Automated Sleep Stage Classification

Objectives: We have developed a system to perform automated sleep staging. The system is a re implementation of the sleep staging method originally developed by Jansen in late 70’s. We extended this system by incorporating transient detection techniques. Method: The basic method divides the EEG into short intervals (about 1 s) from which autoregressive (AR) model coefficient are extracted. The AR coefficients serve as features in a clustering process to establish a library of elementary patterns. These are used to classify the 1 s EEG segments. Next, histograms showing how frequently each elementary pattern occurs in 30 s interval are obtained, and used for sleep staging. Special transient detection rules are then applied to detect spindles and rapid eye movements. Results: The method was developed using nine nights of sleep from three subjects and tested on sleep from seven other subjects. An overall classification rate of around 80 % was obtained. We observed that the method for spindle detection did not work as expected, but the REM detection method worked satisfactorily and it increased the classification percentage of stage 1 and REM. Conclusion: Acceptable sleep staging results were obtained, but improvement may be possible using better spindle and K-complex detection.