Piecewise analysis of EEGs using AR-modeling and clustering

✍ Scribed by Ben H. Jansen; Arie Hasman; Roel Lenten

• Publisher: Elsevier Science
• Year: 1981
• Tongue: English
• Weight: 724 KB
• Volume: 14
• Category: Article
• ISSN: 0010-4809
• DOI: 10.1016/0010-4809(81)90034-3

✦ Synopsis

In this paper, a method is described to evaluate the EEG by means of a piecewise analysis; i.e., divide the EEG in short, stationary intervals, extract features and lump identical intervals together. The procedure involves the recursive computation of a fifthorder autoregressive model by means of a Kalman filter. The coefficients are averaged for 1.28-set intervals and used, together with the amplitude range, in a clustering procedure. As a result of this clustering, elementary patterns are determined. These patterns represent the "letters" of an "alphabet" that can generate an EEG. The letter statistics (or classification profiles), expressing the number of times each pattern occurs per EEG, are used in a second analysis to determine the EEG classification category. The results of applying this method to sleep recordings are described in this paper. About 80% agreement with visual classification was obtained using one recording for training and five other EEGs (four of which were recorded from two other subjects) for testing. These results indicate that the method is useful in extracting elementary patterns from an EEG and that the piecewise analysis approach is feasible.