A Kalman filter procedure for the processing of the electroencephalogram

In the present paper a Kalman filter procedure is illustrated for the reduction of muscular noise superimposed to the electroencephalografic traces (EEG). Such a noise, in fact, has a bandwidth which overlaps the signal carrying the information content useful for the clinical standpoint and, therefore, can not be removed by means of classical digital filtering. A Markov model is used for identifying the signal model (supposed generated by an ARMA process) and the noise model (conceived on the basis of experiments of neurophysiological evidence). The experimental results show a good performance of the filter on the discrete-time EEG signal which is also quantified by the spectral information and the values of the prediction error of the filter itself. Comparison is then carried on with a classical low-pass FIR filter (mostly used in practice) which can not be aggressive enough towards the noise contained in the signal bandwidth but which can undoubtly ameliorate the performance of the Kalman filter.