their respective counterparts for MEA signals, which may be considered mainly non-Gaussian and correlated. Cyclic activity was also observed for MEA signals associated with neighboring electrodes, whereas signals recorded from small groups of neurons present a white-noise behaviour.