Deep Neural Network Based Sleep Staging And Blood Glucose Prediction

The sleep quality affects the blood glucose levels,while abnormalities in endocrine physiological indicators can also change sleep structure.Therefore,the study of sleep staging and blood glucose prediction based on deep neural networks has important clinical significance.Sleep staging is an essential part for the diagnosis of sleep disorders.Sleep stages are classified by specific doctors according to the characteristics of the polysomnogram signals such as electrophysiological signals.However,manual classification methods are time-consuming and labor-intensive.Deep neural network models could free up physician labor.However,deep neural networks are data-driven models that heavily rely on high quality and large amounts of data to train the latent parameters.Currently,the number of labeled sleep data is not yet able to support deep neural network models for large-scale training,such as natural image classification and language processing tasks.Therefore,this thesis introduces a semi-supervised algorithm to alleviate the dilemma caused by the lack of labeled sleep data.This thesis focuses on semi-supervised deep learning methods to build the sleep staging models.Blood glucose is an important physiological indicator of the human endocrine system.The monitoring of blood glucose levels can assist in determining cardiovascular and cerebrovascular function.Therefore,accurate and rapid detection of blood glucose concentration has a positive effect on the prevention and control of many diseases.At present,most of the clinical methods use invasive devices.However,these methods may make patients suffer,and there is also a risk of infection during prolonged measurement.Therefore,non-invasive blood glucose measurement is receiving more and more attention from researchers.This thesis proposes a video-based deep learning model to measure the blood glucose levels.Accordingly,the main contributions are as follows:1.On the single-channel EEG sleep staging,this thesis introduces the pseudo label and consistent regularization for semi-supervised training,which uses the adversarial learning to improve the efficiency of semi-supervised learning.It is difficult to establish discriminative classification boundary which could be realized by linear classifer in common feature space on the single electrophysiological channel based sleep staging task.Therefore,the feature fusion mechanism based poincare ball module is used to enhance the feature extraction ability for fine-grained information.Moreover,the symmetric positive definite matrix based manifold learning module and spectrogram are introduced to improve the performacne of semi-supervised learning.2.The obstructive sleep apnea disease could influence the performance of sleep stage classification models.The physiological characteristics between diverse obstructive sleep apnea population are different,which makes specific feature representation could be ignored by deep learning models.Accordingly,the semi-supervised multi-scale deep neural network is proposed to focus on sleep staging within diverse obstructive sleep apnea severities.In this thesis,a cross attention module based on higher-order statistics is introduced to enhance the model ability to extract fine-grained features.In order to effectively combine fine-grained convolutional features and coarse-grained global information,a Transformer based multi-scale classifier is also proposed for feature fusion.This structure can automatically adjust the corresponding feature gradient information for different obstructive sleep apnea states to achieve discriminative feature screening.For multi-modal electrophysiological information learning,a contrast-smoothing loss function based on adaptive weights is utilized to fuse the feature information.The proposed model could leverage multi-modal information to improve the performance compared to single modality based methods.The model is evaluated against several deep learning models on a local hospital dataset,and achieves an average accuracy of 81% and an average precision of 76%.3.Glucose concentration measurement based on near-infrared spectroscopy is theoretically demonstrated by recent researchers.Hence,the deep learning model could extract the blood glucose feature representations implicitly from fingertip videos.For video based non-invasive blood glucose measurement,the photoplethysmography signals are recovered from physiological videos.Then,a novel end-to-end hybrid multi-view deep neural network framework is proposed to predict the blood glucose values.The temporalfrequency map is proposed to rebuild the connnection within long-term temporal feature representation by fusing both temporal information and frequency information.This method could improve the feature extraction ability for both fine-grained feature representations and multi-scale global information.The feature attention module is introduced to fuse multi-view feature information into a unified discrete frequency space for feature selection.Thus,the extraction ability of neural networks for fine-grained features is enhanced.The thesis concludes with a comprehensive comparison experiment using the proposed model on a clinical dataset,and the prediction results achieve an average accuracy of 98%,which validates the effectiveness of the model.Besides,the framework is also running in Jetson Xavier NX for the simulation of blood glucose prediction within67 seconds.In summary,this thesis focuses on deep neural network-based sleep staging and blood glucose prediction.Based on the clinical needs,several efficient deep neural networks are proposed in this thesis and the effectiveness of the proposed models is verified on several clinical datasets,respectively.This study provides a reliable solution for automatic sleep staging based on semi-supervised learning and non-invasive blood glucose monitoring in home environment.It also provides a referenceable method for corresponding clinical applications.