Table 2 - uploaded by Sirvan Khalighi
Content may be subject to copyright.
Source publication
To facilitate the performance comparison of new methods for sleep patterns analysis, datasets with quality content, publicly-available, are very important and useful. We introduce an open-access comprehensive sleep dataset, called ISRUC-Sleep. The data were obtained from human adults, including healthy subjects, subjects with sleep disorders, and s...
Contexts in source publication
Context 1
... the need and usefulness of publicly available sleep datasets, which can be used as a common reference for researchers, some sleep-related datasets were developed by sleep research groups. As shown in Table 2, these datasets contain multiple signals from some healthy and patient sub- jects. The sleep datasets of PhysioBank [35] have been used in a few works (see Table 2). ...
Context 2
... shown in Table 2, these datasets contain multiple signals from some healthy and patient sub- jects. The sleep datasets of PhysioBank [35] have been used in a few works (see Table 2). Even though MIT-BIH, Sleep-EDF and Extended Sleep-EDF are general purpose datasets, these do not have enough subjects for generalization purposes. ...
Context 3
... there exist some access restrictions regarding different kinds of information of the dataset. As detailed in Table 2, the sub- groups of this dataset have significant differences in terms of number of channels, filtering methods applied to the signals, acquisition software, annotations, scoring criteria and epoch size. ...
Context 4
... summary, all the dataset detailed in Table 2 have limita- tions in some aspects and as far as we know, except Sleep-EDF dataset (expanded), which were recorded in two subsequent day-night at the subject's home, in the others only one acqui- sition session (one recording) per subject is available. ...
Citations
... Experiments were conducted on two subsets of the ISRUC-Sleep database [25]: ISRUC-S1 was used to evaluate the classification performance of the proposed model, while ISRUC-S3 was utilized to assess its generalization capability. We selected 62 adults with diagnosed sleep disorders from the ISRUC-S1 subgroup (33 males and 29 females, aged 20-85 years). ...
Objective. Obstructive sleep apnea (OSA) is a prevalent sleep disorder. Accurate sleep staging is one of the prerequisites in the study of sleep-related disorders and the evaluation of sleep quality. We introduce a novel GraphSleepFormer (GSF) network designed to effectively capture global dependencies and node characteristics in graph-structured data. Approach. The network incorporates centrality coding and spatial coding into its architecture. It employs adaptive learning of adjacency matrices for spatial encoding between channels located on the head, thereby encoding graph structure information to enhance the model’s representation and understanding of spatial relationships. Centrality encoding integrates the degree matrix into node features, assigning varying degrees of attention to different channels. Ablation experiments demonstrate the effectiveness of these encoding methods. The Shapley Additive Explanations (SHAP) method was employed to evaluate the contribution of each channel in sleep staging, highlighting the necessity of using multimodal data. Main results. We trained our model on overnight polysomnography data collected from 28 OSA patients in a clinical setting and achieved an overall accuracy of 80.10%. GSF achieved performance comparable to state-of-the-art methods on two subsets of the ISRUC database. Significance. The GSF Accurately identifies sleep periods, providing a critical basis for diagnosing and treating OSA, thereby contributing to advancements in sleep medicine.
... In another study, Almuhammadi, Aboalayon & Faezipour (2015), obtained frequency subbands and used energy and entropy values instead of using the entire EEG signals and achieved a success ratio of 97.14%. Khalighi et al. (2016), developed an automatic sleep stage classification model, and four signal channels (two EEG and two EOG) were selected according to their proximity to the electrode locations of the new model tested in the used dataset. The signals were converted into spectrogram images and classified by using a convolutional neural network (CNN) model, supporting not only its performance on a standard PSG dataset but also the transferability of the model to a dataset measured with the new system. ...
In this study, we focus on classifying sleep apnea syndrome by using the spectrograms obtained from electroencephalogram (EEG) signals taken from polysomnography (PSG) recordings and the You Only Look Once (YOLO) v8 deep learning model. For this aim, the spectrograms of segments obtained from EEG signals with different apnea-hypopnea values (AHI) using a 30-s window function are obtained by short-time Fourier transform (STFT). The spectrograms are used as inputs to the YOLOv8 model to classify sleep apnea syndrome as mild, moderate, severe apnea, and healthy. For four-class classification models, the standard reference level is 25%, assuming equal probabilities for all classes or an equal number of samples in each class. In this context, this information is an important reference point for the validity of our study. Deep learning methods are frequently used for the classification of EEG signals. Although ResNet64 and YOLOv5 give effective results, YOLOv8 stands out with fast processing times and high accuracy. In the existing literature, parameter reduction approaches in four-class EEG classification have not been adequately addressed and there are limitations in this area. This study evaluates the performance of parameter reduction methods in EEG classification using YOLOv8, fills gaps in the existing literature for four-class classification, and reduces the number of parameters of the used models. Studies in the literature have generally classified sleep apnea syndrome as binary (apnea/healthy) and ignored distinctions between apnea severity levels. Furthermore, most of the existing studies have used models with a high number of parameters and have been computationally demanding. In this study, on the other hand, the use of spectrograms is proposed to obtain higher correct classification ratios by using more accurate and faster models. The same classification experiments are reimplemented for widely used ResNet64 and YOLOv5 deep learning models to compare with the success of the proposed model. In the implemented experiments, total correct classification (TCC) ratios are 93.7%, 93%, and 88.2% for YOLOv8, ResNet64, and YOLOv5, respectively. These experiments show that the YOLOv8 model reaches higher success ratios than the ResNet64 and YOLOv5 models. Although the TCC ratios of the YOLOv8 and ResNet64 models are comparable, the YOLOv8 model uses fewer parameters and layers than the others, providing a faster processing time and a higher TCC ratio. The findings of the study make a significant contribution to the current state of the art. As a result, this study gives rise to the idea that the YOLOv8 deep learning model can be used as a new tool for classification of sleep apnea syndrome from EEG signals.
... We use the publicly available datasets ISRUC-S3 and Sleep-EDF-20 in this study. The ISRUC-S3 was recorded by Khalighi et al. from 10 healthy subjects during sleep [12]. The recordings were divided into 30second epochs and labelled, resulting in a total of 8,589 annotated samples. ...
Despite significant advances in deep learning-based sleep stage classification, the clinical adoption of automatic classification models remains slow. One key challenge is the lack of explainability, as many models function as black boxes with millions of parameters. In response, recent work has increasingly focussed on enhancing model explainability. This study contributes to these efforts by globally explaining spectral processing of individual EEG channels. Specifically, we introduce a method to retrieve the filter spectrum of low-level convolutional feature extraction and compare it with the classification-relevant spectral information in the data. We evaluate our approach on the MSA-CNN model using the ISRUC-S3 and Sleep-EDF-20 datasets. Our findings show that spectral processing plays a significant role in the lower frequency bands. In addition, comparing the correlation between filter spectrum and data-based spectral information with univariate performance indicates that the model naturally prioritises the most informative channels in a multimodal setting. We specify how these insights can be leveraged to enhance model performance. The code for the filter spectrum retrieval and its analysis is available at https://github.com/sgoerttler/MSA-CNN.
... In this study, we utilized two subsets of the ISRUC-Sleep dataset [39]: ISRUC-S1 and ISRUC-S3. The ISRUC-S3 subset contains overnight PSG recordings from 10 healthy participants (including 9 males and 1 female, with ages ranging from 30 to 58 years), whereas ISRUC-S1 involves 100 individuals with sleep disorders (comprising 55 males and 45 females, aged between 20 and 85 years). ...
For diagnosing mental health conditions and assessing sleep quality, the classification of sleep stages is essential. Although deep learning-based methods are effective in this field, they often fail to capture sufficient features or adequately synthesize information from various sources. For the purpose of improving the accuracy of sleep stage classification, our methodology includes extracting a diverse array of features from polysomnography signals, along with their transformed graph and time-frequency representations. We have developed specific feature extraction modules tailored for each distinct view. To efficiently integrate and categorize the features derived from these different perspectives, we propose a cross-attention fusion mechanism. This mechanism is designed to adaptively merge complex sleep features, facilitating a more robust classification process. More specifically, our strategy includes the development of an efficient fusion network with multi-view features for classifying sleep stages that incorporates brain connectivity and combines both temporal and spectral elements for sleep stage analysis. This network employs a systematic approach to extract spatio-temporal-frequency features and uses cross-attention to merge features from different views effectively. In the experiments we conducted on the ISRUC public datasets, we found that our approach outperformed other proposed methods. In the ablation experiments, there was also a 2% improvement over the baseline model. Our research indicates that multi-view feature fusion methods with a cross-attention mechanism have strong potential in sleep stage classification.
... A summary of the three datasets is given in Table I. The first dataset, ISRUC-S3, was recorded by Khalighi et al. from 10 healthy subjects during sleep [31]. The recordings were divided into 30-second epochs, resulting in a total of 8,589 samples. ...
Recent advancements in machine learning-based signal analysis, coupled with open data initiatives, have fuelled efforts in automatic sleep stage classification. Despite the proliferation of classification models, few have prioritised reducing model complexity, which is a crucial factor for practical applications. In this work, we introduce Multi-Scale and Attention Convolutional Neural Network (MSA-CNN), a lightweight architecture featuring as few as ~10,000 parameters. MSA-CNN leverages a novel multi-scale module employing complementary pooling to eliminate redundant filter parameters and dense convolutions. Model complexity is further reduced by separating temporal and spatial feature extraction and using cost-effective global spatial convolutions. This separation of tasks not only reduces model complexity but also mirrors the approach used by human experts in sleep stage scoring. We evaluated both small and large configurations of MSA-CNN against nine state-of-the-art baseline models across three public datasets, treating univariate and multivariate models separately. Our evaluation, based on repeated cross-validation and re-evaluation of all baseline models, demonstrated that the large MSA-CNN outperformed all baseline models on all three datasets in terms of accuracy and Cohen's kappa, despite its significantly reduced parameter count. Lastly, we explored various model variants and conducted an in-depth analysis of the key modules and techniques, providing deeper insights into the underlying mechanisms. The code for our models, baselines, and evaluation procedures is available at https://github.com/sgoerttler/MSA-CNN.
... We utilized two publicly available datasets, namely ISRUC [106] and the DREAMS Patients databases [36]. The ISRUC dataset comprises three subgroups out of which we used subgroup-3, which includes PSG data of 10 healthy subjects. ...
... and DREAMS Patients databases [36], respectively. The three subgroups of ISRUC-Sleep dataset [106], S1, S2, and S3 form datasets III, IV, and V, and contain data from patients with sleep disorders, patients under the effect of sleep medication, and healthy subjects, respectively. ...
This thesis aims to classify different stages of sleep in healthy controls as well as patients with sleep disorders. It provides a hierarchical model which combines multiple modalities in a hierarchical fashion to improve the performance of sleep staging classifier. Also, the last part of this work aims at diagnosing multiple sleep disorders including insomnia, periodic leg movement syndrome, nocturnal frontal lobe epilepsy, REM behavior disorders, etc. by proposing a LEE classifier utilizing LightGBM model with EOG and EEG signals for multi-class and binary classification problems respectively.
... While some researchers prefer to build their datasets from scratch, many prefer benchmark datasets. Three datasets have been used widely in the literature, namely, St. Vincent's University Hospital/University College Dublin sleep apnea database [35], MIT-BIH polysomnographic database [36], and SRUC-Sleep [37]. In addition, a few other publicly available datasets have physiological signals that can be used for analysis; these include the "PhysioNet," a large collection of physiologic signals that are captured and made available online for free [38]. ...
Raw polysomnography (PSG) preprocessing is one of the first steps in
any sleep disorder detection using artificial intelligence (AI) and data
science (DS). This chapter mainly discusses the process of transforming
raw PSG at the very beginning in a way that can be fed into a machine
learning (ML) or deep learning (DL) model. This includes essential
steps that come before building the actual model: starting from defining
the problem, collecting raw PSG, then data exploration, and finally,
preparing the data. PSG preprocessing is often highly specific to a
particular dataset at hand, the main expected result of the learning
model, and the equipment used for signal acquisition. For this reason,
it is common in the literature to overlook raw PSG preprocessing or
to mention it briefly without specifying details. Hence, giving a set
of universally applicable steps is not easy. This chapter discusses the
possible preprocessing steps that could be applied to the raw PSG data,
which were tested empirically or proven theoretically.
... It generates different views of input samples using data augmentation techniques, and then learns representations by maximizing the similarity between views of the same sample while minimizing the similarity between views of different samples. SeqCLR [19] introduces a set of data augmentation techniques specifically for EEG and [37], SEED [38], SleepEDF [39], ISRUC-S3 [40] Accuracy TS-TCC [21] Weak & strong augmentation Transformer Sleep & seizure detection HAR [41], SleepEDF [39], ESR [42], FD [43] Accuracy, F1 SSCL for EEG [22] Signal transformation CNN Sleep stage classification SleepEDF [39], DOD [44] Accuracy, F1 MulEEG [23] Multi-view contrast CNN Sleep stage classification SleepEDF [39], SHHS [45] Accuracy, Kappa, F1 ContraWR [26] Non-negative contrast CNN Sleep stage classification SHHS [45], SleepEDF [39], MGH [46] Accuracy COMET [27] Multi-level contrast CNN Disease detection AD [47], PTB [48], TDBRAIN [49] Accuracy, F1, AUROC, AUPRC SleepPriorCL [28] TS-TCC [21] generates different views of input data using both strong and weak augmentation methods. Weak augmentation employs jittering and scaling strategies, while strong augmentation uses permutation and jittering strategies, applying them to the temporal contrast module of EEG signals for temporal representation learning. ...
In recent years, the field of electroencephalography (EEG) analysis has witnessed remarkable advancements, driven by the integration of machine learning and artificial intelligence. This survey aims to encapsulate the latest developments, focusing on emerging methods and technologies that are poised to transform our comprehension and interpretation of brain activity. The structure of this paper is organized according to the categorization within the machine learning community, with representation learning as the foundational concept that encompasses both discriminative and generative approaches. We delve into self-supervised learning methods that enable the robust representation of brain signals, which are fundamental for a variety of downstream applications. Within the realm of discriminative methods, we explore advanced techniques such as graph neural networks (GNN), foundation models, and approaches based on large language models (LLMs). On the generative front, we examine technologies that leverage EEG data to produce images or text, offering novel perspectives on brain activity visualization and interpretation. This survey provides an extensive overview of these cutting-edge techniques, their current applications, and the profound implications they hold for future research and clinical practice. The relevant literature and open-source materials have been compiled and are consistently updated at https://github.com/wpf535236337/LLMs4TS.
... As shown in Table 1, we evaluated our approach on three publicly available datasets: Sleep-EDF (Kemp et al. 2000), ISRUC (Khalighi et al. 2016), and HMC (Alvarez-Estevez and Rijsman 2021). For each dataset, we utilized both EEG and EOG channels as input. ...
Sleep staging is crucial for assessing sleep quality and diagnosing related disorders. Recent deep learning models for automatic sleep staging using polysomnography often suffer from poor generalization to new subjects because they are trained and tested on the same labeled datasets, overlooking individual differences. To tackle this issue, we propose a novel Source-Free Unsupervised Individual Domain Adaptation (SF-UIDA) framework. This two-step adaptation scheme allows the model to effectively adjust to new unlabeled individuals without needing source data, facilitating personalized customization in clinical settings. Our framework has been applied to three established sleep staging models and tested on three public datasets, achieving state-of-the-art performance.
... To evaluate the performance of our proposed modules, we conducted experiments on three publicly available datasets: (1) ISRUC-S3 dataset (Khalighi, Sousa, Santos, & Nunes, 2016). It contains 6 EEG channels, 3 EMG channels, 2 EOG channels, and 1 ECG channel, recorded from 10 healthy subjects (9 male and 1 female). ...
