| dc.description.abstract | Sleep is not a single uniform state, but rather a cyclical pattern involving multiple stages such as rapid eye movement (REM) and non-REM (NREM stages N1, N2, and N3). Polysomnography (PSG) is considered as the gold standard for sleep scoring; however, it requires experts to visually inspect overnight data of about 8 hrs and annotate as per the standard guidelines. This manual scoring is tedious, time-consuming, and expensive, and also suffers from inter-rater variability. Among the different signals recorded in PSG, electroencephalogram (EEG) is considered to be the most useful and informative for differentiating the different stages of sleep. Electrooculogram (EOG) and electromyogram (EMG) are also used in PSG to record eye movements and muscle activities, respectively, which can help distinguish NREM from REM sleep stages. This thesis explores the modalities of EEG, EOG, and EMG for accurately classifying the multiple stages of sleep. Also, it aims to diagnose different sleep disorders by extracting effective features from the overnight sleep recordings of patients.
In the first part, the results of our studies on the binary classification of sleep and wake states are presented. We have evaluated the performance of single-channel EEG using handcrafted features, such as bandpower ratios, Lempel-Zev complexity, sample entropy, and Hjorth's parameters. We also performed Poincare plot (PP) analysis and derived various features such as length and width of PP, area, ratio of length to width, asymmetry index of PP, and entropy of gridded PP. We then investigated the utility of single-channel EOG or EMG in classifying sleep from wake states in healthy subjects as well as clinical population. With an EMG signal, we obtained an average classification accuracy of 85% on 10 healthy controls and 70% on 25 patients with different sleep disorders. An adaptive data-driven approach called ensemble empirical mode decomposition is used to obtain the mode functions of the EOG signal. Mean, mode, standard deviation, kurtosis, and skewness are computed for each of the intrinsic mode functions (IMFs) and are used as features. Apart from these, we also calculate the instantaneous frequency and energy of IMFs using Hilbert Huang transform. For feature selection, we have used mutual information criteria to rank the features according to their importance and considered top K features (value of K is selected such that after including these top K features, there is no further improvement in classification performance), which are consistently present across all the runs. By using the single channel EOG, we obtain an accuracy of about 95% on healthy controls and 91% on patients with sleep disorders. These studies show that a single-channel EMG or EOG provides results at par with EEG and can be used to identify sleep and wake states in a clinical setting.
In part two of the thesis, we have considered the multi-class classification of sleep stages using a single-channel EEG. In this work, we used the handcrafted features that were investigated in our earlier work on binary classification of sleep-wake states. We utilized random undersampling with boosting technique (RUSBoost) to deal with the class imbalance issue, which is innate to the sleep stage classification problem. This RUSBoost classifier uses a decision tree as the base classifier. We have evaluated the performance of the proposed method on three publicly available datasets of overnight PSG recordings of healthy individuals. The results are reported using two different approaches: (i)subject-independent testing (SIT) with leave-one-subject-out and 50%-holdout strategies and (ii)subject-dependent testing (SDT). The proposed approach is able to outperform most of the existing studies in the literature on automated speech stage classification (ASSC).
In the third part of the thesis, we aim to further improve the performance of ASSC by a) decomposing the multi-class classification problem into multiple binary classification tasks and b) combining multiple modalities, i.e., EEG, EMG, and EOG. This work proposes a hierarchical model (HM) in which the different levels of hierarchy are designed such that a) each level deals with only one binary classification task, and b) sleep stages that are difficult to disambiguate are passed through multiple levels before reaching the final decision. In this work, we also investigate the effectiveness of temporal context (TC) and data augmentation (DA) in the classification performance of the proposed HM. The model is evaluated on seven publicly available datasets comprising healthy subjects as well as patients with diverse sleep disorders. We found that the combination of DA and TC provides a consistent improvement across all the datasets. We also reported the results of the cross-dataset evaluation in which the model is trained and tested on different datasets. The proposed model achieves average accuracies of 83.1%, 90.0%, 84.4%, 82.1%, 81.5%, 79.9%, and 73.7% on Sleep-EDF, Exp Sleep-EDF, ISRUC-S1, S2 and S3, DRMS-SUB, and DRMS-PAT datasets, respectively. For all the datasets except DRMS-SUB, the proposed method outperforms the state-of-the-art approaches for the automated identification of sleep stages.
In the last part of the thesis, we aim to classify various sleep disorders, namely insomnia, narcolepsy (NAR), periodic leg movement syndrome (PLM), nocturnal frontal lobe epilepsy (NFLE), REM behavior disorder (RBD), bruxism and sleep-disordered breathing (SDB) using EEG, EOG and/or EMG signals. We used gradient boosting decision tree model called LightGBM to classify healthy controls and different sleep disorders, and compared its performance with that of SVM. The proposed approach is evaluated on the publicly available CAP dataset of 108 subjects, including healthy controls and seven sleep disorders. We first addressed the problem of binary classification of specific pathologies from healthy controls. A single feature called gridded distribution entropy derived from Poincare plots of EEG signal is able to provide 100% accuracy in distinguishing healthy controls from each pathology, except insomnia and PLM. However, by using EOG channel, these two groups are distinguished from healthy controls with 100% accuracy, indicating the efficacy of EOG channel in disambiguating insomnia and PLM. An accuracy of 83.3% is obtained for the seven-class classification of sleep disorders using only an EOG channel. We have further utilized the binary EEG classifiers for disambiguating the classes that are misclassified by the seven-class EOG classifier. The proposed approach is referred to as the LightGBM-EOG-EEG (LEE) method since it first utilizes the EOG channel for performing multi-class classification (healthy controls and six sleep disorders) and then the EEG channel (P4-O2) for correcting the confused classes. The performance improves from 83.3% to 93.3% by using the LEE method, which combines the powers of EOG and EEG channels. Finally, we used simple threshold-based postprocessing to resolve the confusion between NAR and SDB, taking the accuracy to 94.4%, the best in the literature. This threshold is based on the value of the ratio of the duration of two sleep stages. | en_US |
