Automatic Detection of Chronic Insomnia from Polysomnographic and Clinical Variables Using Machine Learning.

Insomnia is defined subjectively by the presence and frequency of specific clinical symptoms and an association with distress. Although sleep study data has shown some weak associations, no objective test can currently be used to predict insomnia. The purpose of this study was to use previously reported and relatively crafted insomnia-related polysomnographic variables in machine learning models to classify groups with and without insomnia. Demographics, diagnosed depression, Epworth Sleepiness Scale (ESS), and features derived from electroencephalography (EEG), arousals, and sleep stages from 3,407 sleep clinic patients (2,617 without insomnia and 790 insomnia patients based on responses to a set of questions) were included in this analysis. The number of features were reduced using pair-wise correlation and recursive feature elimination. Predictive value of three machine learning models (logistic regression, neural network, and support vector machine) was investigated, and the best performance was achieved with logistic regression, yielding a balanced accuracy of 71%. The most important features in predicting insomnia were depression, age, sex, duration of longest arousal, ESS score, and EEG power in theta and sigma bands across all sleep stages. Results indicate potential of machine learning-based screening for insomnia using clinical variables and EEG.

Automatic scoring of drug-induced sleep endoscopy for obstructive sleep apnea using deep learning.

BACKGROUND: Treatment of obstructive sleep apnea is crucial for long term health and reduced economic burden. For those considered for surgery, drug-induced sleep endoscopy (DISE) is a method to characterize location and pattern of sleep-related upper airway collapse. According to the VOTE classification system, four upper airway sites of collapse are characterized: velum (V), oropharynx (O), tongue (T), and epiglottis (E). The degree of obstruction per site is classified as 0 (no obstruction), 1 (partial obstruction), or 2 (complete obstruction). Here we propose a deep learning approach for automatic scoring of VOTE obstruction degrees from DISE videos. METHODS: We included 281 DISE videos with varying durations (6 s-16 min) from two sleep clinics: Copenhagen University Hospital and Stanford University Hospital. Examinations were split into 5-s clips, each receiving annotations of 0, 1, 2, or X (site not visible) for each site (V, O, T, and E), which was used to train a deep learning model. Predicted VOTE obstruction degrees per examination was obtained by taking the highest predicted degree per site across 5-s clips, which was evaluated against VOTE degrees annotated by surgeons. RESULTS: Mean F1 score of 70% was obtained across all DISE examinations (V: 85%, O: 72%, T: 57%, E: 65%). For each site, sensitivity was highest for degree 2 and lowest for degree 0. No bias in performance was observed between videos from different clinicians/hospitals. CONCLUSIONS: This study demonstrates that automating scoring of DISE examinations show high validity and feasibility in degree of upper airway collapse.

Detection of Cheyne-Stokes Breathing using a transformer-based neural network.

Annotation of sleep disordered breathing, including Cheyne-Stokes Breathing (CSB), is an expensive and time-consuming process for the clinician. To solve the problem, this paper presents a deep learning-based algorithm for automatic sample-wise detection of CSB in nocturnal polysomnographic (PSG) recordings. 523 PSG recordings were retrieved from four different sleep cohorts and subsequently scored for CSB by three certified sleep technicians. The data was pre-processed and 16 time domain features were extracted and passed into a neural network inspired by the transformer unit. Finally, the network output was post-processed to achieve physiologically meaningful predictions. The algorithm reached a F1-score of 0.76, close to the certified sleep technicians showing that it is possible to automatically detect CSB with the proposed model. The algorithm had difficulties distinguishing between severe obstructive sleep apnea and CSB but this was not dissimilar to technician performance. In conclusion, the proposed algorithm showed promising results and a confirmation of the performance could make it relevant as a screening tool in a clinical setting.

Proteomic Biomarkers of the Apnea Hypopnea Index and Obstructive Sleep Apnea: Insights into the Pathophysiology of Presence, Severity, and Treatment Response.

Obstructive sleep apnea (OSA), a disease associated with excessive sleepiness and increased cardiovascular risk, affects an estimated 1 billion people worldwide. The present study examined proteomic biomarkers indicative of presence, severity, and treatment response in OSA. Participants (n = 1391) of the Stanford Technology Analytics and Genomics in Sleep study had blood collected and completed an overnight polysomnography for scoring the apnea−hypopnea index (AHI). A highly multiplexed aptamer-based array (SomaScan) was used to quantify 5000 proteins in all plasma samples. Two separate intervention-based cohorts with sleep apnea (n = 41) provided samples pre- and post-continuous/positive airway pressure (CPAP/PAP). Multivariate analyses identified 84 proteins (47 positively, 37 negatively) associated with AHI after correction for multiple testing. Of the top 15 features from a machine learning classifier for AHI ≥ 15 vs. AHI < 15 (Area Under the Curve (AUC) = 0.74), 8 were significant markers of both AHI and OSA from multivariate analyses. Exploration of pre- and post-intervention analysis identified 5 of the 84 proteins to be significantly decreased following CPAP/PAP treatment, with pathways involving endothelial function, blood coagulation, and inflammatory response. The present study identified PAI-1, tPA, and sE-Selectin as key biomarkers and suggests that endothelial dysfunction and increased coagulopathy are important consequences of OSA, which may explain the association with cardiovascular disease and stroke.

Automatic Segmentation to Cluster Patterns of Breathing in Sleep Apnea.

Annotation of polysomnography (PSG) recordings for diagnosis of obstructive sleep apnea (OSA) is a standard procedure but an expensive and time-consuming process for clinicians. To aid clinicians in this process we present a data driven unsupervised hierarchical clustering approach for detection and visual presentation of breathing patterns in PSG recordings. The aim was to develop a model independent of manual annotations to detect and visualize respiratory events related to OSA. 10 recordings from the Sleep Heart Health Study database were used, and the proposed algorithm was evaluated based on the manually annotated events for each recording. The algorithm reached an F1-score of 0.58 across the 10 recordings when detecting the presence of an event vs. no event and a 100% correct diagnosis prediction of OSA when predicting if apnea-hypopnea index (AHI) ≥ 15, which is a clinically meaningful cut-off. The F1-score may be due to imprecise placement of events, difficulty distinguishing between hypopneas and stable breathing, and variations in scoring. In conclusion the performance can be improved despite the strong agreement in diagnostics. The method is a proof of concept that a clustering method can detect and visualize breathing patterns related to OSA while maintaining a correct diagnosis.

Polysomnographic Plethysmography Excursions are Reduced in Obese Elderly Men.

Sleep apnea is a widespread disorder and is defined by the complete or partial cessation of breathing. Obstructive sleep apnea (OSA) is caused by an obstruction in the upper airway while central sleep apnea (CSA) is characterized by a diminished or absent respiratory effort. It is crucial to differentiate between these respiratory subtypes as they require radically different treatments. Currently, diagnostic polysomnography (PSG) is used to determine respiratory thoracic and abdominal movement patterns using plethysmography belt signals, to distinguish between OSA and CSA. There is significant manual technician interrater variability between these classifications, especially in the evaluation of CSA. We hypothesize that an increased body mass index (BMI) will cause decreased belt signal excursions that increase false scorings of CSA. The hypothesis was investigated by calculating the envelope as a continuous signal of belt signals in 2833 subjects from the MrOS Sleep Study and extracting a mean value of each of the envelopes for each subject. Using linear regression, we found that an increased BMI was associated with lower excursions during REM sleep (-0.013 [mV] thoracic and -0.018 [mV] abdominal, per BMI) and non-REM (-0.014 [mV] thoracic and -0.012 [mV] abdominal, per BMI). We conclude that increased BMI leads to lower excursions in the belt signals during event-free sleep, and that OSA and CSA events are harder to distinguish in subjects with high BMI. This has a major implication for the correct identification of CSA/OSA and its treatment.

Upper Airway Classification in Sleep Endoscopy Examinations using Convolutional Recurrent Neural Networks.

Assessing the upper airway (UA) of obstructive sleep apnea patients using drug-induced sleep endoscopy (DISE) before potential surgery is standard practice in clinics to determine the location of UA collapse. According to the VOTE classification system, UA collapse can occur at the velum (V), oropharynx (O), tongue (T), and/or epiglottis (E). Analyzing DISE videos is not trivial due to anatomical variation, simultaneous UA collapse in several locations, and video distortion caused by mucus or saliva. The first step towards automated analysis of DISE videos is to determine which UA region the endoscope is in at any time throughout the video: V (velum) or OTE (oropharynx, tongue, or epiglottis). An additional class denoted X is introduced for times when the video is distorted to an extent where it is impossible to determine the region. This paper is a proof of concept for classifying UA regions using 24 annotated DISE videos. We propose a convolutional recurrent neural network using a ResNet18 architecture combined with a two-layer bidirectional long short-term memory network. The classifications were performed on a sequence of 5 seconds of video at a time. The network achieved an overall accuracy of 82% and F1-score of 79% for the three-class problem, showing potential for recognition of regions across patients despite anatomical variation. Results indicate that large-scale training on videos can be used to further predict the location(s), type(s), and degree(s) of UA collapse, showing potential for derivation of automatic diagnoses from DISE videos eventually.

Estimation of Apnea-Hypopnea Index Using Deep Learning On 3-D Craniofacial Scans.

Obstructive sleep apnea (OSA) is characterized by decreased breathing events that occur through the night, with severity reported as the apnea-hypopnea index (AHI), which is associated with certain craniofacial features. In this study, we used data from 1366 patients collected as part of Stanford Technology Analytics and Genomics in Sleep (STAGES) across 11 US and Canadian sleep clinics and analyzed 3D craniofacial scans with the goal of predicting AHI, as measured using gold standard nocturnal polysomnography (PSG). First, the algorithm detects pre-specified landmarks on mesh objects and aligns scans in 3D space. Subsequently, 2D images and depth maps are generated by rendering and rotating scans by 45-degree increments. Resulting images were stacked as channels and used as input to multi-view convolutional neural networks, which were trained and validated in a supervised manner to predict AHI values derived from PSGs. The proposed model achieved a mean absolute error of 11.38 events/hour, a Pearson correlation coefficient of 0.4, and accuracy for predicting OSA of 67% using 10-fold cross-validation. The model improved further by adding patient demographics and variables from questionnaires. We also show that the model performed at the level of three sleep medicine specialists, who used clinical experience to predict AHI based on 3D scan displays. Finally, we created topographic displays of the most important facial features used by the model to predict AHI, showing importance of the neck and chin area. The proposed algorithm has potential to serve as an inexpensive and efficient screening tool for individuals with suspected OSA.

Prediction of Patient Demographics using 3D Craniofacial Scans and Multi-view CNNs.

3D data is becoming increasingly popular and accessible for computer vision tasks. A popular format for 3D data is the mesh format, which can depict a 3D surface accurately and cost-effectively by connecting points in the (x, y, z) plane, known as vertices, into triangles that can be combined to approximate geometrical surfaces. However, mesh objects are not suitable for standard deep learning techniques due to their non-euclidean structure. We present an algorithm which predicts the sex, age, and body mass index of a subject based on a 3D scan of their face and neck. This algorithm relies on an automatic pre-processing technique, which renders and captures the 3D scan from eight different angles around the x-axis in the form of 2D images and depth maps. Subsequently, the generated data is used to train three convolutional neural networks, each with a ResNet18 architecture, to learn a mapping between the set of 16 images per subject (eight 2D images and eight depth maps from different angles) and their demographics. For age and body mass index, we achieved a mean absolute error of 7.77 years and 4.04 kg/m2 on the respective test sets, while Pearson correlation coefficients of 0.76 and 0.80 were obtained, respectively. The prediction of sex yielded an accuracy of 93%. The developed framework serves as a proof of concept for prediction of more clinically relevant variables based on 3D craniofacial scans stored in mesh objects.

Non-invasive machine learning estimation of effort differentiates sleep-disordered breathing pathology.

OBJECTIVE: Obstructive sleep-disordered breathing (SDB) events, unlike central events, are associated with increased respiratory effort. Esophageal pressure (P es) monitoring is the gold standard for measuring respiratory effort, but it is typically poorly tolerated because of its invasive nature. The objective was to investigate whether machine learning can be applied to routinely collected non-invasive, polysomnography (PSG) measures to accurately model peak negative P es. APPROACH: One thousand one hundred and nineteen patients from the Stanford Sleep Clinic with PSGs containing P es served as the sample. The selected non-invasive PSG signals included nasal pressure, oral airflow, thoracoabdominal effort, and snoring. A long short-term memory neural network was implemented to achieve a context-based mapping between the non-invasive features and the P es values. A hold-out dataset served as a prospective validation of the algorithm without needing to undertake a costly new study with the impractically invasive P es. MAIN RESULTS: The median difference between the measured and predicted P es was 0.61 cmH2O with an interquartile range (IQR) of 2.99 cmH2O and 5th and 95th percentiles of -5.85 cmH2O and 5.47 cmH2O, respectively. The model performed well when compared to actual esophageal pressure signal (ρ median = 0.581, p  = 0.01; IQR = 0.298; ρ 5% = 0.106; ρ 95% = 0.843). SIGNIFICANCE: A significant difference in predicted P es was shown between normal breathing and all obstructive SDB events; whereas, central apneas did not significantly differ from normal breathing. The developed system may be used as a tool for quantifying respiratory effort from the existing clinical practice of PSG without the need for P es, improving characterization of SDB events as obstructive or not.

Quantification of muscle activity during sleep for patients with neurodegenerative diseases.

Idiopathic REM sleep behavior disorder (iRBD) is a very strong predictor for later development of Parkinson's disease (PD), and is characterized by REM sleep without atonia (RSWA), resulting in increased muscle activity during REM sleep. Abundant studies have shown the loss of atonia during REM sleep, but our aim was to investigate whether iRBD and PD patients have increased muscle activity in both REM and NREM sleep compared to healthy controls. This was achieved by developing a semi-automatic algorithm for quantification of mean muscle activity per second during all sleep stages for the enrolled patients. The three groups examined included patients suffering from iRBD, PD and healthy control subjects (CO). To determine muscle activity, a baseline and threshold were established after pre-processing of the raw surface electromyography (sEMG) signal. The signal was then segmented according to the different sleep stages and muscle activity beyond the threshold was counted. The results were evaluated statistically using the two-sided Mann-Whitney U-test. The results suggested that iRBD patients also exhibit distinctive muscle activity characteristics in NREM sleep, however not as evident as in REM sleep, leading to the conclusion that RSWA still is the most distinct characteristic of RBD. Furthermore, the muscle activity of PD patients was comparable to that of controls with only slightly elevated amplitudes.