| The quality of sleep is essential for an individual’s physical and mental well-being.Sleep data,which contains valuable information about sleep patterns,can objectively reflect sleep conditions and serve as a critical basis for diagnosing sleep disorders.Sleep data analysis can be divided into sleep staging and sleep arousal recognition.Currently,sleep arousal and sleep stage are simultaneously labeled through visual inspection of polysomographic recording by physicians,which is a time-consuming and tedious work due to the diverse range of formats and the large of amount of data to be analyzed.Forthermore,the results are susceptible to the technician’s personal experience.Therefore,automated analysis of sleep data holds significant research and practical value.In light of the aforementioned context,this article proposes a sleep data analysis algorithm based on the time-frequency images of electroencephalography(EEG).This method places emphasis on the design of an effective deep neural network model that combines the signal characteristics of both sleep stages and wakefulness.This enables the model to extract specific features from the time-frequency map of EEG,thereby allowing it to be classified according to the scoring habits of technicians and meet the required clinical accuracy standards.The main research work of this article is as follows:(1)To address the issue of the existing sleep staging models requiring a large amount of annotated data and limited accuracy due to insufficient feature extraction,this article proposes a semi-supervised sleep staging model based on a hybrid neural network that utilizes time-frequency images.To reduce the workload of technician annotation,it is advisable to employ a semi-supervised learning approach for feature extraction from unlabeled data in the first step.Secondly,in order to fully explore the characteristics of sleep data,both time-domain and frequency-domain signals are used as inputs to provide the spatial-temporal feature data foundation for the subsequent model.Afterward,a multichannel convolutional neural network was utilized to extract features from sleep data,and the extraction of prominent features was further enhanced by incorporating an attention mechanism.The above mixed features are finally fused and classified.In comparison with fully supervised learning on three public datasets and one private dataset,the semisupervised learning model achieved an average accuracy of 81.0% and a kappa value of73.2%.The results show that the proposed model can achieve comparable performance with supervised sleep staging models,while significantly reducing the workload of technician annotation data.(2)Existing sleep data analysis models are limited in their focus and do not meet the practical needs of clinical annotations.This paper proposes a multi-task sleep staging and sleep arousal model based on a time-frequency residual network.Firstly,utilizing time-frequency images as input enables the comprehensive utilization of sleep data.Secondly,the residual network is used as the backbone to extract spatiotemporal features,and attention mechanism is introduced to capture contextual information,thereby extracting distinctive features while preventing gradient vanishing.Finally,a multi-task loss was designed to achieve synchronous identification of sleep stages and sleep arousal.The results show that the multi-task model proposed in this article can achieve comparable performance with existing single-task recognition models on two public datasets and one private dataset.(3)In order to promote the practical application of medical-engineering collaborative projects,this article designed a sleep data analysis and visualization system.Firstly,the system was analyzed for its requirements secondly,the overall framework was designedfinally,the database storage and functional modules of the system were implemented.The test results show that the system can assist doctors in analyzing and diagnosing sleep data online,and has certain clinical application value. |
PDF Full Text Request