Classification of EEG signals based on mean-square error optimal time-frequency features

This paper illustrates the improvement in accuracy of classification for electroencephalogram (EEG) signals measured during a memory encoding task, by using features based on a mean square error (MSE) optimal time-frequency estimator. The EEG signals are modelled as Locally Stationary Processes, based on the modulation in time of an ordinary stationary covariance function. After estimating the model parameters, we compute the MSE optimal kernel for the estimation of the Wigner-Ville spectrum. We present a simulation study to evaluate the performance of the derived optimal spectral estimator, compared to the single windowed Hanning spectrogram and the Welch spectrogram. Further, the estimation procedure is applied to the measured EEG and the time-frequency features extracted from the spectral estimates are used to feed a neural network classifier. Consistent improvement in classification accuracy is obtained by using the features from the proposed estimator, compared to the use of existing methods.