In this project, we have continued to focus our efforts on the automated detection of state changes and arousals from sleep. We have refined the method that was described in the previous year's report. This method is based on on-line estimation of the underlying probability density functions representing sleep and arousal. Rather than using the high-frequency power (alpha + beta) of the EEG, the method now uses the high-frequency power to low-frequency power ratio (HLR) as the main index for determining the presence or absence of arousal. In addition, the instantaneous total EEG power is also compared to the 30-second moving average of total EEG power in order to determine the presence of large transient increases in EEG amplitude. Thus far, the algorithm has been tested primarily against sleep data obtained from normal adults and adult patients with sleep apnea. A parallel approach that we have taken has been to employ a fuzzy-logic-based algorithm for classifying sleep-wake state. The algorithm was developed in the Matlab environment. The EEG, ECG, EOG and respiration signals recorded from premature infants were first processed on an epoch-to-epoch (30 secs) basis to extract the following variables for input into the algorithm: delta EEG relative power, theta EEG relative power, alpha+beta EEG relative power, total EEG power, mean respiratory frequency, dispersion of respiratory spectral power about the mean frequency and EOG variance. The output variable was sleep-wake state, divided into 3 categories: quiet sleep, active sleep (equivalent to rapid eye movement sleep in adults) and wakefulness. Gaussian membership functions were employed in the fuzzification and defuzzification stages of the algorithm.
Showing the most recent 10 out of 84 publications
