Michael Grözinger

An E-Health Solution for Automatic Sleep Classification according to Rechtschaffen and Kales: Validation Study of the Somnolyzer 24 × 7 Utilizing the Siesta Database

To date, the only standard for the classification of sleep-EEG recordings that has found worldwide acceptance are the rules published in 1968 by Rechtschaffen and Kales. Even though several attempts have been made to automate the classification process, so far no method has been published that has proven its validity in a study including a sufficiently large number of controls and patients of all adult age ranges. The present paper describes the development and optimization of an automatic classification system that is based on one central EEG channel, two EOG channels and one chin EMG channel. It adheres to the decision rules for visual scoring as closely as possible and includes a structured quality control procedure by a human expert. The final system (Somnolyzer 24 × 7™) consists of a raw data quality check, a feature extraction algorithm (density and intensity of sleep/wake-related patterns such as sleep spindles, delta waves, SEMs and REMs), a feature matrix plausibility check, a classifier designed as an expert system, a rule-based smoothing procedure for the start and the end of stages REM, and finally a statistical comparison to age- and sex-matched normal healthy controls (Siesta Spot Report™). The expert system considers different prior probabilities of stage changes depending on the preceding sleep stage, the occurrence of a movement arousal and the position of the epoch within the NREM/REM sleep cycles. Moreover, results obtained with and without using the chin EMG signal are combined. The Siesta polysomnographic database (590 recordings in both normal healthy subjects aged 20–95 years and patients suffering from organic or nonorganic sleep disorders) was split into two halves, which were randomly assigned to a training and a validation set, respectively. The final validation revealed an overall epoch-by-epoch agreement of 80% (Cohen’s kappa: 0.72) between the Somnolyzer 24 × 7 and the human expert scoring, as compared with an inter-rater reliability of 77% (Cohen’s kappa: 0.68) between two human experts scoring the same dataset. Two Somnolyzer 24 × 7 analyses (including a structured quality control by two human experts) revealed an inter-rater reliability close to 1 (Cohen’s kappa: 0.991), which confirmed that the variability induced by the quality control procedure, whereby approximately 1% of the epochs (in 9.5% of the recordings) are changed, can definitely be neglected. Thus, the validation study proved the high reliability and validity of the Somnolyzer 24 × 7 and demonstrated its applicability in clinical routine and sleep studies.

Increased bioavailability of oral melatonin after fluvoxamine coadministration

Fluvoxamine, a selective serotonin reuptake inhibitor, is known to elevate melatonin serum concentrations. It has not been clear whether these effects might be attributed to an increased melatonin production or to an decreased elimination of melatonin. The latter hypothesis was tested by this study.

Human sleep EEG under the influence of pulsed radio frequency electromagnetic fields. Results from polysomnographies using submaximal high power flux densities

Former exploratory investigations of sleep alterations due to global system for mobile communications (GSM) signals have shown a hypnotic and REM-suppressive effect under field exposure. This effect was observed in a first study using a power flux density of 0.5 W/m2, and the same trend occurred in a second study with a power flux density of 0.2 W/m2. For the present study, we applied a submaximal power flux density of 50 W/m2. To investigate putative effects of radio frequency electromagnetic fields (EMFs) of cellular GSM phones on human sleep EEG pattern, all-night polysomnographies of 20 healthy male subjects both with and without exposure to a circularly polarized EMF (900 MHz, pulsed with a frequency of 217 Hz, pulse duration 577 μs) were recorded. The results showed no significant effect of the field application either on conventional sleep parameters or on sleep EEG power spectra.

Effects of REM sleep awakenings and related wakening paradigms on the ultradian sleep cycle and the symptoms in depression

In 1975 Vogel and coworkers published their classical study where they compared selective rapid eye movement (REM) sleep deprivation by brief awakenings to a control intervention paradigm in depressed patients. The superior antidepressive impact of the first procedure was attributed to the REM pressure accumulating during the treatment period. The laborious procedure and the considerable effort necessary to evaluate the sleep profiles in real time have prevented similar experiments so far. Based on artificial neural networks we developed a software for the real time detection of REM sleep. In combination with an alarm system the algorithm allowed us to wake up subjects automatically and to reduce REM sleep by about 50%. The procedure was then compared to a modified nonREM intervention paradigm for a treatment period of ten consecutive nights in depressed patients (n(1)=14, n(2)=13). These simultaneously received moderate dosages of Trimipramine. We found a strong and robust but not significantly different reduction of the average Hamilton rating scores (33 and 41% of baseline levels). While the REM sleep awakenings shortened the sleep cycle duration considerably, our nonREM intervention paradigm lengthened the ultradian alternations. Both effects might be interpreted as a challenge imposed on the nonREM-REM alternating mechanism possibly responsible for the antidepressive impact. A different timing of the control interventions might have caused the discrepancy between our findings and those of Vogel and coworkers.

Automatic recognition of rapid eye movement (REM) sleep by artificial neural networks

Artificial neural networks are well known for their good performance in pattern recognition. Their suitability for detecting REM sleep periods on the basis of preprocessed EEG data in humans under clinical conditions was tested and their performance compared with the manual evaluation. A single channel of the EEG signal was analysed in time periods of 20 s and preprocessed into a vector of six real numbers, which served as input to the network. EOG and EMG information was ignored. Backpropagation was used as a learning rule for the network, which consisted of 12 neurons and 39 synapses. Training datasets were put together from the input vectors and the corresponding sleep stages were scored manually. In working mode different networks were compared in terms of the rate of misclassified time periods for data not belonging to the training sets. The indicator function of REM sleep was well approximated by the network output in the course of the night, which was especially true for REM onsets. The average rate of correctly classified time periods was 89%. The errors were analysed and suggestions for improvements developed.

A replication study on P300 single trial analysis in schizophrenia: confirmation of a reduced number of ‘true positive’ P300 waves

A single trial analysis of event-related potentials (auditory odd-ball paradigm) of 20 schizophrenics was performed in comparison to matched healthy controls. The purpose of the study was to test the hypothesis that in schizophrenia the well-known P300 amplitude reduction of averaged event-related potentials is due to fewer elicited single trial P300 waves. The results of the present study support this finding of our previous exploratory investigation and point to the view that schizophrenics reveal basal disturbances in information processing due to inadequately elicited electrophysiological responses to target stimuli.

Online detection of rem sleep based on the comprehensive evaluation of short adjacent eeg segments by artificial neural networks

1. For scientific and clinical requirements the present objective is a robust automatic online algorithm to detect rapid eye movement (REM) sleep from single channel sleep EEG data without using EMG or EOG information. 2. For data preprocessing 20 seconds time periods of the continuous EEG activity are digitally filtered in 7 frequency bands. Then the RMS values of these filtered signals are calculated along segments of 2.5 seconds. The resulting matrix of RMS values is representing information on the power of the signal localized in time and frequency and serves as input to an artificial neural network. A pooled set of EEG data together with the corresponding manual evaluation of the recordings was used in the training process. 3. Afterwards more than 90% of the time periods not belonging to the training set could be correctly labeled into REM and nonREM periods. In comparison to an older algorithm based on RMS values calculated along segments of 20 seconds, the error rate could be reduced by 20%.

Recognition of rapid-eye-movement sleep from single-channel EEG data by artificial neural networks : A study in depressive patients with and without amitriptyline treatment

An automatic procedure for the online recognition of REM sleep appears to be a necessary tool for selective REM sleep deprivation in depressive patients. To develop such a procedure we applied an artificial neural network to preprocessed single-channel EEG activity. EOG and EMG information was purposely not provided as input to the network. A generalized back-propagation algorithm was used for computer simulation. The sleep profile scored manually according to Rechtschaffen and Kales served as the desired output during the training period and as standard for the judgement of the network output during working mode. Polysomnographic recordings from 5 healthy subjects were pooled to train the network, whereas second-night EEG recordings from the same subjects were used as independent working data sets. We further applied the network to the data of 5 depressive patients without medication and 6 depressive patients treated with amitriptyline. For these groups between 84.9 and 88.6% out of all time periods consisting of 20 s of continuous EEG activity were correctly classified. The indicator function of REM sleep was well approximated by the network output in the course of the night. Especially the REM onset was excellently recognized. The inclusion of patient data in the training set yielded a different network, which was evaluated and compared.

Neural net classification of REM sleep based on spectral measures as compared to nonlinear measures.

Abstract. In various studies the implementation of nonlinear and nonconventional measures has significantly improved EEG (electroencephalogram) analyses as compared to using conventional parameters alone. A neural network algorithm well approved in our laboratory for the automatic recognition of rapid eye movement (REM) sleep was investigated in this regard. Originally based on a broad range of spectral power inputs, we additionally supplied the nonlinear measures of the largest Lyapunov exponent and correlation dimension as well as the nonconventional stochastic measures of spectral entropy and entropy of amplitudes. No improvement in the detection of REM sleep could be achieved by the inclusion of the new measures. The accuracy of the classification was significantly worse, however, when supplied with these variables alone. In view of results demonstrating the efficiency of nonconventional measures in EEG analysis, the benefit appears to depend on the nature of the problem.

Alterations of continuous MEG measures during mental activities.

In a pilot study, we investigated the topography of 11 continuous MEG measures for the eyes-opened and eyes-closed condition together with three simple mental tasks (mental arithmetic, visual imagery, word generation). One-minute recordings for each condition from 16 right-handed subjects were analyzed. The electrophysiological measures consisted of 6 spectral band measures together with spectral edge frequency and spectral entropy, plus the time-domain-based entropy of amplitudes (ENA) and the nonlinear measures correlation dimension D2 and Lyapunov exponent L1. In summary, our results indicate a pronounced task-dependent difference between the anterior and the posterior region, but no lateralization effects. Although the nonlinear measures ranged in the middle field with respect to the number of significant contrasts, they were the only ones to be partially successful in discriminating the mental tasks from each other. The most efficient measure turned out to be the ENA. Under mental activation the ENA was larger than in both no task conditions (eyes opened and eyes closed). This finding reflects lower variations of the maximum amplitude during performance of mental tasks than during no task states.