On-line classification and segmentation of the electro-encephalogram signal using an adaptive lattice predictor

This paper proposes an adaptive approach, using the least-mean-square lattice (LMSL) predictor, to classification, segmentation and tracking of the electro-encephalogram (EEG) signal. The LMSL approach is an all-zero lattice predictor consisting of cascaded similar first-order sections whose coefficients are updated in the least-mean square (LMS) sense. These predictors are independent due to the orthogonality principle linked to the LMS algorithm. Therefore, on-line adding a new section (i.e., increasing the predictor model order by one) has no effect on the coefficients of the preceding sections. In the proposed approach, the time-trajectories of the reflection coefficients of the all-zero lattice predictor as well as the on-line power spectrum estimates are employed as classification, segmentation and tracking parameters. Because of section independence an additional parameter can be used when needed for improving the classification and segmentation accuracy. Results of computer generated and real-world EEG data are provided to show the significant usefulness of the proposed approach.