Automated Diagnose Of Sleep Apnea Syndrome Based On Deep Learning

Sleep Apnea Syndrome（SAS）is a common sleep disorder with a prevalence of 4% in my country,but the diagnosis rate is only 10%.The traditional diagnosis method of SAS requires the use of a professional sleep monitor（polysomnography,PSG）,usually in the hospital’s sleep center to monitor the patient’s sleep throughout the night（no less than 7hours）,during which the patient needs to paste multiple electrodes.Therefore,the monitoring process itself has certain sleep disturbances.The current diagnostic method mainly relies on doctors to interpret and evaluate PSG throughout the night.This process is time-consuming and labor-intensive,and has problems such as low efficiency,strong subjectivity,and high professional requirements.The development of big data and artificial intelligence technology has greatly improved the assessment and prediction methods of diseases.However,the methods based on manual feature extraction still rely on the prior knowledge of domain experts in the feature labeling stage,and problems such as the need for a large amount of data preprocessing in the evaluation and prediction stages are not well resolved.Based on the field of big data and deep learning,this thesis avoids the PSG segmentation signal,which requires manual annotation and a large amount of data preprocessing,and uses the entire signal as the research object to evaluate and predict SAS.（1）Collect SAS clinical data,analyze the signal characteristics of PSG sleep data,and complete the extraction and expression of sleep data.In order to avoid the problems of manual labeling and preprocessing,and realize the full automation of evaluation and prediction,the whole PSG signal characteristics are used for research;in order to be replicated in non-professional medical institutions,the signal characteristics of blood oxygen saturation are used for research.The characteristics of the oxygen saturation signal are directly related to the SAS.Moreover,compared with signals such as ECG,EEG,and EMG,the blood oxygen saturation signal is easy to collect and has strong anti-interference ability.（2）An automatic diagnosis model of sleep apnea syndrome based on one-dimensional residual convolutional neural network is proposed.The model aims to assess and predict the severity of SAS in patients,namely: normal,mild,moderate and severe.Among them,in terms of making full use of raw sleep data,convolutional neural network is used to automatically extract the characteristics of blood oxygen saturation of patients with different severity.In terms of deepening the structure of the deep neural network,the residual block structure of short-circuit connection is realized by using the residual neural network,which solves the problems of gradient explosion or gradient disappearance of the deep model,and improves the performance of the model.（3）In order to verify and further optimize the model,several sets of experiments were carried out.In the SAS diagnosis and prediction of actual clinical data,the overall accuracy rate of 86.09% was obtained.The classification accuracy rates for each category are as follows: normal 87.84%,mild 85.88%,moderate 67.27%,severe 91.00%.In addition,a model performance comparison experiment was carried out,and it was found that the diagnostic models based on different characteristic signals showed that the SAS automatic diagnosis model performed well on normal and severe types,and the model performed in general on moderate types.At the same time,it was found that the moderate type was more likely to be misjudged as the mild type.The performance of the model based on the single-channel blood oxygen saturation signal is the best,and the blood oxygen saturation signal plays a key role for the method based on the whole-segment signal.The network hyperparameters are tuned by the Bayesian search algorithm.Finally,the ensemble learning method based on the voting method combined with the strategy is used to greatly improve the performance of the model.The overall accuracy of the final model reaches 95.56%,and the classification of each category is accurate.The rates are as follows:normal 100%,mild 92.94%,moderate 85.19%,and severe 99%.This result is close to the level of human expert diagnosis.（4）Finally,a SAS diagnosis and prediction system is developed in the Python environment,which consists of a sleep monitoring data management system and a rapid diagnosis and prediction system for sleep apnea syndrome.The former is responsible for processing sleep monitoring data,including data selection,interception,downsampling,noise reduction and other functions;the latter is responsible for receiving the characteristic signals of the subjects and calling the deep learning model to complete the diagnosis and prediction of SAS.