Machine learning-empowered sleep staging classification using multi-modality signals

The goal is to enhance an automated sleep staging system's performance by leveraging the diverse signals captured through multi-modal polysomnography recordings. Three modalities of PSG signals, namely electroencephalogram (EEG), electrooculogram (EOG), and electromyogram (EMG), were considered to obtain the optimal fusions of the PSG signals, where 63 features were extracted. These include frequency-based, time-based, statistical-based, entropy-based, and non-linear-based features. We adopted the ReliefF (ReF) feature selection algorithms to find the suitable parts for each signal and superposition of PSG signals. Twelve top features were selected while correlated with the extracted feature sets' sleep stages. The selected features were fed into the AdaBoost with Random Forest (ADB + RF) classifier to validate the chosen segments and classify the sleep stages. This study's experiments were investigated by obtaining two testing schemes: epoch-wise testing and subject-wise testing. The suggested research was conducted using three publicly available datasets: ISRUC-Sleep subgroup1 (ISRUC-SG1), sleep-EDF(S-EDF), Physio bank CAP sleep database (PB-CAPSDB), and S-EDF-78 respectively. This work demonstrated that the proposed fusion strategy overestimates the common individual usage of PSG signals.

Interhemispheric differences of electroencephalography signal characteristics in different sleep stages

OBJECTIVE

The current electroencephalography (EEG) measurement setup is complex, laborious to set up, and uncomfortable for patients. We hypothesize that differences in EEG signal characteristics for sleep staging between the left and right hemispheres are negligible; therefore, there is potential to simplify the current measurement setup. We aimed to investigate the technical hemispheric differences in EEG signal characteristics along with electrooculography (EOG) signals during different sleep stages.

METHODS

Type II portable polysomnography (PSG) recordings of 50 patients were studied. Amplitudes and power spectral densities (PSDs) of the EEG and EOG signals were compared between the left (C3-M2, F3-M2, O1-M2, and E1-M2) and the right (C4-M1, F4-M1, O2-M1, and E2-M2) hemispheres. Regression analysis was performed to investigate the potential influence of sleep stages on the hemispheric differences in PSDs. Wilcoxon signed-rank tests were also employed to calculate the effect size of hemispheres across different frequency bands and sleep stages.

RESULTS

The results showed statistically significant differences in signal characteristics between hemispheres, but the absolute differences were minor. The median hemispheric differences in amplitudes were smaller than 3 μv with large interquartile ranges during all sleep stages. The absolute and relative PSD characteristics were highly similar between hemispheres in different sleep stages. Additionally, there were negligible differences in the effect size between hemispheres across all sleep stages.

CONCLUSIONS

Technical signal differences between hemispheres were minor across all sleep stages, indicating that both hemispheres contain similar information needed for sleep staging. A reduced measurement setup could be suitable for sleep staging without the loss of relevant information.

Multi-night home assessment of sleep structure in OSA with and without insomnia

OBJECTIVE

To explore sleep structure in participants with obstructive sleep apnea (OSA) and comorbid insomnia (COMISA) and participants with OSA without insomnia (OSA-only) using both single-night polysomnography and multi-night wrist-worn photoplethysmography/accelerometry.

METHODS

Multi-night 4-class sleep-staging was performed with a validated algorithm based on actigraphy and heart rate variability, in 67 COMISA (23 women, median age: 51 years) and 50 OSA-only (15 women, median age: 51) participants. Sleep statistics were compared using linear regression models and mixed-effects models. Multi-night variability was explored using a clustering approach and between- and within-participant analysis.

RESULTS

Polysomnographic parameters showed no significant group differences. Multi-night measurements, during 13.4 ± 5.2 nights per subject, demonstrated a longer sleep onset latency and lower sleep efficiency for the COMISA group. Detailed analysis of wake parameters revealed longer mean durations of awakenings in COMISA, as well as higher numbers of awakenings lasting 5 min and longer (WKN≥5min) and longer wake after sleep onset containing only awakenings of 5 min or longer. Within-participant variance was significantly larger in COMISA for sleep onset latency, sleep efficiency, mean duration of awakenings and WKN≥5min. Unsupervised clustering uncovered three clusters; participants with consistently high values for at least one of the wake parameters, participants with consistently low values, and participants displaying higher variability.

CONCLUSION

Patients with COMISA more often showed extended, and more variable periods of wakefulness. These observations were not discernible using single night polysomnography, highlighting the relevance of multi-night measurements to assess characteristics indicative for insomnia.

Research and application of deep learning-based sleep staging: Data, modeling, validation, and clinical practice

Over the past few decades, researchers have attempted to simplify and accelerate the process of sleep stage classification through various approaches; however, only a few such approaches have gained widespread acceptance. Artificial intelligence technology, particularly deep learning, is promising for earning the trust of the sleep medicine community in automated sleep-staging systems, thus facilitating its application in clinical practice and integration into daily life. We aimed to comprehensively review the latest methods that are applying deep learning for enhancing sleep staging efficiency and accuracy. Starting from the requisite "data" for constructing deep learning algorithms, we elucidated the current landscape of this domain and summarized the fundamental modeling process, encompassing signal selection, data pre-processing, model architecture, classification tasks, and performance metrics. Furthermore, we reviewed the applications of automated sleep staging in scenarios such as sleep-disorder screening, diagnostic procedures, and health monitoring and management. Finally, we conducted an in-depth analysis and discussion of the challenges and future in intelligent sleep staging, particularly focusing on large-scale sleep datasets, interdisciplinary collaborations, and human-computer interactions.

Studying sleep: towards the identification of hypnogram features that drive expert interpretation

STUDY OBJECTIVES

Hypnograms contain a wealth of information and play an important role in sleep medicine. However, interpretation of the hypnogram is a difficult task and requires domain knowledge and "clinical intuition." This study aimed to uncover which features of the hypnogram drive interpretation by physicians. In other words, make explicit which features physicians implicitly look for in hypnograms.

METHODS

Three sleep experts evaluated up to 612 hypnograms, indicating normal or abnormal sleep structure and suspicion of disorders. ElasticNet and convolutional neural network classification models were trained to predict the collected expert evaluations using hypnogram features and stages as input. The models were evaluated using several measures, including accuracy, Cohen's kappa, Matthew's correlation coefficient, and confusion matrices. Finally, model coefficients and visual analytics techniques were used to interpret the models to associate hypnogram features with expert evaluation.

RESULTS

Agreement between models and experts (Kappa between 0.47 and 0.52) is similar to agreement between experts (Kappa between 0.38 and 0.50). Sleep fragmentation, measured by transitions between sleep stages per hour, and sleep stage distribution were identified as important predictors for expert interpretation.

CONCLUSIONS

By comparing hypnograms not solely on an epoch-by-epoch basis, but also on these more specific features that are relevant for the evaluation of experts, performance assessment of (automatic) sleep-staging and surrogate sleep trackers may be improved. In particular, sleep fragmentation is a feature that deserves more attention as it is often not included in the PSG report, and existing (wearable) sleep trackers have shown relatively poor performance in this aspect.

Validity and reliability of the Oura Ring Generation 3 (Gen3) with Oura sleep staging algorithm 2.0 (OSSA 2.0) when compared to multi-night ambulatory polysomnography: A validation study of 96 participants and 421,045 epochs

PURPOSE

To evaluate the validity and the reliability of the Oura Ring Generation 3 (Gen3) with Oura Sleep Staging Algorithm 2.0 (OSSA 2.0) through multi-night polysomnography (PSG).

PARTICIPANTS AND METHODS

Participants were 96 generally healthy Japanese men and women aged between 20 and 70 years contributing with 421,045 30-s epochs. Sleep scoring was performed according to American Academy of Sleep Medicine criteria. Each participant could contribute with a maximum of three polysomnography (PSG) nights. Within-participant means were created for each sleep measure and paired t-tests were used to compare equivalent measures obtained from the PSG and Oura Rings (non-dominant and dominant hand). Agreement between sleep measures were assessed using Bland-Altman plots. Interrater reliability for epoch accuracy was determined by prevalence-adjusted and bias-adjusted kappa (PABAK).

RESULTS

The Oura Ring did not significantly differ from PSG for the measures time in bed, total sleep time, sleep onset latency, sleep period time, wake after sleep onset, time spent in light sleep, and time spent in deep sleep. Oura Rings worn on the non-dominant- and dominant-hand underestimated sleep efficiency by 1.1 %-1.5 % and time spent in REM sleep by 4.1-5.6 min. The Oura Ring had a sensitivity of 94.4 %-94.5 %, specificity of 73.0 %-74.6 %, a predictive value for sleep of 95.9 %-96.1 %, a predictive value for wake of 66.6 %-67.0 %, and accuracy of 91.7 %-91.8 %. PABAK was 0.83-0.84 and reliability was 94.8 %. Sleep staging accuracy ranged between 75.5 % (light sleep) and 90.6 % (REM sleep).

CONCLUSIONS

The Oura Ring Gen3 with OSSA 2.0 shows good agreement with PSG for global sleep measures and time spent in light and deep sleep.

InsightSleepNet: the interpretable and uncertainty-aware deep learning network for sleep staging using continuous Photoplethysmography

BACKGROUND

This study was conducted to address the existing drawbacks of inconvenience and high costs associated with sleep monitoring. In this research, we performed sleep staging using continuous photoplethysmography (PPG) signals for sleep monitoring with wearable devices. Furthermore, our aim was to develop a more efficient sleep monitoring method by considering both the interpretability and uncertainty of the model's prediction results, with the goal of providing support to medical professionals in their decision-making process.

METHOD

The developed 4-class sleep staging model based on continuous PPG data incorporates several key components: a local attention module, an InceptionTime module, a time-distributed dense layer, a temporal convolutional network (TCN), and a 1D convolutional network (CNN). This model prioritizes both interpretability and uncertainty estimation in its prediction results. The local attention module is introduced to provide insights into the impact of each epoch within the continuous PPG data. It achieves this by leveraging the TCN structure. To quantify the uncertainty of prediction results and facilitate selective predictions, an energy score estimation is employed. By enhancing both the performance and interpretability of the model and taking into consideration the reliability of its predictions, we developed the InsightSleepNet for accurate sleep staging.

RESULT

InsightSleepNet was evaluated using three distinct datasets: MESA, CFS, and CAP. Initially, we assessed the model's classification performance both before and after applying an energy score threshold. We observed a significant improvement in the model's performance with the implementation of the energy score threshold. On the MESA dataset, prior to applying the energy score threshold, the accuracy was 84.2% with a Cohen's kappa of 0.742 and weighted F1 score of 0.842. After implementing the energy score threshold, the accuracy increased to a range of 84.8-86.1%, Cohen's kappa values ranged from 0.75 to 0.78 and weighted F1 scores ranged from 0.848 to 0.861. In the case of the CFS dataset, we also noted enhanced performance. Before the application of the energy score threshold, the accuracy stood at 80.6% with a Cohen's kappa of 0.72 and weighted F1 score of 0.808. After thresholding, the accuracy improved to a range of 81.9-85.6%, Cohen's kappa values ranged from 0.74 to 0.79 and weighted F1 scores ranged from 0.821 to 0.857. Similarly, on the CAP dataset, the initial accuracy was 80.6%, accompanied by a Cohen's kappa of 0.73 and weighted F1 score was 0.805. Following the application of the threshold, the accuracy increased to a range of 81.4-84.3%, Cohen's kappa values ranged from 0.74 to 0.79 and weighted F1 scores ranged from 0.813 to 0.842. Additionally, by interpreting the model's predictions, we obtained results indicating a correlation between the peak of the PPG signal and sleep stage classification.

CONCLUSION

InsightSleepNet is a 4-class sleep staging model that utilizes continuous PPG data, serves the purpose of continuous sleep monitoring with wearable devices. Beyond its primary function, it might facilitate in-depth sleep analysis by medical professionals and empower them with interpretability for intervention-based predictions. This capability can also support well-informed clinical decision-making, providing valuable insights and serving as a reliable second opinion in medical settings.

An explainable deep-learning model to stage sleep states in children and propose novel EEG-related patterns in sleep apnea

Automatic deep-learning models used for sleep scoring in children with obstructive sleep apnea (OSA) are perceived as black boxes, limiting their implementation in clinical settings. Accordingly, we aimed to develop an accurate and interpretable deep-learning model for sleep staging in children using single-channel electroencephalogram (EEG) recordings. We used EEG signals from the Childhood Adenotonsillectomy Trial (CHAT) dataset (n = 1637) and a clinical sleep database (n = 980). Three distinct deep-learning architectures were explored to automatically classify sleep stages from a single-channel EEG data. Gradient-weighted Class Activation Mapping (Grad-CAM), an explainable artificial intelligence (XAI) algorithm, was then applied to provide an interpretation of the singular EEG patterns contributing to each predicted sleep stage. Among the tested architectures, a standard convolutional neural network (CNN) demonstrated the highest performance for automated sleep stage detection in the CHAT test set (accuracy = 86.9% and five-class kappa = 0.827). Furthermore, the CNN-based estimation of total sleep time exhibited strong agreement in the clinical dataset (intra-class correlation coefficient = 0.772). Our XAI approach using Grad-CAM effectively highlighted the EEG features associated with each sleep stage, emphasizing their influence on the CNN's decision-making process in both datasets. Grad-CAM heatmaps also allowed to identify and analyze epochs within a recording with a highly likelihood to be misclassified, revealing mixed features from different sleep stages within these epochs. Finally, Grad-CAM heatmaps unveiled novel features contributing to sleep scoring using a single EEG channel. Consequently, integrating an explainable CNN-based deep-learning model in the clinical environment could enable automatic sleep staging in pediatric sleep apnea tests.

GAC-SleepNet: A dual-structured sleep staging method based on graph structure and Euclidean structure

Sleep staging is a precondition for the diagnosis and treatment of sleep disorders. However, how to fully exploit the relationship between spatial features of the brain and sleep stages is an important task. Many current classical algorithms only extract the characteristic information of the brain in the Euclidean space without considering other spatial structures. In this study, a sleep staging network named GAC-SleepNet is designed. GAC-SleepNet uses the characteristic information in the dual structure of the graph structure and the Euclidean structure for the classification of sleep stages. In the graph structure, this study uses a graph convolutional neural network to learn the deep features of each sleep stage and converts the features in the topological structure into feature vectors by a multilayer perceptron. In the Euclidean structure, this study uses convolutional neural networks to learn the temporal features of sleep information and combine attention mechanism to portray the connection between different sleep periods and EEG signals, while enhancing the description of global features to avoid local optima. In this study, the performance of the proposed network is evaluated on two public datasets. The experimental results show that the dual spatial structure captures more adequate and comprehensive information about sleep features and shows advancement in terms of different evaluation metrics.

Automatic sleep staging by a hybrid model based on deep 1D-ResNet-SE and LSTM with single-channel raw EEG signals

Sleep staging is crucial for assessing sleep quality and diagnosing sleep disorders. Recent advances in deep learning methods with electroencephalogram (EEG) signals have shown remarkable success in automatic sleep staging. However, the use of deeper neural networks may lead to the issues of gradient disappearance and explosion, while the non-stationary nature and low signal-to-noise ratio of EEG signals can negatively impact feature representation. To overcome these challenges, we proposed a novel lightweight sequence-to-sequence deep learning model, 1D-ResNet-SE-LSTM, to classify sleep stages into five classes using single-channel raw EEG signals. Our proposed model consists of two main components: a one-dimensional residual convolutional neural network with a squeeze-and-excitation module to extract and reweight features from EEG signals, and a long short-term memory network to capture the transition rules among sleep stages. In addition, we applied the weighted cross-entropy loss function to alleviate the class imbalance problem. We evaluated the performance of our model on two publicly available datasets; Sleep-EDF Expanded consists of 153 overnight PSG recordings collected from 78 healthy subjects and ISRUC-Sleep includes 100 PSG recordings collected from 100 subjects diagnosed with various sleep disorders, and obtained an overall accuracy rate of 86.39% and 81.97%, respectively, along with corresponding macro average F1-scores of 81.95% and 79.94%. Our model outperforms existing sleep staging models in terms of overall performance metrics and per-class F1-scores for several sleep stages, particularly for the N1 stage, where it achieves F1-scores of 59.00% and 55.53%. The kappa coefficient is 0.812 and 0.766 for the Sleep-EDF Expanded and ISRUC-Sleep datasets, respectively, indicating strong agreement with certified sleep experts. We also investigated the effect of different weight coefficient combinations and sequence lengths of EEG epochs used as input to the model on its performance. Furthermore, the ablation study was conducted to evaluate the contribution of each component to the model's performance. The results demonstrate the effectiveness and robustness of the proposed model in classifying sleep stages, and highlights its potential to reduce human clinicians' workload, making sleep assessment and diagnosis more effective. However, the proposed model is subject to several limitations. Firstly, the model is a sequence-to-sequence network, which requires input sequences of EEG epochs. Secondly, the weight coefficients in the loss function could be further optimized to balance the classification performance of each sleep stage. Finally, apart from the channel attention mechanism, incorporating more advanced attention mechanisms could enhance the model's effectiveness.

ProductGraphSleepNet: Sleep staging using product spatio-temporal graph learning with attentive temporal aggregation

The classification of sleep stages plays a crucial role in understanding and diagnosing sleep pathophysiology. Sleep stage scoring relies heavily on visual inspection by an expert, which is a time-consuming and subjective procedure. Recently, deep learning neural network approaches have been leveraged to develop a generalized automated sleep staging and account for shifts in distributions that may be caused by inherent inter/intra-subject variability, heterogeneity across datasets, and different recording environments. However, these networks (mostly) ignore the connections among brain regions and disregard modeling the connections between temporally adjacent sleep epochs. To address these issues, this work proposes an adaptive product graph learning-based graph convolutional network, named ProductGraphSleepNet, for learning joint spatio-temporal graphs along with a bidirectional gated recurrent unit and a modified graph attention network to capture the attentive dynamics of sleep stage transitions. Evaluation on two public databases: the Montreal Archive of Sleep Studies (MASS) SS3; and the SleepEDF, which contain full night polysomnography recordings of 62 and 20 healthy subjects, respectively, demonstrates performance comparable to the state-of-the-art (Accuracy: 0.867;0.838, F1-score: 0.818;0.774 and Kappa: 0.802;0.775, on each database respectively). More importantly, the proposed network makes it possible for clinicians to comprehend and interpret the learned spatial and temporal connectivity graphs for sleep stages.

[Automatic sleep staging based on power spectral density and random forest]

The method of using deep learning technology to realize automatic sleep staging needs a lot of data support, and its computational complexity is also high. In this paper, an automatic sleep staging method based on power spectral density (PSD) and random forest is proposed. Firstly, the PSDs of six characteristic waves (K complex wave, δ wave, θ wave, α wave, spindle wave, β wave) in electroencephalogram (EEG) signals were extracted as the classification features, and then five sleep states (W, N1, N2, N3, REM) were automatically classified by random forest classifier. The whole night sleep EEG data of healthy subjects in the Sleep-EDF database were used as experimental data. The effects of using different EEG signals (Fpz-Cz single channel, Pz-Oz single channel, Fpz-Cz + Pz-Oz dual channel), different classifiers (random forest, adaptive boost, gradient boost, Gaussian naïve Bayes, decision tree, K-nearest neighbor), and different training and test set divisions (2-fold cross-validation, 5-fold cross-validation, 10-fold cross-validation, single subject) on the classification effect were compared. The experimental results showed that the effect was the best when the input was Pz-Oz single-channel EEG signal and the random forest classifier was used, no matter how the training set and test set were transformed, the classification accuracy was above 90.79%. The overall classification accuracy, macro average F1 value, and Kappa coefficient could reach 91.94%, 73.2% and 0.845 respectively at the highest, which proved that this method was effective and not susceptible to data volume, and had good stability. Compared with the existing research, our method is more accurate and simpler, and is suitable for automation.

Automatic sleep staging for the young and the old - Evaluating age bias in deep learning

BACKGROUND

Various deep-learning systems have been proposed for automated sleep staging. Still, the significance of age-specific underrepresentation in training data and the resulting errors in clinically used sleep metrics are unknown.

METHODS

We adopted XSleepNet2, a deep neural network for automated sleep staging, to train and test models using polysomnograms of 1232 children (7.0 ± 1.4 years) and 3757 adults (56.9 ± 19.4 years) and 2788 older adults (mean 80.7 ± 4.2 years). We developed four separate sleep stage classifiers using exclusively pediatric (P), adult (A), older adults (O) as well as PSG from mixed cohorts: pediatric, adult, and older adult (PAO). Results were compared against an alternative sleep stager (DeepSleepNet) for validation purposes.

RESULTS

When pediatric PSG was classified by XSleepNet2 exclusively trained on pediatric PSG, the overall accuracy was 88.9%, dropping to 78.9% when subjected to a system trained exclusively on adult PSG. Errors performed by the system staging PSG of older people were comparably lower. However, all systems produced significant errors in clinical markers when considering individual PSG. Results obtained with DeepSleepNet showed similar patterns.

CONCLUSION

Underrepresentation of age groups, in particular children, can significantly lower the performance of automatic deep-learning sleep stagers. In general, automated sleep stagers may behave unexpectedly, limiting clinical use. Future evaluation of automated systems must pay attention to PSG-level performance and overall accuracy.

LWSleepNet: A lightweight attention-based deep learning model for sleep staging with singlechannel EEG

Introduction

Sleep is vital to human health, and sleep staging is an essential process in sleep assessment. However, manual classification is an inefficient task. Along with the increased demand for portable sleep quality detection devices, lightweight automatic sleep staging needs to be developed.

Methods

This study proposes a novel attention-based lightweight deep learning model called LWSleepNet. A depthwise separable multi-resolution convolutional neural network is introduced to analyze the input feature map and captures features at multiple frequencies using two different sized convolutional kernels. The temporal feature extraction module divides the input into patches and feeds them into a multi-head attention block to extract time-dependent information from sleep recordings. The model's convolution operations are replaced with depthwise separable convolutions to minimize its number of parameters and computational cost. The model's performance on two public datasets (Sleep-EDF-20 and Sleep-EDF-78) was evaluated and compared with those of previous studies. Then, an ablation study and sensitivity analysis were performed to evaluate further each module.

Results

LWSleepNet achieves an accuracy of 86.6% and Macro-F1 score of 79.2% for the Sleep-EDF-20 dataset and an accuracy of 81.5% and Macro-F1 score of 74.3% for the Sleep-EDF-78 dataset with only 55.3 million floating-point operations per second and 180 K parameters.

Conclusion

On two public datasets, LWSleepNet maintains excellent prediction performance while substantially reducing the number of parameters, demonstrating that our proposed Light multiresolution convolutional neural network and temporal feature extraction modules can provide excellent portability and accuracy and can be easily integrated into portable sleep monitoring devices.

Comparison of deep transfer learning algorithms and transferability measures for wearable sleep staging

Background

Obtaining medical data using wearable sensors is a potential replacement for in-hospital monitoring, but the lack of data for such sensors poses a challenge for development. One solution is using in-hospital recordings to boost performance via transfer learning. While there are many possible transfer learning algorithms, few have been tested in the domain of EEG-based sleep staging. Furthermore, there are few ways for determining which transfer learning method will work best besides exhaustive testing. Measures of transferability do exist, but are typically used for selection of pre-trained models rather than algorithms and few have been tested on medical signals. We tested several supervised transfer learning algorithms on a sleep staging task using a single channel of EEG (AF7-Fpz) captured from an in-home commercial system.

Results

Two neural networks—one bespoke and another state-of-art open-source architecture—were pre-trained on one of six source datasets comprising 11,561 subjects undergoing clinical polysomnograms (PSGs), then re-trained on a target dataset of 75 full-night recordings from 24 subjects. Several transferability measures were then tested to determine which is most effective for assessing performance on unseen target data. Performance on the target dataset was improved using transfer learning, with re-training the head layers being the most effective in the majority of cases (up to 63.9% of cases). Transferability measures generally provided significant correlations with accuracy (up to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r = -0.53$$\end{document}r=-0.53).

Conclusion

Re-training the head layers provided the largest performance boost. Transferability measures are useful indicators of transfer learning effectiveness.

A lightweight automatic sleep staging method for children using single-channel EEG based on edge artificial intelligence

With the development of telemedicine and edge computing, edge artificial intelligence (AI) will become a new development trend for smart medicine. On the other hand, nearly one-third of children suffer from sleep disorders. However, all existing sleep staging methods are for adults. Therefore, we adapted edge AI to develop a lightweight automatic sleep staging method for children using single-channel EEG. The trained sleep staging model will be deployed to edge smart devices so that the sleep staging can be implemented on edge devices which will greatly save network resources and improving the performance and privacy of sleep staging application. Then the results and hypnogram will be uploaded to the cloud server for further analysis by the physicians to get sleep disease diagnosis reports and treatment opinions. We utilized 1D convolutional neural networks (1D-CNN) and long short term memory (LSTM) to build our sleep staging model, named CSleepNet. We tested the model on our childrens sleep (CS) dataset and sleep-EDFX dataset. For the CS dataset, we experimented with F4-M1 channel EEG using four different loss functions, and the logcosh performed best with overall accuracy of 83.06% and F1-score of 76.50%. We used Fpz-Cz and Pz-Oz channel EEG to train our model in Sleep-EDFX dataset, and achieved an accuracy of 86.41% without manual feature extraction. The experimental results show that our method has great potential. It not only plays an important role in sleep-related research, but also can be widely used in the classification of other time sequences physiological signals.

Improving Machine Learning Technology in the Field of Sleep

The authors discuss the challenges of machine- and deep learning-based automatic analysis of obstructive sleep apnea with respect to known issues with the signal interpretation, patient physiology, and the apnea-hypopnea index. Their goal is to provide guidance for sleep and machine learning professionals working in this area of sleep medicine. They suggest that machine learning approaches may well be better targeted at examining and attempting to improve the diagnostic criteria, in order to build a more nuanced understanding of the detailed circumstances surrounding OSA, rather than merely attempting to reproduce human scoring.

Cross-Gram matrices and their use in transfer learning: Application to automatic REM detection using heart rate

BACKGROUND AND OBJECTIVES

while traditional sleep staging is achieved through the visual - expert-based - annotation of a polysomnography, it has the disadvantages of being unpractical and expensive. Alternatives have been developed over the years to relieve sleep staging from its heavy requirements, through the collection of more easily assessable signals and its automation using machine learning. However, these alternatives have their limitations, some due to variabilities among and between subjects, other inherent to their use of sub-discriminative signals. Many new solutions rely on the evaluation of the Autonomic Nervous System (ANS) activation through the assessment of the heart-rate (HR); the latter is modulated by the aforementioned variabilities, which may result in data and concept shifts between what was learned and what we want to classify. Such adversary effects are usually tackled by Transfer Learning, dealing with problems where there are differences between what is known (source) and what we want to classify (target). In this paper, we propose two new kernel-based methods of transfer learning and assess their performances in Rapid-Eye-Movement (REM) sleep stage detection, using solely the heart rate.

METHODS

our first contribution is the introduction of Kernel-Cross Alignment (KCA), a measure of similarity between a source and a target, which is a direct extension of Kernel-Target Alignment (KTA). To our knowledge, KCA has currently never been studied in the literature. Our second contribution is two alignment-based methods of transfer learning: Kernel-Target Alignment Transfer Learning (KTATL) and Kernel-Cross Alignment Transfer Learning (KCATL). Both methods differ from KTA, whose traditional use is kernel-tuning: in our methods, the kernel has been fixed beforehand, and our objective is the improvement of the estimation of unknown target labels by taking into account how observations relate to each other, which, as it will be explained, allows to transfer knowledge (transfer learning).

RESULTS

we compare performances with transfer learning (KCATL, KTATL) to performances without transfer using a fixed classifier (a Support Vector Classifier - SVC). In most cases, both transfer learning methods result in an improvement of performances (higher detection rates for a fixed false-alarm rate). Our methods do not require iterative computations.

CONCLUSION

we observe improved performances using our transfer methods, which are computationally efficient, as they only require the computation of a kernel matrix and are non-iterative. However, some optimisation aspects are still under investigation.

'SleepCycles' package for R - A free software tool for the detection of sleep cycles from sleep staging

The detection of NREM-REM sleep cycles in human sleep data (i.e., polysomnographically assessed sleep stages) enables fine-grained analyses of ultradian variations in sleep microstructure (e.g., sleep spindles, and arousals), or other amplitude- and frequency-specific electroencephalographic features during sleep. While many laboratories have software that is used internally, reproducibility requires the availability of open-source software. Therefore, we here introduce the ‘SleepCycles’ package for R, an open-source software package that identifies sleep cycles and their respective (non-) rapid eye movement ([N]REM) periods from sleep staging data. Additionally, each (N)REM period is subdivided into parts of equal duration (percentiles), which may be useful for further fine-grained analyses. The detection criteria used in the package are, with some adaptations, largely based on criteria originally proposed by Feinberg and Floyd (1979). The latest version of the package can be downloaded from the Comprehensive R Archives Network (CRAN).•

The package ‘SleepCycles’ for R allows to identify sleep cycles and their respective NREM and REM periods from sleep staging results.

•

Besides the cycle detection, NREM and REM periods are also split into parts of equal duration (percentiles) thereby allowing for a better temporal resolution across the night and comparisons of sleep cycles with different durations amongst different night recordings.

A deep learning algorithm based on 1D CNN-LSTM for automatic sleep staging

BACKGROUND

Sleep staging is an important part of sleep research. Traditional automatic sleep staging based on machine learning requires extensive feature extraction and selection.

OBJECTIVE

This paper proposed a deep learning algorithm without feature extraction based on one-dimensional convolutional neural network and long short-term memory.

METHODS

The algorithm can automatically divide sleep into 5 phases including awake period, non-rapid eye movement sleep period (N1 ∼ N3) and rapid eye movement using the electroencephalogram signals. The raw signal was processed by the wavelet transform. Then, the processed signal was directly input into the deep learning algorithm to obtain the staging result.

RESULTS

The accuracy of staging is 93.47% using the Fpz-Cz electroencephalogram signal. When using the Fpz-Cz and electroencephalogram signal, the algorithm can obtain the highest accuracy of 94.15%.

CONCLUSION

These results show that this algorithm is suitable for different physiological signals and can realize end-to-end automatic sleep staging without any manual feature extraction.

Investigation of low dimensional feature spaces for automatic sleep staging

BACKGROUND

Automatic sleep stage classification depends crucially on the selection of a good set of descriptive features. However, the selection of a feature set with an appropriate low computational cost without compromising classification performance is still a challenge. This study attempts to represent sleep EEG patterns using a minimum number of features, without significant performance loss.

METHODS

Three feature selection algorithms were applied to a high dimensional feature space comprising 84 features. These methods were based on a bootstrapping approach guided by Gini ranking and mutual information between the features. The algorithms were tested on three scalp electroencephalography (EEG) and one ear-EEG datasets. The relationship between the information carried by different features was investigated using mutual information and illustrated by a graphical clustering tool.

RESULTS

The minimum number of features that can represent the whole feature set without performance loss was found to range between 5 and 11 for different datasets. In ear-EEG, 7 features based on Continuous Wavelet Transform (CWT) resulted in similar performance as the whole set whereas in two scalp EEG datasets, the difference between minimal CWT set and the whole set was statistically significant (0.008 and 0.017 difference in average kappa). Features were divided into groups according to the type of information they carry. The group containing relative power features was identified as the most informative feature group in sleep stage classification, whereas the group containing non-linear features was found to be the least informative.

CONCLUSIONS

This study showed that EEG sleep staging can be performed based on a low dimensional feature space without significant decrease in sleep staging performance. This is especially important in the case of wearable devices like ear-EEG where low computational complexity is needed. The division of the feature space into groups of features, and the analysis of the distribution of feature groups for different feature set sizes, is helpful in the selection of an appropriate feature set.

Hybrid manifold-deep convolutional neural network for sleep staging

Analysis of electroencephalogram (EEG) is a crucial diagnostic criterion for many sleep disorders, of which sleep staging is an important component. Manual stage classification is a labor-intensive process and usually suffered from many subjective factors. Recently, more and more computer-aided techniques have been applied to this task, among which deep convolutional neural network has been performing well as an effective automatic classification model. Despite some comprehensive models have been developed to improve classification results, the accuracy for clinical applications has not been reached due to the lack of sufficient labeled data and the limitation of extracting latent discriminative EEG features. Therefore, we propose a novel hybrid manifold-deep convolutional neural network with hyperbolic attention. To overcome the shortage of labeled data, we update the semi-supervised training scheme as an optimal solution. In order to extract the latent feature representation, we introduce the manifold learning module and the hyperbolic module to extract more discriminative information. Eight subjects from the public dataset are utilized to evaluate our pipeline, and the model achieved 89% accuracy, 70% precision, 80% sensitivity, 72% f1-score and kappa coefficient of 78%, respectively. The proposed model demonstrates powerful ability in extracting feature representation and achieves promising results by using semi-supervised training scheme. Therefore, our approach shows strong potential for future clinical development.

Uncovering the structure of clinical EEG signals with self-supervised learning

OBJECTIVE

Supervised learning paradigms are often limited by the amount of labeled data that is available. This phenomenon is particularly problematic in clinically-relevant data, such as electroencephalography (EEG), where labeling can be costly in terms of specialized expertise and human processing time. Consequently, deep learning architectures designed to learn on EEG data have yielded relatively shallow models and performances at best similar to those of traditional feature-based approaches. However, in most situations, unlabeled data is available in abundance. By extracting information from this unlabeled data, it might be possible to reach competitive performance with deep neural networks despite limited access to labels.

APPROACH

We investigated self-supervised learning (SSL), a promising technique for discovering structure in unlabeled data, to learn representations of EEG signals. Specifically, we explored two tasks based on temporal context prediction as well as contrastive predictive coding on two clinically-relevant problems: EEG-based sleep staging and pathology detection. We conducted experiments on two large public datasets with thousands of recordings and performed baseline comparisons with purely supervised and hand-engineered approaches.

MAIN RESULTS

Linear classifiers trained on SSL-learned features consistently outperformed purely supervised deep neural networks in low-labeled data regimes while reaching competitive performance when all labels were available. Additionally, the embeddings learned with each method revealed clear latent structures related to physiological and clinical phenomena, such as age effects.

SIGNIFICANCE

We demonstrate the benefit of SSL approaches on EEG data. Our results suggest that self-supervision may pave the way to a wider use of deep learning models on EEG data.

Direct application of an ECG-based sleep staging algorithm on reflective photoplethysmography data decreases performance

OBJECTIVE

The maturation of neural network-based techniques in combination with the availability of large sleep datasets has increased the interest in alternative methods of sleep monitoring. For unobtrusive sleep staging, the most promising algorithms are based on heart rate variability computed from inter-beat intervals (IBIs) derived from ECG-data. The practical application of these algorithms is even more promising when alternative ways of obtaining IBIs, such as wrist-worn photoplethysmography (PPG) can be used. However, studies validating sleep staging algorithms directly on PPG-based data are limited.

RESULTS

We applied an automatic sleep staging algorithm trained and validated on ECG-data directly on inter-beat intervals derived from a wrist-worn PPG sensor, in 389 polysomnographic recordings of patients with a variety of sleep disorders. While the algorithm reached moderate agreement with gold standard polysomnography, the performance was significantly lower when applied on PPG- versus ECG-derived heart rate variability data (kappa 0.56 versus 0.60, p < 0.001; accuracy 73.0% versus 75.9% p < 0.001). These results show that direct application of an algorithm on a different source of data may negatively affect performance. Algorithms need to be validated using each data source and re-training should be considered whenever possible.

Sleep staging algorithm based on multichannel data adding and multifeature screening

BACKGROUND

Sleep staging is an important basis of sleep research, which is closely related to both normal sleep physiology and sleep disorders. Many studies have reported various sleep staging algorithms of which the framework generally consists of three parts: signal preprocessing, feature extraction and classification. However, there are few studies on the superposition of signals and feature screening for sleep staging.

OBJECTIVE

The objectives were to (1) Analyze the effective signal enhancement based on the superposition of homologous and heterogeneous signals, (2) Find a better way to use multichannel signals, (3) Study a systematic method of feature screening for sleep staging, and (4) Improve the performance of automatic sleep staging.

METHODS

In this paper, a novel method of signal preprocessing and feature screening was proposed. In the signal preprocessing, multi-channel signal superposition was applied to improve the effective information contained in the original signal. In the feature screening, 62 features were initially selected including the time-domain features, frequency-domain features and nonlinear features, and a ReliefF algorithm was employed to select 14 features highly correlated to sleep stages from the former 62 features. Then, Pearson correlation coefficients were used to remove 2 redundant features from the 14 features to eventually obtain 12 features. Next, with the aforementioned signal preprocessing method, the 12 selected features and a support vector machine (SVM) classifier were used for sleep staging based on thirty recordings.

RESULTS

Comparing the performance of sleep staging using different single-channel signals and different multi-channel superposition signals, we found that the best performance was obtained while using the superposition of two electroencephalogram (EEG) signals. The overall accuracies of sleep staging with 2-6 classes obtained by superposing the two EEG signals reach 98.28%, 95.50%, 94.28%, 93.08% and 92.34%, respectively, and the kappa coefficient of sleep staging with 6 classes reaches 84.07%.

CONCLUSIONS

Among the proposed sleep staging methods of using single-channel signal and multi-channel signal superposition, the best performance and consistency were obtained while using the superposition of two electroencephalogram (EEG) signals. The multichannel signal superposition method pointed out a valuable direction for improving the performance of automatic sleep staging in both theoretical research and engineering applications, and the proposed systematical feature screening method opened up a reasonable pathway for better selecting type and number of features for sleep staging.

Automated multi-model deep neural network for sleep stage scoring with unfiltered clinical data

Purpose

To develop an automated framework for sleep stage scoring from PSG via a deep neural network.

Methods

An automated deep neural network was proposed by using a multi-model integration strategy with multiple signal channels as input. All of the data were collected from one single medical center from July 2017 to April 2019. Model performance was evaluated by overall classification accuracy, precision, recall, weighted F1 score, and Cohen’s Kappa.

Results

Two hundred ninety-four sleep studies were included in this study; 122 composed the training dataset, 20 composed the validation dataset, and 152 were used in the testing dataset. The network achieved human-level annotation performance with an average accuracy of 0.8181, weighted F1 score of 0.8150, and Cohen’s Kappa of 0.7276. Top-2 accuracy (the proportion of test samples for which the true label is among the two most probable labels given by the model) was significantly improved compared to the overall classification accuracy, with the average being 0.9602. The number of arousals affected the model’s performance.

Conclusion

This research provides a robust and reliable model with the inter-rater agreement nearing that of human experts. Determining the most appropriate evaluation parameters for sleep staging is a direction for future research.

Electronic supplementary material

The online version of this article (10.1007/s11325-019-02008-w) contains supplementary material, which is available to authorized users.

Sleep staging from single-channel EEG with multi-scale feature and contextual information

PURPOSE

Portable sleep monitoring devices with less-attached sensors and high-accuracy sleep staging methods can expedite sleep disorder diagnosis. The aim of this study was to propose a single-channel EEG sleep staging model, SleepStageNet, which extracts sleep EEG features by multi-scale convolutional neural networks (CNN) and then infers the type of sleep stages by capturing the contextual information between adjacent epochs using recurrent neural networks (RNN) and conditional random field (CRF).

METHODS

To verify the feasibility of our model, two datasets, one composed by two different single-channel EEGs (Fpz-Cz and Pz-Oz) on 20 healthy people and one composed by a single-channel EEG (F4-M1) on 104 obstructive sleep apnea (OSA) patients with different severities, were examined. The corresponding sleep stages were scored as four states (wake, REM, light sleep, and deep sleep). The accuracy measures were obtained from epoch-by-epoch comparison between the model and PSG scorer, and the agreement between them was quantified with Cohen's kappa (ҡ).

RESULTS

Our model achieved superior performance with average accuracy (Fpz-Cz, 0.88; Pz-Oz, 0.85) and ҡ (Fpz-Cz, 0.82; Pz-Oz, 0.77) on the healthy people. Furthermore, we validated this model on the OSA patients with average accuracy (F4-M1, 0.80) and ҡ (F4-M1, 0.67). Our model significantly improved the accuracy and ҡ compared to previous methods.

CONCLUSIONS

The proposed SleepStageNet has proved feasible for assessment of sleep architecture among OSA patients using single-channel EEG. We suggest that this technological advancement could augment the current use of home sleep apnea testing.

Modeling brain dynamic state changes with adaptive mixture independent component analysis

There is a growing interest in neuroscience in assessing the continuous, endogenous, and nonstationary dynamics of brain network activity supporting the fluidity of human cognition and behavior. This non-stationarity may involve ever-changing formation and dissolution of active cortical sources and brain networks. However, unsupervised approaches to identify and model these changes in brain dynamics as continuous transitions between quasi-stable brain states using unlabeled, noninvasive recordings of brain activity have been limited. This study explores the use of adaptive mixture independent component analysis (AMICA) to model multichannel electroencephalographic (EEG) data with a set of ICA models, each of which decomposes an adaptively learned portion of the data into statistically independent sources. We first show that AMICA can segment simulated quasi-stationary EEG data and accurately identify ground-truth sources and source model transitions. Next, we demonstrate that AMICA decomposition, applied to 6-13 channel scalp recordings from the CAP Sleep Database, can characterize sleep stage dynamics, allowing 75% accuracy in identifying transitions between six sleep stages without use of EEG power spectra. Finally, applied to 30-channel data from subjects in a driving simulator, AMICA identifies models that account for EEG during faster and slower response to driving challenges, respectively. We show changes in relative probabilities of these models allow effective prediction of subject response speed and moment-by-moment characterization of state changes within single trials. AMICA thus provides a generic unsupervised approach to identifying and modeling changes in EEG dynamics. Applied to continuous, unlabeled multichannel data, AMICA may likely be used to detect and study any changes in cognitive states.

Automated sleep staging of OSAs based on ICA preprocessing and consolidation of temporal correlations

An automated sleep staging based on analyzing long-range time correlations in EEG is proposed. These correlations, indicating time-scale invariant property or self-similarity at different time scales, are known to be salient dynamical characteristics of stage succession for a sleeping brain even when the subject suffers a destructive disorder such as Obstructive Sleep Apnea (OSA). The goal is to extract a set of complementary features from cerebral sources mapped onto the scalp electrodes or from a number of denoised EEG channels. For this purpose, source localization/extraction and noise reduction approaches based on Independent Component Analysis were used prior to correlation analysis. Feature extracted segments were then classified in one of the five classes including WAKE, STAGE1, STAGE2, SWS and REM via an ensemble neuro-fuzzy classifier. Some techniques were employed to improve the classifier's performance including Scaled Conjugate Gradient Method to speed up learning the ANFIS classifiers, a pruning algorithm to eliminate irrelevant fuzzy rules and the 10-fold cross-validation technique to train and test the system more efficiently. The performance of classification for two strategies including (1) feature extraction from effective cerebral sources and (2) feature extraction from selected channels of denoised EEG signals was compared and contrasted in terms of training errors and test accuracies. The first and second strategies achieved 92.23 and 88.74% agreement with human expert respectively which indicates the effectiveness of the staging system based on cerebral sources of activity. Our results further indicate that the misclassification rates were almost below 10%. The proposed automated sleep staging system is reliable due to the fact that it is based on the underlying dynamics of sleep staging which is less likely to be affected by sleep fragmentations occurred repeatedly in OSA.

Accuracy of a smartphone application in estimating sleep in children

PURPOSE

Chronic sleep problems can lead to difficulties for both the individual and society at large, making it important to effectively measure sleep. This study assessed the accuracy of an iPhone application (app) that could potentially be used as a simple, inexpensive means to measure sleep over an extended period of time in the home.

METHODS

Twenty-five subjects from the ages of 2-14 who were undergoing overnight polysomnography (PSG) were recruited. The phone was placed on the mattress, near their pillow, and recorded data simultaneously with the PSG. The data were then downloaded and certain parameters were compared between the app and PSG, including total sleep time, sleep latency, and time spent in various defined "stages."

RESULTS

Although there seemed to be a visual relationship between the graphs generated by the app and PSG, this was not confirmed on numerical analysis. There was no correlation between total sleep time or sleep latency between the app and PSG. Sleep latency from the PSG and latency to "deep sleep" from the app had a significant relationship (p = 0.03). No combination of PSG sleep stages corresponded with app "stages" in a meaningful way.

CONCLUSIONS

The Sleep Cycle App may have value in increasing the user's awareness of sleep issues, but it is not yet accurate enough to be used as a clinical tool.

Case comparison of sleep features from ear-EEG and scalp-EEG

Background

We investigate the potential usability of a novel in-the-ear electroencephalography recording device for sleep staging.

Methods

In one healthy subject we compare simultaneous earelectroencephalography to standard scalp EEG visually and using power spectrograms. Hypnograms independently derived from the records are compared.

Results

We find that alpha activity, K complexes, sleep spindles and slow wave sleep can be visually distinguished using earelectroencephalography. Spectral peaks are shared between the two records. Hypnograms are 90.9% similar.

Conclusion

The results indicate that ear-electroencephalography can be used for sleep staging.

Multi-channel EEG-based sleep stage classification with joint collaborative representation and multiple kernel learning

BACKGROUND

Electroencephalography (EEG) based sleep staging is commonly used in clinical routine. Feature extraction and representation plays a crucial role in EEG-based automatic classification of sleep stages. Sparse representation (SR) is a state-of-the-art unsupervised feature learning method suitable for EEG feature representation.

NEW METHOD

Collaborative representation (CR) is an effective data coding method used as a classifier. Here we use CR as a data representation method to learn features from the EEG signal. A joint collaboration model is established to develop a multi-view learning algorithm, and generate joint CR (JCR) codes to fuse and represent multi-channel EEG signals. A two-stage multi-view learning-based sleep staging framework is then constructed, in which JCR and joint sparse representation (JSR) algorithms first fuse and learning the feature representation from multi-channel EEG signals, respectively. Multi-view JCR and JSR features are then integrated and sleep stages recognized by a multiple kernel extreme learning machine (MK-ELM) algorithm with grid search.

RESULTS

The proposed two-stage multi-view learning algorithm achieves superior performance for sleep staging. With a K-means clustering based dictionary, the mean classification accuracy, sensitivity and specificity are 81.10 ± 0.15%, 71.42 ± 0.66% and 94.57 ± 0.07%, respectively; while with the dictionary learned using the submodular optimization method, they are 80.29 ± 0.22%, 71.26 ± 0.78% and 94.38 ± 0.10%, respectively.

COMPARISON WITH EXISTING METHODS

The two-stage multi-view learning based sleep staging framework outperforms all other classification methods compared in this work, while JCR is superior to JSR.

CONCLUSIONS

The proposed multi-view learning framework has the potential for sleep staging based on multi-channel or multi-modality polysomnography signals.

Inter-expert and intra-expert reliability in sleep spindle scoring

OBJECTIVES

To measure the inter-expert and intra-expert agreement in sleep spindle scoring, and to quantify how many experts are needed to build a reliable dataset of sleep spindle scorings.

METHODS

The EEG dataset was comprised of 400 randomly selected 115s segments of stage 2 sleep from 110 sleeping subjects in the general population (57±8, range: 42-72 years). To assess expert agreement, a total of 24 Registered Polysomnographic Technologists (RPSGTs) scored spindles in a subset of the EEG dataset at a single electrode location (C3-M2). Intra-expert and inter-expert agreements were calculated as F1-scores, Cohen's kappa (κ), and intra-class correlation coefficient (ICC).

RESULTS

We found an average intra-expert F1-score agreement of 72±7% (κ: 0.66±0.07). The average inter-expert agreement was 61±6% (κ: 0.52±0.07). Amplitude and frequency of discrete spindles were calculated with higher reliability than the estimation of spindle duration. Reliability of sleep spindle scoring can be improved by using qualitative confidence scores, rather than a dichotomous yes/no scoring system.

CONCLUSIONS

We estimate that 2-3 experts are needed to build a spindle scoring dataset with 'substantial' reliability (κ: 0.61-0.8), and 4 or more experts are needed to build a dataset with 'almost perfect' reliability (κ: 0.81-1).

SIGNIFICANCE

Spindle scoring is a critical part of sleep staging, and spindles are believed to play an important role in development, aging, and diseases of the nervous system.