Baynesian estimation techniques for the extraction of event - related potentials in neuroscience
Costanza D’Avanzo · Padua Research Archive (University of Padova) · 2011
The study of the Event-Related Potentials (ERPs) represents a classic topic in neuroscience research. In fact, as discussed in Chapter 1 of this work, ERPs measured in response to sensory, cognitive and motor stimuli are crucial in the comprehension of many aspects of neurophysiology and, since their acquisition is relatively simple and non invasive, they also have several clinical applications. On the other hand, ERPs extraction is not a trivial question: ERPs signals are embedded in background spontaneous electroencephalographic (EEG) activity with much larger amplitude and common spectral content. The approach traditionally used for ERP extraction, the so-called Conventional Averaging (CA), presents well-known limitations. In the light of this, more sophisticated methodologies have been proposed in order to improve the average estimate and to provide a single-trial description of the ERPs; in Chapter 2 some of these techniques are described. In this thesis a Bayesian approach for the improved estimation of the average ERP and for the single-trial ERPs extraction is proposed. In particular, in Chapter 3 and 4 two new methods, with different degree of sophistication, are implemented. The first method is based on a two-stage procedure. In the first stage, an average ERP is determined as the weighted average of the available sweeps, previously processed by an individual optimal filter, in which a 2nd order a priori statistical information on the involved signals is exploited. In the second stage, single-sweep estimation is dealt with within the same framework, by using the average ERP estimated in the previous stage as a priori expected response. The second method is based on a one-stage multi task learning procedure. Differently from the most estimation approaches, the method provides the estimate of the average and the single-trial responses by processing just once all the available sweeps simultaneously. The unknown single-trial ERPs are treated as the “individuals” of a homogenous population and the information available for a sweep is considered informative with respect to the other ones. The method assumes that the generic sweep can be modeled as the sum of three independent stochastic processes: an average curve of population that is common to all the sweeps, an individual shift that differentiates each sweep from the others, a background EEG noise component varying sweep by sweep. Simulated datasets with different levels of signal-to-noise ratio (SNR) have been employed in order to test the performance of the proposed approaches in estimating the average ERP and the single-trial ERPs, also by varying the number of available sweeps. Results, discussed in Chapter 5, point out that the proposed approaches provide significantly better estimates of the average ERP with respect to the CA technique, for each of the tested SNR levels. In particular, with the new approaches, the number of sweeps needed for the average ERP estimation can be reduced of about 50 %. As far as the single-trial estimation is concerned, the proposed methods provide a significantly better reconstruction of the single-trial responses in comparison with a representative literature method, while results comparable with the above-mentioned method are obtained with regard to the estimation of latency and amplitude of P300 component. In Chapter 6, the proposed methods are applied to a real data set. This data set consists of EEG signals recorded on cirrhotic and normal subjects during a Simon task, a two-choice paradigm in which the subject is required to evaluate which stimulus, between the two possible target stimuli, appears on a monitor. In particular, the availability of the single trial P300 latencies and amplitudes has allowed to better understand the causes of the reduction in cirrhosis of the CA-based P300 amplitude and it has made possible the investigation of the relationship between the variability of P300 component and the variability of the behavioral measures.