Julie Anja Engelhard Christensen

Automatic sleep classification using a data-driven topic model reveals latent sleep states

BACKGROUND The golden standard for sleep classification uses manual scoring of polysomnography despite points of criticism such as oversimplification, low inter-rater reliability and the standard being designed on young and healthy subjects. NEW METHOD To meet the criticism and reveal the latent sleep states, this study developed a general and automatic sleep classifier using a data-driven approach. Spectral EEG and EOG measures and eye correlation in 1s windows were calculated and each sleep epoch was expressed as a mixture of probabilities of latent sleep states by using the topic model Latent Dirichlet Allocation. Model application was tested on control subjects and patients with periodic leg movements (PLM) representing a non-neurodegenerative group, and patients with idiopathic REM sleep behavior disorder (iRBD) and Parkinsons Disease (PD) representing a neurodegenerative group. The model was optimized using 50 subjects and validated on 76 subjects. RESULTS The optimized sleep model used six topics, and the topic probabilities changed smoothly during transitions. According to the manual scorings, the model scored an overall subject-specific accuracy of 68.3 ± 7.44 (% μ ± σ) and group specific accuracies of 69.0 ± 4.62 (control), 70.1 ± 5.10 (PLM), 67.2 ± 8.30 (iRBD) and 67.7 ± 9.07 (PD). COMPARISON WITH EXISTING METHOD Statistics of the latent sleep state content showed accordances to the sleep stages defined in the golden standard. However, this study indicates that sleep contains six diverse latent sleep states and that state transitions are continuous processes. CONCLUSIONS The model is generally applicable and may contribute to the research in neurodegenerative diseases and sleep disorders.

Data-driven modeling of sleep EEG and EOG reveals characteristics indicative of pre-Parkinson's and Parkinson's disease.

BACKGROUND Manual scoring of sleep relies on identifying certain characteristics in polysomnograph (PSG) signals. However, these characteristics are disrupted in patients with neurodegenerative diseases. NEW METHOD This study evaluates sleep using a topic modeling and unsupervised learning approach to identify sleep topics directly from electroencephalography (EEG) and electrooculography (EOG). PSG data from control subjects were used to develop an EOG and an EEG topic model. The models were applied to PSG data from 23 control subjects, 25 patients with periodic leg movements (PLMs), 31 patients with idiopathic REM sleep behavior disorder (iRBD) and 36 patients with Parkinsons disease (PD). The data were divided into training and validation datasets and features reflecting EEG and EOG characteristics based on topics were computed. The most discriminative feature subset for separating iRBD/PD and PLM/controls was estimated using a Lasso-regularized regression model. RESULTS The features with highest discriminability were the number and stability of EEG topics linked to REM and N3, respectively. Validation of the model indicated a sensitivity of 91.4% and a specificity of 68.8% when classifying iRBD/PD patients. COMPARISON WITH EXISTING METHOD The topics showed visual accordance with the manually scored sleep stages, and the features revealed sleep characteristics containing information indicative of neurodegeneration. CONCLUSIONS This study suggests that the amount of N3 and the ability to maintain NREM and REM sleep have potential as early PD biomarkers. Data-driven analysis of sleep may contribute to the evaluation of neurodegenerative patients.

Validation of a novel automatic sleep spindle detector with high performance during sleep in middle aged subjects

Many of the automatic sleep spindle detectors currently used to analyze sleep EEG are either validated on young subjects or not validated thoroughly. The purpose of this study is to develop and validate a fast and reliable sleep spindle detector with high performance in middle aged subjects. An automatic sleep spindle detector using a bandpass filtering approach and a time varying threshold was developed. The validation was done on sleep epochs from EEG recordings with manually scored sleep spindles from 13 healthy subjects with a mean age of 57.9 ± 9.7 years. The sleep spindle detector reached a mean sensitivity of 84.6% and a mean specificity of 95.3%. The sleep spindle detector can be used to obtain measures of spindle count and density together with quantitative measures such as the mean spindle frequency, mean spindle amplitude, and mean spindle duration.

Sleep stability and transitions in patients with idiopathic REM sleep behavior disorder and patients with Parkinson's disease.

OBJECTIVE Patients with idiopathic rapid eye movement (REM) sleep behavior disorder (iRBD) are at high risk of developing Parkinsons disease (PD). As wake/sleep-regulation is thought to involve neurons located in the brainstem and hypothalamic areas, we hypothesize that the neurodegeneration in iRBD/PD is likely to affect wake/sleep and REM/non-REM (NREM) sleep transitions. METHODS We determined the frequency of wake/sleep and REM/NREM sleep transitions and the stability of wake (W), REM and NREM sleep as measured by polysomnography (PSG) in 27 patients with PD, 23 patients with iRBD, 25 patients with periodic leg movement disorder (PLMD) and 23 controls. Measures were computed based on manual scorings and data-driven labeled sleep staging. RESULTS Patients with PD showed significantly lower REM stability than controls and patients with PLMD. Patients with iRBD had significantly lower REM stability compared with controls. Patients with PD and RBD showed significantly lower NREM stability and significantly more REM/NREM transitions than controls. CONCLUSIONS We conclude that W, NREM and REM stability and transitions are progressively affected in iRBD and PD, probably reflecting the successive involvement of brain stem areas from early on in the disease. SIGNIFICANCE Sleep stability and transitions determined by a data-driven approach could support the evaluation of iRBD and PD patients.

Detection of K-complexes based on the wavelet transform.

Sleep scoring needs computational assistance to reduce execution time and to assure high quality. In this pilot study a semi-automatic K-Complex detection algorithm was developed using wavelet transformation to identify pseudo-K-Complexes and various feature thresholds to reject false positives. The algorithm was trained and tested on sleep EEG from two databases to enhance its general applicability. When testing on data from subjects from the DREAMS© database, a mean true positive rate of 74 % and a positive predictive value of 65 % were achieved. After adjusting a few thresholds to adapt to the second database, the Danish Center for Sleep Medicine, a similar performance was achieved. The algorithm performs at the level of the State of the Art and surpasses the inter-rater agreement rate.

Classification of iRBD and Parkinson's disease patients based on eye movements during sleep

Patients suffering from the sleep disorder idiopathic rapid-eye-movement sleep behavior disorder (iRBD) have been observed to be in high risk of developing Parkinsons disease (PD). This makes it essential to analyze them in the search for PD biomarkers. This study aims at classifying patients suffering from iRBD or PD based on features reflecting eye movements (EMs) during sleep. A Latent Dirichlet Allocation (LDA) topic model was developed based on features extracted from two electrooculographic (EOG) signals measured as parts in full night polysomnographic (PSG) recordings from ten control subjects. The trained model was tested on ten other control subjects, ten iRBD patients and ten PD patients, obtaining a EM topic mixture diagram for each subject in the test dataset. Three features were extracted from the topic mixture diagrams, reflecting “certainty”, “fragmentation” and “stability” in the timely distribution of the EM topics. Using a Naive Bayes (NB) classifier and the features “certainty” and “stability” yielded the best classification result and the subjects were classified with a sensitivity of 95 %, a specificity of 80% and an accuracy of 90 %. This study demonstrates in a data-driven approach, that iRBD and PD patients may exhibit abnorm form and/or timely distribution of EMs during sleep.

Separation of Parkinson's patients in early and mature stages from control subjects using one EOG channel

In this study, polysomnographic left side EOG signals from ten control subjects, ten iRBD patients and ten Parkinsons patients were decomposed in time and frequency using wavelet transformation. A total of 28 features were computed as the means and standard deviations in energy measures from different reconstructed detail subbands across all sleep epochs during a whole night of sleep. A subset of features was chosen based on a cross validated Shrunken Centroids Regularized Discriminant Analysis, where the controls were treated as one group and the patients as another. Classification of the subjects was done by a leave-one-out validation approach using same method, and reached a sensitivity of 95%, a specificity of 70% and an accuracy of 86.7%. It was found that in the optimal subset of features, two hold lower frequencies reflecting the rapid eye movements and two hold higher frequencies reflecting EMG activity. This study demonstrates that both analysis of eye movements during sleep as well as EMG activity measured at the EOG channel hold potential of being biomarkers for Parkinsons disease.

Automatic detection of REM sleep in subjects without atonia

Idiopathic Rapid-Eye-Movement (REM) sleep Behavior Disorder (iRBD) is a strong early marker of Parkinsons Disease and is characterized by REM sleep without atonia (RSWA) and increased phasic muscle activity. Current proposed methods for detecting RSWA assume the presence of a manually scored hypnogram. In this study a full automatic REM sleep detector, using the EOG and EEG channels, is proposed. Based on statistical features, combined with subject specific feature scaling and post-processing of the classifier output, it was possible to obtain an mean accuracy of 0.96 with a mean sensititvity and specificity of 0.94 and 0.96 respectively.

Classification of iRBD and Parkinson's patients using a general data-driven sleep staging model built on EEG

Sleep analysis is an important diagnostic tool for sleep disorders. However, the current manual sleep scoring is time-consuming as it is a crude discretization in time and stages. This study changes Esbroeck and Westovers [1] latent sleep staging model into a global model. The proposed data-driven method trained a topic mixture model on 10 control subjects and was applied on 10 other control subjects, 10 iRBD patients and 10 Parkinsons patients. In that way 30 topic mixture diagrams were obtained from which features reflecting distinct sleep architectures between control subjects and patients were extracted. Two features calculated on basis of two latent sleep states classified subjects as “control” or “patient” by a simple clustering algorithm. The mean sleep staging accuracy compared to classical AASM scoring was 72.4% for control subjects and a clustering of the derived features resulted in a sensitivity of 95% and a specificity of 80%. This study demonstrates that frequency analysis of sleep EEG can be used for data-driven global sleep classification and that topic features separates iRBD and Parkinsons patients from control subjects.

Sleep spindle density in narcolepsy

BACKGROUND Patients with narcolepsy type 1 (NT1) show alterations in sleep stage transitions, rapid-eye-movement (REM) and non-REM sleep due to the loss of hypocretinergic signaling. However, the sleep microstructure has not yet been evaluated in these patients. We aimed to evaluate whether the sleep spindle (SS) density is altered in patients with NT1 compared to controls and patients with narcolepsy type 2 (NT2). METHODS All-night polysomnographic recordings from 28 NT1 patients, 19 NT2 patients, 20 controls (C) with narcolepsy-like symptoms, but with normal cerebrospinal fluid hypocretin levels and multiple sleep latency tests, and 18 healthy controls (HC) were included. Unspecified, slow, and fast SS were automatically detected, and SS densities were defined as number per minute and were computed across sleep stages and sleep cycles. The between-cycle trends of SS densities in N2 and NREM sleep were evaluated within and between groups. RESULTS Between-group comparisons in sleep stages revealed no significant differences in any type of SS. Within-group analyses of the SS trends revealed significant decreasing trends for NT1, HC, and C between first and last sleep cycle. Between-group analyses of SS trends between first and last sleep cycle revealed that NT2 differ from NT1 patients in the unspecified SS density in NREM sleep, and from HC in the slow SS density in N2 sleep. CONCLUSIONS SS activity is preserved in NT1, suggesting that the ascending neurons to thalamic activation of SS are not significantly affected by the hypocretinergic system. NT2 patients show an abnormal pattern of SS distribution.