EEG Signal Based Sleep Disordered Breathing Recognition Research

Sleep disorders could be divided into seven major categories according to different clinical causes,and sleep disordered breathing is one of them that could not be ignored.Researches have shown that the prevalence of sleep disordered breathing is increasing year by year,however the public has not paid enough attention to it.Many medical studies have pointed out that OSA events would increase the risk of cardiovascular and cerebrovascular diseases.Therefore,timely diagnosis and treatment is very important.In clinical practice,the polysomnography(PSG)is the gold standard for diagnosing sleep breathing disorders,which could determine the type of illness and the severity of the condition.In a collaborative communication with one third-class hospital in Beijing,clinicians pointed out that there is a certain correlation between brain activity and OSA events.For example,patients with long-term sleep apnea may have an increased risk of stroke or even Parkinson’s disease.In current researches,OSA events are generally judged by blood oxygen saturation and oronasal airflow,which raises the question that whether OSA events could be judged solely by EEG signals.As the overall reflection of the cerebral cortex electrophysiological activities,EEG signals in PSG have the advantages of sustainable monitoring and no impact on the human body,compared to other monitoring technologies.Therefore,analyzing and distinguishing sleep disordered breathing events through EEG signals has important medical value and engineering significance.In addition,clinicians have also proposed that there is a correlation between brain activity,sleep stages,and OSA events.However,based on existing technologies and researches,it is still difficult to systematically analyze this association.We hope to adopt new technologies such as artificial intelligence to mine this potential relationship.More importantly,all current studies have ignored this potential relationship,regarding sleep staging and OSA events detection as mutually independent tasks.Therefore,the works of this thesis are as follows:·To settle the problem that complex semantic information in OSA events,this thesis first proposed a sleep breathing event detection algorithm based on weakly supervised multi-instance learning(EEGMILNet).Through the multi-instance learning framework,EEGMILNet transformed the complex labels classification task based on feature set into predicting the overall attributes of this set,effectively overcoming the feature space ambiguity problem and underfitting problem.Validation and evaluation were implemented on three clinical datasets.The sleep breathing event detection method proposed in this thesis achieved 77.3%,74.1%,and 78.6%in accuracy,respectively,and got 4.2%,2.0%,and 2.3%performance improvements compared to the baseline model.The results indicated that EEGMILNet could effectively alleviate feature space ambiguity through multi-instance feature fusion and feature space optimization.At the same time,we expanded the model to effectively predict the duration of OSA events,exploring the scalability of future clinical engineering applications.·Most current researches implemented sleep staging and OSA events detection as separate tasks and ignored the potential correlation between the two.Inspired by clinical medical researches,this thesis proposed a sleep disordered breathing detection network based on multi-task collaborative learning(i.e.SleepMTCLNet).SleepMTCLNet regarded two tasks as a whole,and explored relevant features from the whole through local parameter sharing and cross task knowledge distillation,rather than just sharing parameters or narrowing the distance between different tasks.This thesis has conducted sufficient experimental verification on two dataset-s,the proposed model achieved 73%on the sleep staging task and 74%on the OSA events detection task.Compared to other works,SleepMTCLNet based on multitask collaborative learning got a 1%-5%performance improvement while effectively reduced parameters and overcome task bias problem.Therefore,this method could be applied as potential solution for sleep lightweight monitoring models in the future.·Considering the correlation between brain activity,sleep stages,and OSA events,this thesis adopted multi-task collaborative technology to mine and learn this complex relationship.The results observed that the clustering distribution of OSA event features extracted through deep learning model in the feature space is roughly consistent with its corresponding sleep stages.In the unsupervised cross task experiment,the features extracted for OSA events detection task achieved an overall accuracy of 53%on sleep staging task,with F1 scores of 76%and 64%achieved in WAKE and N3 stages,respectively.This work further proved a certain correlation between brain activity,sleep stages,and OSA events.