Automatic adaptive segmentation of clinical EEGs.

A method of automatic adaptive segmentation of EEGs, whereby the boundaries between different patterns of activity appearing in a given channel are identified and demarcated, has been applied to a group of clinical EEGs. The EEGs were especially selected from a library of 70 recordings so as to include a diversity of normal and abnormal EEG patterns. In the subsequent steps of the analysis, like segments were automatically clustered and a dendrogram plotted, from which the principal clusters or types of activity were visually selected. For the latter, a temporal profile was then plotted, indicating which type of activity was present at any given time during the respective recording. Summary parameters of mean amplitude and a measure of mean frequency were plotted alongside the temporal profile. The findings are as follows. (1) A single set of segmentation parameters was found to be clinically satisfactory for the entire group. (2) The inappropriateness of adaptive segmentation for the isolation of spikes and sharp waves, which had been anticipated in view of the short duration of such transients in relation to the length of the window (1.2 sec) used for the autocorrelation functions employed in the segmentation algorithm, was confirmed. A separate spike/sharp-wave detection algorithm is therefore planned. (3) Longer transients (i.e., some 300 msec or greater) are segmented and separately clustered. (4) For an individual EEG, the number of clinically significant clusters was 5 or less. (5) If a given cluster included activity in more than one frequency band (for example, simultaneous alpha and slow activity) the simple summary parameters of mean amplitude and a measure of mean frequency may not be sufficient for a clinically adequate description; however, the original autocorrelation function, or, more conventionally, the power spectrum (which can be estimated from the autocorrelation function) provides the necessary information. (6) The use of automatic adaptive EEG segmentation minimized human bias in the selection of portions of EEG recordings for computer analysis.