Feature selection for sleep/wake stages classification using data driven methods

This paper focuses on the problem of selecting relevant features extracted from human polysomnographic (PSG) signals to perform accurate sleep/wake stages classification. Extraction of various features from the electroencephalogram (EEG), the electro-oculogram (EOG) and the electromyogram (EMG) processed in the frequency and time domains was achieved using a database of 47 night sleep recordings obtained from healthy adults in laboratory settings. Multiple iterative feature selection and supervised classification methods were applied together with a systematic statistical assessment of the classification performances. Our results show that using a simple set of features such as relative EEG powers in five frequency bands yields an agreement of 71% with the whole database classification of two human experts. These performances are within the range of existing classification systems. The addition of features extracted from the EOG and EMG signals makes it possible to reach about 80% of agreement with the expert classification. The most significant improvement on classification accuracy is obtained on NREM sleep stage I, a stage of transition between sleep and wakefulness.     

A model-based monitor of human sleep stages

Stochastic models are proposed for sleep and for the sleep related electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG). The evolution of sleep through its various stages is described as a Markov chain. The EEG is modelled using Wiener processes. The EOG and EMG are modelled as combinations of Poisson point processes and Gaussian processes, respectively. The EEG models contain a feedback structure that is based on physiological data. The maximum likelihood sleep stage monitor, that uses the sleep-related observations, has been derived and implemented. The agreement between automatic and human stage classifications of six sleep recordings was 70.6%, which was 4.5% worse than the average agreement between six human classifiers. Monitoring of simulated sleep suggests that the difficulty in separating wakefulness from stage 1 is due to poor modelling. If one ignores this difference, which, from a diagnostic point of view is fairly unimportant, the above mentioned agreement reaches 81.8%, which is 0.5% better than the corresponding average human vs human agreement.     

Spindle activity during sleep in subjects with a deficit in pallesthetic sensibility

Sleep spindle activity was studied on four subjects affected by pallesthesic deficit due to injury the posterior funiculi of the cord. The spindle activity was studied as density (number of spindles min.), duration and percent of sleep time utilized in spindle activity. The polygraph sleep records included EOG, EMG and 6 monopolar EEG recordings, 3 for each side, on the frontal, parietal and occipital regions. The records showed a spindle activity which was similar in different subjects and that was significantly higher than the physiological values. In fact, the spindle density was about 250%, the duration was about 130% and the spindle percent was about 280%, with respect to the physiological values assumed to be 100%.     

Midlatency auditory evoked responses: Differential effects of sleep in the human

Middle latency responses (MLRs) in the 10–100 msec latency range, evoked by click stimuli, were studied in 14 adult volunteer subjects during sleep-wakefulness to determine whether such changes in state were reflected by any MLR component. Evoked potentials were collected in 500 trial averages during continuos presentation of 1/sec clicks during initial awake recordings and thereafter during a 2 h afternoon nap or all-night sleep session. Continuously recorded EEG, EOG and EMG were scored for wakefulness, stages 2–4 of slow wave sleep (SWS), and rapid eye movement (REM) sleep during each evoked potential epoch. The major components included in this study and their latency ranges, as determined by peak latency measurements from the awake records, were: ABR V, 5–8 msec, Pa, 30–40 msec, Nb, 45–55 msec, and P1, 55–80 msec. In agreement with previous reports, ABR V and Pa showed no amplitude changes from wakefulness to either SWS or REM. Not previously reported, however, was the dramatic decrease and disappearance of P1 during SWS and its reappearance during REM to an amplitude similar to that during wakefulness. This unique linkage between a particular evoked potential component and sleep-wakefulness indicates that its generator system must be functionally related to states of arousal. Relevant data from the cat model suggest that the generator substrate for P1 may be within the ascending reticular activating system.     

Changes in sleep quality of athletes under normobaric hypoxia equivalent to 2,000-m altitude: a polysomnographic study.

This study evaluated the sleep quality of athletes in normobaric hypoxia at a simulated altitude of 2,000 m. Eight male athletes slept in normoxic condition (NC) and hypoxic conditions equivalent to those at 2,000-m altitude (HC). Polysomnographic recordings of sleep included the electroencephalogram (EEG), electrooculogram, chin surface electromyogram, and electrocardiogram. Thoracic and abdominal motion, nasal and oral airflow, and arterial blood oxygen saturation (Sa(O(2))) were also recorded. Standard visual sleep stage scoring and fast Fourier transformation analyses of the EEG were performed on 30-s epochs. Subjective sleepiness and urinary catecholamines were also monitored. Mean Sa(O(2)) decreased and respiratory disturbances increased with HC. The increase in respiratory disturbances was significant, but the increase was small and subclinical. The duration of slow-wave sleep (stage 3 and 4) and total delta power (<3 Hz) of the all-night non-rapid eye movement sleep EEG decreased for HC compared with NC. Subjective sleepiness and amounts of urinary catecholamines did not differ between the conditions. These results indicate that acute exposure to normobaric hypoxia equivalent to that at 2,000-m altitude decreased slow-wave sleep in athletes, but it did not change subjective sleepiness or amounts of urinary catecholamines.     

Sleep and EEG spectra in the pigeon (Columba livia) under baseline conditions and after sleep deprivation

Summary Sleep in adult domestic pigeons was studied by continuous 24-h recording of the EEG, EMG and EOG. Vigilance states were scored on the basis of behavioral observations, visual scoring of the polygraph records, and EEG power spectra.The animals showed a clear nocturnal preference for sleep. Throughout the dark period, EEG slow-wave activity was at a uniform level, whereas REM sleep (REMS) showed an increasing trend.EEG power density values differed significantly between the vigilance states. In general the values were highest in nonREM sleep (NREMS), intermediate in waking (W) and lowest in REMS.Twenty-four hour sleep deprivation reduced W and increased REMS, effects that are well documented in mammals. Unlike in mammals, EEG slow-wave activity remained unchanged, whereas EOG activity in W and NREMS was enhanced.Abbreviations EEG electroencephalogram - EMG electromyogram - EOG electrooculogram - SD sleep deprivation - L light - D dark - LD light dark - NREMS non rapid eye movement sleep - REMS REM sleep     

Behavioral and electrophysiological patterns of wakefulness-sleep states in a lizard.

Four individuals of the lizard Ctenosaura pectinata were chronically implanted for electroencephalographic (EEG), electromyographic (EMG) and electro-oculographic (EOG) recordings. Four different vigilance states were observed throughout the nyctohemeral cycle. These states were: Active wakefulness (Aw), quiet wakefulness (Qw), quiet sleep (Qs) and active sleep (As). Each state displayed its own behavioral and electrophysiological characteristics. EEG waves were similar during Aw and Qw but they diminished in amplitude and frequency when passing from these states to Qs, and both parameters increased during As. Muscular activity was intense in Aw, it decreased during Qw and almost disappeared during Qs. This activity reappeared in a phasic way during As, coinciding with generalized motor manifestations. Ocular activity was intense during Aw but minimal during Qw, it disappeared in Qs and was present intermittently in As. Aw, Qw, Qs and As occupied 5.9%, 25.7%, 67.7% and 0.6% of the 24 hr period, respectively. The frequency and duration of As episodes showed great inter-animal variability and the mean duration was of 12.9 sec. Stimuli reaction threshold was highest during sleep. In conclusion, the lizard Ctenosaura pectinata exhibit two sleep phases (Qs and As) that may be assimilated to slow wave sleep (SWS) and paradoxical sleep (PS) of birds and mammals.     

Midlatency auditory evoked responses: Differential effects of sleep in the cat

Middle latency responses (MLRs) in the 10–100 msec latency range, evoked by click stimuli, were studied in 8 adult cats during sleep-wakefulness to determine whether such changes in state were reflected by any MLR component. In particular, we wanted to determine whether the 20–22 msec positivity recorded at the vertex, ‘wave A,’ shown in previous studies to reflect a generator substrate within the ascending reticular formation, was tightly linked to changes in sleep-wakefulness, as reported for single neurons in the ascending reticular activating system. Evoked potentials were collected in 100 trial averages during continuous presentation of 1/sec clicks during initial awake recordings and thereafter during all-night sleep sessions. Continuously recorded EEG, EOG and EMG were scored for wakefulness, slow wave sleep (SWS), and rapid eye movement (REM) sleep during each evoked potential epoch. Recordings were obtained from electrodes implanted at the vertex and overlying the primary auditory cortex referenced to frontal sinus or to neck. In agreement with others, components of the auditory brain-stem response and the 12 msec primary cortical response showed no change in amplitude from wakefulness to either SWS or REM. Only wave A, among the components evaluated in the 1–100 msec range, decreased and disappeared during SWS and dramatically reappeared during REM to an amplitude equal to that during wakefulness. These data lend particular support to a functional relation between wave A and the ascending reticular activating system and suggest that this potential may provide a unique and dynamic probe of tonic brain activity. Moreover, this animal model provides a hypothetical basis for expecting a similar surface recorded potential in the human, a potential which has consequently been discovered.     

Circadian rhythms ofMacaca mulatta: Sleep,EEG, body and eye movement,and temperature

EEG, EOG, EMG, gross activity, and temperature were continuously recorded over 24 hours from 38 maleMacaca mulatta monkeys. EEG, EOG, and EMG were translated into standard Sleep Stages. The EEG also was automatically filtered, rectified, integrated, digitized, and plotted. Results are presented first as 24-hour plots for temperature, Sleep Stages Awake, 1–2, 3–4, and REM, for EMG, gross motion, EOG, and for the occurrence of EEG bands delta, theta, alpha, sigma, and beta.Twenty-four-hour cosine curves were fitted to the data by least squares, demonstrating and quantifying with confidence intervals a circadian rhythm in each function. During the 12-hour dark span the acrophases, or peaks of the cosines fitted to the data, appeared in the following clockwise order: Stage 3–4, delta, Stage 1–2, Stage REM, total EEG, and theta. The clockwise order of acrophases appearing in the light span was: alpha, gross motion, sigma, beta, EMG, EOG, temperature, and Stage Awake. Circadian amplitudes are given for each rhythm.Three of these measures of brain function had circadian rises or falls which appeared to be influenced by the daily times of light-on or -off. Lighting acted as circadian phase-synchronizing stimulus for temperature and the EEG bands beta and sigma.The data demonstrate circadian rhythms for certain parameters associated with sleep—delta, theta, Stage REM, and Stage 3–4 and important differences in phase. This finding constitutes another line of evidence that sleep is not unitary, consisting rather of related but separately controlled rhythmic functions.Considerable phylogenetic constancy appears when these data on circadian phasing are compared with similar data from other primates, including man.The experimental work was part of Aeromedical Research Laboratory Project 6892, Holloman Air Force Base, U.S.A.F. Biochemical assays were supported by the U.S.P.H.S. (MH 15413). Data analysis performed at the University of Minnesota was supported by NASA (NAS 2-2738 and NGR-24-005-006), the U.S. Air Force (F29608-69-C-0011), and by the USPHS (CA 5-K6-GM 13981). Further reproduction of this article is authorized as needed to meet the requirements of the U.S. government. The animals used in this study were handled in accordance with the Guide for Laboratory Animal Facilities and Care published by the National Academy of Science-National Research Council. Drs.Crowley andKripke were on active duty, and Dr.Pegram was a civilian employee with the USAF during part of this project.     

Electromyogram processing for sleep research

A computer method for quantifying the submental electromyographic surface interference pattern (EMG) during sleep and wakefulness by amplitude envelope measurement for consecutive 2-sec intervals is described. The method is largely insensitive to electrocardiogram (EKG) artifact. Though this algorithm was developed as part of a program to detect electroencephalographic (EEG), electrooculographic (EOG), tonic and phasic EMG changes during sleep, the method is applicable by itself wherever the envelope width of the EMG interference pattern is of interest. The results obtained correlate well with visual estimates of the amplitude envelope of the raw EMG. It offers increased speed, accuracy and reproducibility compared to visual EMG evaluation and enables a high degree of information extraction. The simplicity of the algorithm permits implementation and on-line processing on a small laboratory computer.     

Sleep staging with movement-related signals

Body movement related signals (i.e., activity due to postural changes and the ballistocardiac effort) were recorded from six normal volunteers using the static-charge-sensitive bed (SCSB). Visual sleep staging was performed on the basis of simultaneously recorded EEG, EMG and EOG signals. A statistical classification technique was used to determine if reliable sleep staging could be performed using only the SCSB signal. A classification rate of between 52% and 75% was obtained for sleep staging in the five conventional sleep stages and the awake state. These rates improved from 78% to 89% for classification between awake, REM and non-REM sleep and from 86% to 98% for awake versus asleep classification.     

Automatic EEG arousal detection for sleep apnea syndrome

Electroencephalographic (EEG) arousals are seen in EEG recordings as an awakening response of the human brain. Sleep apnea is a serious sleep disorder. Severe sleep apnea brings about EEG arousals and sleep for patients with sleep apnea syndrome (SAS) is thus frequently interrupted. The number of respiratory-related arousals during the whole night on PSG recordings is directly related to the quality of sleep. Detecting EEG arousals in the PSG record is thus a significant task for clinical diagnosis in sleep medicine. In this paper, a method for automatic detection of EEG arousals in SAS patients was proposed. To effectively detect respiratory-related arousals, threshold values were determined according to pathological events as sleep apnea and electromyogram (EMG). If resumption of ventilation (end of the apnea interval) was detected, much lower thresholds were adopted for detecting EEG arousals, including relatively doubtful arousals. Conversely, threshold was maintained high when pathological events were undetected. The proposed method was applied to polysomnographic (PSG) records of eight patients with SAS and accuracy of EEG arousal detection was verified by comparative visual inspection. Effectiveness of the proposed method in clinical diagnosis was also investigated.     

Relationship of ethanol blood level to REM and non-REM sleep time and distribution in the rat.

Direct electroencephalographic (EEG) and integrated electromyographic (EMG) recordings were analyzed for possible changes in the REM and non-REM sleep time in chronically implanted rats given 0, 1, 2, and 4 g/kg ethanol. REM and non-REM sleep were found, respectively, to be lessened and elevated in a dose-related manner. The degree of disruption of normal sleep-awake patterns was also found to correspond with blood-ethanol concentrations for the different doses of ethanol. These findings are discussed in relation to the influence of ethanol on the sleep of the human subject and the suggestion that the rat with chronic EEG and EMG electrodes may serve as a model for studying the degree of disruption of sleep upon chronic exposure to ethanol.     

Influence of repeated passive body heating on subsequent night sleep in humans

Experiments were carried out on four healthy male subjects in two separate sessions: (a) A baseline period of two consecutive nights, one spent at thermoneutrality [operative temperature (To) = 30 degrees C, dew-point temperature (Tdp) = 7 degrees C, air velocity (Va) = 0.2 m.s-1] and the other in hot condition (To = 35 degrees C, Tdp = 7 degrees C, Va = 0.2 m.s-1). During the day, the subjects lived in their normal housing and were engaged in their usual activities. (b) An acclimation period of seven consecutive daily heat exposures from 1400 to 1700 hours (To = 44 degrees C, Tdp = 29 degrees C, Va = 0.3 m.s-1). During each night, the subjects slept in thermoneutral or in hot conditions. The sleep measurements were: EEG from two sites, EOG from both eyes, EMG and EKG. Esophageal and ten skin temperatures were recorded continuously during the night. In the nocturnal hot conditions, a sweat collection capsule recorded the sweat gland activity in the different sleep stages. Results showed that passive body heating had no significant effect on the sleep structure of subsequent nights at thermoneutrality. In contrast, during nights at To = 35 degrees C an effect of daily heat exposure was observed on sleep. During the 2nd night of the heat acclimation period, sleep was more restless and less efficient than during the baseline night. The rapid eye movement sleep duration was reduced, while the rate of transient activation phases observed in sleep stage 2 increased significantly. On the 7th night, stage 4 sleep increased (+68%) over values observed during the baseline night.(ABSTRACT TRUNCATED AT 250 WORDS)     

Epilepsy: A heuristic model for relating nocturnal sleep EEG spike distributions to the risk of seizure

The cumulative temporal distribution of epileptiform events in the sleep EEG of epileptic patients is represented by a second-degree regression equation as a basis for summarizing properties of the spike distribution. This heuristic regression model attempts to provide a quantitative as well as an intuitive physical basis for understanding the relationship between epileptiform events and the state of the epileptic patient. The mathematical model is derived in this paper and the regression parameters associated with the model are interpreted. Computer implementation of the epileptiform event detection and the regression analysis have been carried out and typical results obtained from all-night sleep EEG recordings are described.     

Electrographic and behavioural correlations of the rest-activity cycle in the sea turtle, Caretta caretta L. (Chelonia)

Behavioural and electrographic (EEG, EOG, EMG, EKG) observations were carried out during the activity-inactivity cycle of the sea turtle, Caretta caretta L. Observations were made on turtles kept in a tank with flowing sea water and under natural lighting. There were no clearcut changes in the amount of activity associated with the dark-light phase, although the peak of activity occurred in the early afternoon hours. During periods of behavioural inactivity, EEG changes were minimum and were not accompanied by consistent changes in the EMG of the neck muscles. Eye movements were absent at that time. Although much information is lacking, we tentatively conclude that the sea turtle does not exhibit signs of sleep, but alternates between states of activity and inactivity that are simultaneous with a non-altered level of responsiveness. Such association of low but responsive activity might be of survival value.     

Feature ranking and rank aggregation for automatic sleep stage classification: a comparative study

Background

Nowadays, sleep quality is one of the most important measures of healthy life, especially considering the huge number of sleep-related disorders. Identifying sleep stages using polysomnographic (PSG) signals is the traditional way of assessing sleep quality. However, the manual process of sleep stage classification is time-consuming, subjective and costly. Therefore, in order to improve the accuracy and efficiency of the sleep stage classification, researchers have been trying to develop automatic classification algorithms. Automatic sleep stage classification mainly consists of three steps: pre-processing, feature extraction and classification. Since classification accuracy is deeply affected by the extracted features, a poor feature vector will adversely affect the classifier and eventually lead to low classification accuracy. Therefore, special attention should be given to the feature extraction and selection process.

Methods

In this paper the performance of seven feature selection methods, as well as two feature rank aggregation methods, were compared. Pz-Oz EEG, horizontal EOG and submental chin EMG recordings of 22 healthy males and females were used. A comprehensive feature set including 49 features was extracted from these recordings. The extracted features are among the most common and effective features used in sleep stage classification from temporal, spectral, entropy-based and nonlinear categories. The feature selection methods were evaluated and compared using three criteria: classification accuracy, stability, and similarity.

Results

Simulation results show that MRMR-MID achieves the highest classification performance while Fisher method provides the most stable ranking. In our simulations, the performance of the aggregation methods was in the average level, although they are known to generate more stable results and better accuracy.

Conclusions

The Borda and RRA rank aggregation methods could not outperform significantly the conventional feature ranking methods. Among conventional methods, some of them slightly performed better than others, although the choice of a suitable technique is dependent on the computational complexity and accuracy requirements of the user.

     

Time-of-day modulation of homeostatic and allostatic sleep responses to chronic sleep restriction in rats

To study sleep responses to chronic sleep restriction (CSR) and time-of-day influences on these responses, we developed a rat model of CSR that takes into account the polyphasic sleep patterns in rats. Adult male rats underwent cycles of 3 h of sleep deprivation (SD) and 1 h of sleep opportunity (SO) continuously for 4 days, beginning at the onset of the 12-h light phase ("3/1" protocol). Electroencephalogram (EEG) and electromyogram (EMG) recordings were made before, during, and after CSR. During CSR, total sleep time was reduced by ～60% from baseline levels. Both rapid eye movement sleep (REMS) and non-rapid eye movement sleep (NREMS) during SO periods increased initially relative to baseline and remained elevated for the rest of the CSR period. In contrast, NREMS EEG delta power (a measure of sleep intensity) increased initially, but then declined gradually, in parallel with increases in high-frequency power in the NREMS EEG. The amplitude of daily rhythms in NREMS and REMS amounts was maintained during SO periods, whereas that of NREMS delta power was reduced. Compensatory responses during the 2-day post-CSR recovery period were either modest or negative and gated by time of day. NREMS, REMS, and EEG delta power lost during CSR were not recovered by the end of the second recovery day. Thus the "3/1" CSR protocol triggered both homeostatic responses (increased sleep amounts and intensity during SOs) and allostatic responses (gradual decline in sleep intensity during SOs and muted or negative post-CSR sleep recovery), and both responses were modulated by time of day.     

Sleep and EEG spectra in the Syrian hamster (Mesocricetus auratus) under baseline conditions and following sleep deprivation

Summary Sleep was studied by continuous 24-h recordings in adult male Syrian hamsters, chronically implanted with EEG and EMG electrodes. Three vigilance states were determined using visual scoring and EEG power spectra (0.25–25 Hz) computed for 4-s episodes.The effects of two methods of total sleep deprivation (SD) were examined on vigilance states and the EEG power spectrum. The animals were subjected to 24 h SD by: (1) forced locomotion in a slowly rotating drum, (2) gentle handling whenever the hamsters attempted a sleeping posture. In addition, the hamsters were subjected to SD by handling during the first 3 h of the L period.Sleep predominated in the L period (78.2% of 12 h) and the D period (51.2%). The power spectra of the 3 vigilance states were similar during the L and D period. In NREM sleep, power density values in the low frequency range (0.25–6.0 Hz) exceeded those of REM sleep and W by a maximum factor of 8.3 and 2.8, respectively. At frequencies above 16 Hz, NREM and REM sleep power density values were significantly lower than during W. A progressive decrease in power density for low EEG frequencies (0.25–7 Hz) during NREM sleep was seen in the course of the L period. Power density values of higher frequencies (8–25 Hz) increased at the end of the L period and remained high during the first hours of the D period.The effect of prolonged SD on vigilance states and EEG spectra was similar by both methods and strikingly small compared to similar results in rats. In contrast, 3 h SD induced a large and more prolonged effect. The similarities and differences of sleep and sleep regulation are summarized for the hamster, rat and man.Abbreviations EEG electroencephalogram - LD light dark - REM rapid eye movements - NREM sleep non REM sleep - W waking - SD sleep deprivation - TST total sleep time - L light - D dark     

Alpha-bursts and K-complex: phasic activation pattern during spontaneous recovery of correct psychomotor performance at difference stages of drowsiness

It is known that phasic activation processes reveal themselves by different electrophysiological patterns depending on the sleep depth. Alpha bursts are an electrophysiological manifestation of arousal at the initial stage of sleep, whereas at the II stage K-complex becomes the main arousal pattern. We have shown earlier that during light drowsiness spontaneous recovery of correct psychomotor test performance (after an error) by a sitting subject is accompanied by EEG alpha bursts. The aim of this work was to study the EEG phasic activation pattern at deeper drowsiness during test performance by a subject in a lying position. Subjects had to press sensitive button in a lying position with closed eyes with self-paced oral counting of pressings. The experiment lasted for 40 min; EEG, EOG, and button pressing were recorded. It was shown that recovery of correct performance after errors at deeper drowsiness was accompanied by two types of EEG phasic activation patterns (PAP-1 and PAP-2). The alpha frequency component was always present in both PAP-1 and PAP-2. PAP-1 were observed at early stages of drowsiness and consisted of high-amplitude alpha bursts and EEG activity of higher frequency. PAP-2 were recorded at deeper stages and consisted of K-complexes with superposition of PAP-1. At first (medium level of drowsiness) the alpha bursts were superposed on the late slow K-complex components. With further deepening of drowsiness the early fast components of K-complex were also observed. The early appearance of K-complex during test performance at drowsiness seems to be associated with the urgent run of brain arousal systems, which at spontaneous falling asleep are in operation at the II sleep stage.