Automatic detection of non-apneic sleep arousal regions from polysomnographic recordings
Introduction
Sleep disorders are implicated in some health problems [1]. American Academy of Sleep Medicine (AASM) defined sleep arousals as abrupt shifts of electroencephalography (EEG) frequency which may include theta, alpha and/or frequencies above 16 Hz but not spindles [2] that last at least 3 s, with at least 10 s of previous stable sleep [3]. Respiratory effort-related arousal (RERA) is a type of arousal characterized by mild upper airway narrowing during sleep, with increased respiratory effort required to maintain a slightly reduced level of airflow (not large enough to be scored as hypopnea).
In most cases, the studies of RERA patients have been largely observational. RERA is preferably recorded with esophageal or nasal manometry, both of these methods are invasive. Patients with RERA often present with a functional somatic syndrome misdiagnosis, including migraine/tension headaches, irritable bowel syndrome, chronic fatigue syndrome, temporomandibular joint (TMJ) syndrome, TMD, and fibromyalgia. Many young women with sleep-disordered breathing are mistakenly treated with hypnotics, antidepressants, pain medication, attention deficit/hyperactivity disorder medication, eugeroics, and muscle relaxants.
Manually scoring, the above mentioned sleep disorders is a hard, time-consuming, subjective, costly, and error prone task [4]. The manual sleep annotation is strongly dependent on not only the standards and guidelines that have been published and revised during the years by AASM but also on consensus and adjudication especially in the absence of high-level evidence. Moreover, there are some optional recommendations in sleep annotation that covered rules which do not need to be followed but maybe at the discretion of the clinician or investigator. These properties cause the sleep annotation to be dependent on the scorer. Furthermore, as awareness among the public and health care providers has increased, referrals for sleep disorder diagnosis have increased. As a result, the waiting time for diagnosis has grown significantly. Also, the lack of trained technologists capable of carrying out proper in-lab neuro-cardio-pulmonary monitoring is a major barrier, especially in developing countries. So, many people with sleep disorders remain undiagnosed and untreated. Therefore, a method that can facilitate diagnosis of RERA automatically can be useful.
Intending to detect the origin of sleep disorders, PhysioNet Computing in Cardiology (CinC) held a challenge in 2018 (hereafter called the challenge) to develop automated methods to classify sleep states into the normal sleep, occurrence of Apnea, and Non-Apnea that labeled with 0, 1, and 1 respectively. It is noted that in this study based on the challenge data, detection of Non-Apnea events is pivoted on detecting the RERA events. Such methods can act as criteria for diagnosis as well as being completely in line with polysomnography (PSG) empirical evidence and with less deviation in results. Briefly, this study aims to detect regions of PSG where the RERA events occurred. Hereafter these regions are called target regions.
Based on the challenge some studies have been conducted, deploying a variety of methods from generalized linear models to deep neural networks. Since a change in various physiological parameters usually arises in the period leading to the arousals, recurrent neural network (RNN) and long short-term memory (LSTM) networks seem to be appropriate approaches to classify the arousals regions automatically. Unidirectional recurrent neural network (URNN) and bidirectional recurrent neural network (BRNN) are the networks that are used in the majority of the studies. Warrick et al. [5] used an RNN architecture for sequence learning using three layers of LSTM. Warrick introduced a different strategy by using a network architecture consists of two components and the scattering transform of raw signals were used as RNN inputs. The proposed structure by Howe-Patterson et al. [6] is composed of multiple dense convolutional units (DCU), bidirectional LSTM layers, and a softmax output layer. In contrast to the above-mentioned study that used preprocessed PSG signals as the input of networks, Marinosson et al. [7] derived time and frequency domain features from the PSG signals and the features were fed into a BRNN, using LSTM units (BRNN-LSTM). The results of this study are compared to Marinosson et al. [7] method. He et al. [8] used a sequence-to-sequence deep neural networks (DNNs) which consists of a series of convolutional layers with residual connections followed by an LSTM layer and two fully connected layers and preprocessed signals were fed into the network. Patane et al. [9] used a deep ensemble neural network architecture. In their study, the merged feature space was fed through a Siamese architecture. In the majority of presented studies, Samples labeled 1 were discarded [10], [11] or re-labeled as 0. Howe-Patterson et al. [6] used a multi-task mechanism and relabeled regions, based on correlated regions.
In the current study, it has been endeavored to put forward a method to find an optimal subset of features to precisely detect the target regions, utilizing filter, wrapper, and embedded methods of feature selection. Besides, much focus has been placed on reducing the dimension of feature space without increasing the misclassification error. In this way, the training process of such a large data became more feasible in terms of computation and training time. Furthermore, by this means multicollinearity and overfitting risk can be avoided. For this purpose, the multi-objectives genetic algorithm (MOGA) was applied as the optimization technique to reduce misclassification error and feature space dimension simultaneously. To overcome the unbalanced data problem, a combination of techniques was implemented. Aiming to create a robust meta-model two algorithms of boosting along with a Bootstrap Aggregating (Bagging) algorithm were employed and their performances were compared.
The rest of the paper is organized as follows. Section 2 demonstrates the dataset and explains the method. The results are brought up in Section 3 and the discussion is discussed in Section 4. The conclusions are summarized in Section 5.
Section snippets
Materials and methods
This section is composed of six subsections. At first, the bio-signals used in this study and preprocessing procedures are described. Then the methods of feature selection are introduced in detail. Finally, the prediction models used to predict the results are explained.
Results
Figs. 2 and 3 depict some of the most informative and discriminative selected features for the EEG of individual , related to channel C3-M2 and C4-M1 in the interval of 1140-1365e In this interval, all of the 1, 1, and 0 events have occurred for 75 s successively.
The number of extracted and selected features through the feature selection procedure using PSG signals are shown in Table 1. In this table, EF indicates the number of extracted features while FE stands for the number of
Discussion
Many common applications of predictive analytics arise from complex relationships between features. Large features and a limited number of samples can cause a classifier to be prone to be overfitted. Feature selection can decrease the complexity of a model so this study relied heavily on the features subset selection. Using interpretable features in the learning process is the advantage of the approach used in the present study. Here interpretability, yield more confidence in the solution and
Conclusion
In this study, we have presented a sleep arousal detector algorithm for polysomnographic data. As the first step, we generate hand-crafted intermediate features. This study illustrates the potential of physiology driven feature selection for machine learning problems in biomedical signal processing. As is illustrated in Figs. 2 and 3, automatic detection of RERAs and Non-RERA-Non-Apneas of sleep arousals are a subtle task and variable in appearance. This causes the task to be a challenge. The
Authors’ contribution
Jamileh Karimi: conceptualization, methodology, software, validation, writing – original draft, writing – review & editing. Babak Mohammadzadeh Asl: conceptualization, methodology, validation, writing – review & editing, supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (32)
- et al.
A comparative review on sleep stage classification methods in patients and healthy individuals
Comput. Methods Programs Biomed.
(2017) - et al.
Dynamics of sleep: exploring critical transitions and early warning signals
Comput. Methods Programs Biomed.
(2020) - et al.
An algorithm for seizure onset detection using intracranial EEG
Epilepsy Behav.
(2011) - et al.
R-ensembler: a greedy rough set based ensemble attribute selection algorithm with kNN imputation for classification of medical data
Comput. Methods Programs Biomed.
(2020) - et al.
Application of genetic algorithm for hemodialysis schedule optimization
Comput. Methods Programs Biomed.
(2017) - et al.
Polysomnography for obstructive sleep apnea should include arousal-based scoring: an American academy of sleep medicine position statement
J. Clin. Sleep Med.
(2018) - et al.
A review of obstructive sleep apnea detection approaches
IEEE J. Biomed. Health Informatics
(2018) - et al.
Sleep arousal detection from polysomnography using the scattering transform and recurrent neural networks
- et al.
Automated detection of sleep arousals from polysomnography data using a dense convolutional neural network
- et al.
Automatic detection of target regions of respiratory effort-related arousals using recurrent neural networks
Identification of arousals with deep neural networks (DNNs) using different physiological signals
Automated recognition of sleep arousal using multimodal and personalized deep ensembles of neural networks
Effectiveness of a convolutional neural network in sleep arousal classification using multiple physiological signals
Deep learning with convolutional neural networks for sleep arousal detection
Atlas of Clinical Sleep Medicine E-Book: Expert Consult-Online
You snooze, you win: the physionet/computing in cardiology challenge 2018
Cited by (3)
Respiratory analysis during sleep using a chest-worn accelerometer: A machine learning approach
2022, Biomedical Signal Processing and ControlCitation Excerpt :Furthermore, the unnatural sleeping environment can cause white-coat syndrome and first-night effects, thus masking the authenticity of regular sleep experienced at home [4]. Unattended PSGs or routines with a reduced number of channels, such as respiratory polygraphy (RP), and automatic processing routines allow for less expensive home assessments [5,6]. These assessments, however, normally encompass only a single night of recording, thereby neglecting the night-to-night variability present in natural sleep [7].
Sleep arousal detection for monitoring of sleep disorders using one-dimensional convolutional neural network-based U-Net and bio-signals
2024, Data Technologies and Applications