Feature learning for EEG-based person identification

Evoked potentials recorded on a multielectrode EEG device are known to be a ected by volume conductance and functional connectivity while a task is performed by a person. Modelling functional connectivity represents neural interactions between electrodes which are distinguishable and genetically identical. However, the representations that are caused by volume conductance are not distinguishable because of unwanted correlations of the signal. Orthogonalisation using autoregressive modelling minimises the conductance component, and the connectivity features can be then extracted from the residuals. The proposed method shows it is possible to reduce the multidimensionality of the predicted AR model coe cients by modelling theresidual from the EEG electrode channel baseline, which makes an important contribution to the functional connectivity. The results show that the required models can be learnt by Machine Learning techniques which are capable of providing the maximal performance in the case of multidimensional EEG data. The proposed method was able to learn accurate identificationwith few EEG recording channels, especially when the channel that is used has a functional connectivity with the interactive task. The study, which has been conducted on a EEG benchmark including 109 participants, shows a signi cant improvement of the identi cation accuracy.