Sleep staging through an unsupervised learning lens

To assess the sleep detection and staging validity of a non-contact, commercially available bedside bio-motion sensing device (S+, ResMed) and evaluate the impact of algorithm updates. Polysomnography data from 27 healthy adult participants was compared epoch-by-epoch to synchronized data that were recorded and staged by actigraphy and S+. An update to the S+ algorithm (common in the rapidly evolving commercial sleep tracker industry) permitted comparison of the original (S+V1) and updated (S+V2) versions. Sleep detection accuracy by S+V1 (93.3%), S+V2 (93.8%), and actigraphy (96.0%) was high; wake detection accuracy by each (69.6%, 73.1%, and 47.9%, respectively) was low. Higher overall S+ specificity, compared to actigraphy, was driven by higher accuracy in detecting wake before sleep onset (WBSO), which differed between S+V2 (90.4%) and actigraphy (46.5%). Stage detection accuracy by the S+ did not exceed 67.6% (for stage N2 sleep, by S+V2) for any stage. Performance is compared to previously established variance in polysomnography scored by humans: a performance standard which commercial devices should ideally strive to reach. Similar limitations in detecting wake after sleep onset (WASO) were found for the S+ as have been previously reported for actigraphy and other commercial sleep tracking devices. S+ WBSO detection was higher than actigraphy, and S+V2 algorithm further improved WASO accuracy. Researchers and clinicians should remain aware of the potential for algorithm updates to impact validity. A commentary on this article appears in this issue on page 935.

The effects of serotonergic agents on respiration neuromodulation may vary according to differences in the serotonin system, such as those linked to depression. This study investigated how sleep-related respiratory disturbances relate to depression and the use of medications commonly prescribed for depression. Retrospective polysomnography was collated for all 363 individuals who met selection criteria out of 2,528 consecutive individuals referred to a specialized sleep clinic (Ottawa, Canada) between 2006 and 2016. The apnea-hypopnea index (AHI), oxygen saturation nadir, and oxygen desaturation index during REM and NREM sleep were analyzed using mixed analyses of covariance comparing 3 main groups: (1) medicated individuals with depressive disorders (antidepressant group; subdivided into the selective serotonin reuptake inhibitor and norepinephrine-dopamine reuptake inhibitor subgroups), (2) non-medicated individuals with depressive disorders (non-medicated group), and (3) mentally healthy control patients (control group). Individuals with depressive disorders (on antidepressants or not) had significantly higher AHIs compared to control patients (both P ≤ .007). The antidepressant group had a lower NREM sleep oxygen saturation nadir and a higher NREM sleep oxygen desaturation index than the control and non-medicated groups (all P ≤ .009). Within individuals with depressive disorders, independent of depression severity, the selective serotonin reuptake inhibitor group had a lower oxygen saturation nadir and a higher oxygen desaturation index during NREM sleep than the norepinephrine-dopamine reuptake inhibitor (both P ≤ .045) and non-medicated groups (both P < .001) and a higher NREM sleep AHI than the non-medicated group (P = .014). These findings suggest that the use of selective serotonin reuptake inhibitors may be associated with impaired breathing and worse nocturnal oxygen saturation in individuals with depressive disorders and sleep complaints, but this needs to be confirmed by prospective studies.

Narcolepsy is a sleep disorder affecting millions of people worldwide and causes serious public health problems. It is hard for doctors to correctly and objectively diagnose narcolepsy. Polysomnography (PSG) recordings, a gold standard for sleep monitoring and quality measurement, can provide abundant and objective cues for the narcolepsy diagnosis. There have been some studies on automatic narcolepsy diagnosis using PSG recordings. However, the sleep stage information, an important cue for narcolepsy diagnosis, has not been fully utilized. For example, some studies have not considered the sleep stage information to diagnose narcolepsy. Although some studies consider the sleep stage information, the stages are manually scored by experts, which is time-consuming and subjective. And the framework using sleep stages scored automatically for narcolepsy diagnosis is designed in a two-phase learning manner, where sleep staging in the first phase and diagnosis in the second phase, causing cumulative error and degrading the performance. To address these challenges, we propose a novel end-to-end framework for automatic narcolepsy diagnosis using PSG recordings. In particular, adopting the idea of multi-task learning, we take the sleep staging as our auxiliary task, and then combine the sleep stage related features with narcolepsy related features for our primary task of narcolepsy diagnosis. We collected a dataset of PSG recordings from 77 participants and evaluated our framework on it. Both of the sleep stage features and the end-to-end fashion contribute to diagnosis performance. Moreover, we do a comprehensive analysis on the relationship between sleep stages and narcolepsy, correlation of different channels, predictive ability of different sensing data, and diagnosis results in subject level.

Our objective was to evaluate a portable device (Somté, Compumedics, Australia), which incorporates 2 neurophysiological channels (electroencephalography and electrooculography) with cardiorespiratory monitoring for the diagnosis of obstructive sleep apnea (OSA). Full polysomnography (PSG) and Somté recordings were simultaneously performed in 68 patients with suspected OSA. Data were analyzed blindly by 2 scorers. A good agreement between methods in sleep efficiency was observed (68.8% [18.4] with PSG vs 68% [19.1] with Somté [p: n.s.] for scorer 1, and 67.5% [19.1] vs 68.4% [18.5; p: n.s.] for scorer 2). The apnea-hypopnea index (AHI) obtained with Somté was lower than with PSG: 19 (17.8) vs 21.7 (19) (p < 0.001) for scorer 1, and 16.6 (16.7) vs 20 (18.8) (p < 0.001) for scorer 2. The sensitivity of Somté for a PSG-AHI > 5 was 91% for scorer 1 and 90% for scorer 2, while specificity was 77% and 90%, respectively. The areas under the receiver operating curve for different PSG-AHI cutoff points (≥ 5, ≥ 15, and ≥ 30) were 0.81, 0.90, and 0.86, respectively, for scorer 1, and 0.90, 0.88, and 0.83 for scorer 2. These data suggest that Somté is an effective device to identify sleep and respiratory variables in patients with suspected OSA.

Sleep disorders affect a large portion of the global population and are strong predictors of morbidity and all-cause mortality. Sleep staging segments a period of sleep into a sequence of phases providing the basis for most clinical decisions in sleep medicine. Manual sleep staging is difficult and time-consuming as experts must evaluate hours of polysomnography (PSG) recordings with electroencephalography (EEG) and electrooculography (EOG) data for each patient. Here, we present U-Sleep, a publicly available, ready-to-use deep-learning-based system for automated sleep staging (sleep.ai.ku.dk). U-Sleep is a fully convolutional neural network, which was trained and evaluated on PSG recordings from 15,660 participants of 16 clinical studies. It provides accurate segmentations across a wide range of patient cohorts and PSG protocols not considered when building the system. U-Sleep works for arbitrary combinations of typical EEG and EOG channels, and its special deep learning architecture can label sleep stages at shorter intervals than the typical 30 s periods used during training. We show that these labels can provide additional diagnostic information and lead to new ways of analyzing sleep. U-Sleep performs on par with state-of-the-art automatic sleep staging systems on multiple clinical datasets, even if the other systems were built specifically for the particular data. A comparison with consensus-scores from a previously unseen clinic shows that U-Sleep performs as accurately as the best of the human experts. U-Sleep can support the sleep staging workflow of medical experts, which decreases healthcare costs, and can provide highly accurate segmentations when human expertize is lacking.

To reveal the characteristics of genioglossus (GG) activation in moderate and severe obstructive sleep apnea (OSA) patients during rapid eye movement (REM) sleep compared with non-rapid eye movement (NREM) sleep and to determine whether continuous positive airway pressure (CPAP) could improve GG activation in OSA patients during sleep. All subjects underwent polysomnography (PSG) with synchronous GG electromyography (GGEMG) recording with intra-oral surface electrodes at baseline on the first night. Only those subjects diagnosed with moderate and severe OSA were included and were manually titrated with CPAP to achieve a therapeutic pressure (Pt) with GGEMG recording on the second night. Nine OSA patients and six normal controls were analyzed in this study. The tonic GGEMG was higher in OSA patients during wakefulness (p = 0.003) and NREM sleep (p = 0.015), but it was not higher in REM sleep (p = 0.862). The average phasic activity of OSA patients was significantly higher in all stages, including wakefulness (p = 0.007), NREM sleep (p = 0.005), and REM sleep (p = 0.021). The peak phasic GGEMG was not different in wakefulness compared with normal controls (p = 0.240), but it was higher in OSA patients in NREM sleep (p = 0.001) and REM sleep (p = 0.021), and it was significantly reduced by using CPAP during sleep (NREM sleep: p = 0.027; REM sleep: p = 0.001). Our results demonstrate that GG activation during NREM and REM sleep is associated with component differences. The tonic component of GGEMG exhibited less of a compensatory increase compared with the phasic component in REM sleep, suggesting that it may be one of the pathological mechanisms of UA collapsibility in REM sleep. In addition, treatment with CPAP can normalize GGEMG activity and mostly reduced the peak phasic GGEMG during sleep.

Study objectiveTraditionally, age-related deterioration of sleep architecture in older individuals has been evaluated by visual scoring of polysomnographic (PSG) recordings with regard to total sleep time and latencies. In the present study, we additionally compared the non-REM sleep (NREM) stage and delta, theta, alpha, and sigma wave stability between young and older subjects to extract features that may explain age-related changes in sleep.MethodsPolysomnographic recordings were performed in 11 healthy older (72.6 ± 2.4 years) and 9 healthy young (23.3 ± 1.1 years) females. In addition to total sleep time, the sleep stage, delta power amplitude, and delta, theta, alpha, and sigma wave stability were evaluated by sleep stage transition analysis and a novel computational method based on a coefficient of variation of the envelope (CVE) analysis, respectively.ResultsIn older subjects, total sleep time and slow-wave sleep (SWS) time were shorter whereas wake after sleep onset was longer. The number of SWS episodes was similar between age groups, however, sleep stage transition analysis revealed that SWS was less stable in older individuals. NREM sleep stages in descending order of delta power were: SWS, N2, and N1, and delta power during NREM sleep in older subjects was lower than in young subjects. The CVE of the delta-band is an index of delta wave stability and showed significant differences between age groups. When separately analyzed for each NREM stage, different CVE clusters in NREM were clearly observed between young and older subjects. A lower delta CVE and amplitude were also observed in older subjects compared with young subjects in N2 and SWS. Additionally, lower CVE values in the theta, alpha and sigma bands were also characteristic of older participants.ConclusionThe present study shows a decrease of SWS stability in older subjects together with a decrease in delta wave amplitude. Interestingly, the decrease in SWS stability coincided with an increase in short-term delta, theta, sigma, and alpha power stability revealed by lower CVE. Loss of electroencephalograms (EEG) variability might be a useful marker of brain age.

The lack of adequate sleep during intensive care unit (ICU) admission is a frequently overlooked complication. Disrupted sleep is associated with immune system dysfunction, impaired resistance to infection, as well as alterations in nitrogen balance and wound healing. The effects of surgical ICU admission on patients' sleep quality and architecture remain poorly defined. The purpose of this study was to describe the quantity and quality of sleep as well as sleep architecture, as defined by polysomnography (PSG), in patients cared for in the surgical ICU. A prospective observational cohort study was performed at our urban Level I trauma center. A convenience sample of surgical or trauma ICU patients underwent continuous PSG for up to 24 hours to evaluate sleep patterns. A certified sleep technician performed, monitored, and scored all PSG recordings. A single neurologist trained in PSG interpretation reviewed all PSG recordings. chi goodness-of-fit analysis was performed to detect differences in the proportion of time spent in stages 1 and 2 (superficial stages), stages 3 and 4 (deep stages), or rapid eye movement (REM) sleep between study patients and healthy historical controls. All PSG recordings were performed greater than 24 hours after the administration of a general anesthetic. Patients with traumatic brain injury were excluded. Sixteen patients were selected to undergo PSG recordings. Median age was 37.5 years (range, 20-83), 81.3% were male patients, 62.5% were injured, and 31.3% were mechanically ventilated. Total PSG recording time was 315 hours (mean, 19.7 hours per patient), total sleep time captured by PSG was 132 hours (mean, 8.28 hours per patient), and there were 6.2 awakenings per hour of sleep measured. ICU patients had an increase in the proportion of time spent in the superficial stages of sleep, and a decrease in the proportion of time spent in the deeper stages of sleep as well as a decrease in REM sleep compared with healthy controls (p < 0.001). Patients do achieve measurable sleep while cared for in a surgical ICU setting. However, sleep is fragmented and the quality of sleep is markedly abnormal with significant reductions in stages 3 and 4 and REM, the deeper restorative stages of sleep. Further studies on the effects of a strategy to promote sleep during ICU care are warranted.

Previous studies showed that obstructive sleep apnea (OSA) is associated with dyslipidemia. However, whether micro-arousals during rapid eye movement (REM) and non-rapid eye movement (NREM) sleep independently associated with dyslipidemia were unknown. 4472 participants with OSA-related symptoms were finally included in our cohort. Various sleep variables including micro-arousal index (MAI) were obtained from standard polysomnography (PSG) recordings. Fasting serum lipid levels were assessed at our hospital laboratory. Linear regression models were employed to investigate relationships between micro-arousals in REM and NREM sleep and lipid profile with adjusting for multiple confounding factors. Fully adjusted models demonstrated a significant dose-dependent positive correlation between the MAI during REM sleep (MAI REM) and lipoprotein(a) (Lp(a)) levels (β = 0.061; P = 0.008), a finding that remained robust in sensitivity analyses (β = 0.068; P = 0.040). Subgroup analyses further revealed that this association was particularly prominent in patients with moderate-to-severe OSA (moderate: β = 0.106, P = 0.046; severe: β = 0.060,P = 0.045) and in male participants (β = 0.075,P = 0.003). MAIREM independently predicted elevated Lp(a) levels (P < 0.01), with peak effects manifested in males and moderate-to-severe OSA patients. The statistical findings regarding the association between REM sleep fragmentation and Lp(a) in our study may offer new insights into the close relationship between OSA and cardiovascular diseases.

PurposeObstructive sleep apnea (OSA) is a prevalent sleep-related breathing disorder. Research conducted on patients with OSA using electroencephalography (EEG) has revealed a noticeable shift in the overnight polysomnography (PSG) power spectrum. To better quantify the effects of OSA on brain function and to identify the most reliable predictors of pathological cortical activation, this study quantified the PSG power and its association with the degree of hypoxia in OSA patients.Patients and MethodsThis retrospective study recruited 93 patients with OSA. OSA patients were divided into three groups based on their apnea-hypopnea index (AHI) scores. The clinical characteristics and sleep macrostructure of these patients were examined, followed by an analysis of PSG signals. Power spectral density (PSD) in five frequency bands was analyzed during nonrapid eye movement (NREM) sleep, rapid eye movement (REM) sleep, and wakefulness. Finally, correlation analysis was conducted to assess the relationships among PSD, PSG parameters, and serum levels of S100β and uric acid.ResultsObstructive sleep apnea occurred during both the NREM and REM sleep phases. Except for a decrease in the duration of N2 sleep and an increase in the microarousal index, there were no significant differences in sleep architecture based on disease severity. Compared to the mild OSA group, the theta and alpha band PSD in the frontal and occipital regions during NREM sleep and wakefulness were significantly decreased in the moderate and severe OSA groups. Correlation analysis revealed that theta PSD in N1 and N3 stages were negatively correlated the AHI, oxygen desaturation index, SaO2<90% and microarousal index.ConclusionThese findings imply that patients with more severe OSA exhibited considerable NREM hypoxia and abnormal brain activity in the frontal and occipital regions. Therefore, sleep EEG oscillation may be a useful neurophysiological indicator for assessing brain function and disease severity in patients with OSA.

To measure the presence of the alpha-sleep anomaly in facioscapulohumeral muscular dystrophy (FSHD) and to evaluate the association between the sleep electroencephalogram (EEG) pattern and the presence of musculoskeletal pain. Cross-sectional study. Sleep laboratory. Fifty-five consecutive adult FSHD patients, 26 women and 29 men, age 49.6 ± 15.1 years (range 18-76). Questionnaires and polysomnography. Patients were asked to indicate if in the 3 months before the sleep study they presented persisting or recurring musculoskeletal pain. Patients who reported pain were asked to fill in the Italian version of the Brief Pain Inventory and the McGill Pain questionnaire, and a 101-point visual analog scale (VAS) for pain intensity. Polysomnographic recordings were performed. EEG was analyzed by means of Fast Fourier Transform. Four power spectra bands (δ 0-4 Hz, θ 4-8 Hz, α 8-14 Hz, β 14-32 Hz) were computed. Sleep macrostructure parameters and alpha/delta EEG power ratio during non rapid eye movement (NREM) sleep were compared between patients with and without pain. Forty-two patients in our sample reported chronic pain. VAS mean score was 55.2 ± 23.8 (range 10-100), pain rating index score was 13.8 ± 10.2, and present pain intensity was 2.5 ± 0.8. The statistical analysis documented an increased occurrence of the alpha and beta rhythms during NREM sleep in FSHD patients with pain. Significant correlations were observed between the alpha/delta power ratio during NREM sleep and pain measures. Chronic musculoskeletal pain is frequent in FSHD patients, and it represents a major mechanism of sleep disruption.

Machine-learning-based sleep staging models have achieved expert-level performance on standard polysomnographic (PSG) data. However, their application to EEG recorded by wearable devices remains limited by non-conventional referencing montage and the lack of benchmarking against PSG. Here, we tested whether an ensemble of state-of-the-art automatic staging algorithms can reliably classify sleep from a customized configuration of the ZMax headband, adapted to record a single fronto-mastoid EEG channel. A total of 35 nights of simultaneous ZMax and PSG recordings were acquired in a home setting, amounting to 250.02 hours of analysable data from 10 healthy participants. PSG data were scored according to AASM criteria by two independent experts from different sleep centres, with discrepancies resolved to obtain a consensus hypnogram. ZMax signal was processed using four machine-learning algorithms (YASA, U-Sleep, SleepTransformer, DeepResNet), whose predictions were further combined into a final ensemble scoring through soft-voting. The ensemble scoring achieved almost perfect agreement with human consensus staging (night-level mean ± SD; accuracy = 88.83% ± 2.84%, Cohen’s κ = 84.10% ± 4.52%, and Matthews Correlation Coefficient = 84.54% ± 4.23%). It showed excellent predictive accuracy for REM (F1-score = 93.99%), N3 (89.53%), N2 (87.93%), and wakefulness (86.37%), with lower performance for N1 (53.20%). These findings support the deployment of an ensemble scoring approach based on state-of-the-art sleep staging algorithms on ultra-minimal, mastoid-referenced EEG setups. This paradigm opens the way to the integration of data from modern wearable technologies into traditional PSG-based sleep research, overcoming longstanding barriers to ecological and large-scale sleep monitoring.

This preliminary study investigated the use of cardiac information or more specifically, heart rate variability (HRV), for automatic deep sleep detection throughout the night. The HRV data can be derived from cardiac signals, which were obtained from polysomnography (PSG) recordings. In total 42 features were extracted from the HRV data of 15 single-night PSG recordings (from 15 healthy subjects) for each 30-s epoch, used to perform epoch-by-epoch classification of deep sleep and non-deep sleep (including wake state and all the other sleep stages except deep sleep). To reduce variation of cardiac physiology between subjects, we normalized each feature per subject using a simple Z-score normalization method by subtracting the mean and dividing by the standard deviation of the feature values. A correlation-based feature selection (CFS) method was employed to select informative features as well as removing feature redundancy and a linear discriminant (LD) classifier was applied for deep and non-deep sleep classification. Results show that the use of Z-score normalization can significantly improve the classification performance. A Cohen's Kappa coefficient of 0.42 and an overall accuracy of 81.3% based on a leave-one-subject-out cross-validation were achieved.

Introduction Sleep monitoring hardware that allows for accurate sleep staging while also being unobtrusive and self-administered has the potential to make reliable EEG-based assessment of sleep quality at home widely accessible. However, for novel hardware utilizing dry EEG electrodes, a question remains regarding the similarity of these signals to traditional polysomnography (PSG), especially when used for therapy development. Here, we investigate how well EEG signals from a wearable dry electrode system can be sleep staged by an algorithm trained on traditional PSG. Methods The Dreem 3 headband, a low-profile sleep monitoring device with dry electrode EEG sensors, was used simultaneously with overnight PSG in 75 individuals. A machine learning model that was trained to stage sleep from the EEG signals of PSG, SleepStageML™, was run without modification on both the PSG and headband EEG signals. Each PSG recording was independently staged by 3 registered polysomnographic technologists to generate consensus ground-truth sleep stage labels. We compared the algorithmic performance of the sleep stages generated from the PSG signals to that of the sleep stages generated from the synchronously recorded headband data. Results Staging of PSG signals resulted in per-stage positive-percent-agreements (PPAs) of 92.4% for W, 58.4% for N1, 90.7% for N2, 76.0% for N3, and 92.9% for R. Negative-percent-agreements (NPAs) for these stages were 98.6%, 98.1%, 88.4%, 97.5%, and 98.4% respectively. When staging the headband recordings, performance remained high with PPAs for W, N1, N2, N3, and R of 94.4%, 34.7%, 88.2%, 78.6%, and 83.7% and NPAs of 96.2%, 98.9%, 86.5%, 96.3%, and 98.4% respectively. Cohen's Kappa on PSG signals was 0.82, and 0.77 on headband data. Conclusion We demonstrate that the EEG signals from a dry-electrode sleep monitoring system are similar enough to the EEG signals of a traditional PSG that an algorithm trained to stage PSG can also stage these dry-electrode signals. These results suggest that dry electrodes configured in a comfortable headband montage can capture clinical grade EEG useful for sleep staging with existing paradigms. These findings are promising for the broader investigation of cerebral functioning in sleep pathology, and could simplify the development of sleep biomarkers of neuropsychiatric disease. Support (if any)

Sleep restriction and emotion, electroencephalography (EEG) and dream recall, and insomnia and punctuality