Hierarchical modeling of EEG signals

The electroencephalogram (EEG) is a complex electrical signal which reflects generalized brain activity. The EEG is utilized in the clinical assessment of many neurological and psychiatric disorders and offers promise for monitoring of patients undergoing anesthesia and operation. Automated analysis and classification of EEG signals is hampered by the unwieldy amounts of data which are generated by detailed descriptors of the nonstationary waveforms. This paper describes a technique for quantitative analysis of EEG signals which is based on a hierarchy of models. These models include 1) recursively estimated autoregressive model, 2) piecewise stationary autoregressive model, 3) composite source model, and 4) character string and syntactic models. This hierarchical modeling approach introduces criteria for segmentation, clustering, and identification of character substrings which reduce the need for ad hoc parameters or subjective decisions. The hierarchical representation is therefore highly operator independent, provides significant data compression of complex signals, and is compatible with a generalized classification procedure. Examples of the application of this modeling approach to clinical patient EEG data illustrate the system capabilities.