Study On Detection Of Sleep Apnea And Hypopnea Event Based On Nasal Flow Signals

Epidemiology investigation shows that the incidence of sleep apnea-hypopnea syndrome in China is rising and has gradually become a serious threat to people’s life.Sleep apnea-hypopnea syndrome is a disorder caused by repeated complete or partial upper airway collapse during sleep.Patients with sleep apnea-hypopnea syndrome have a severely reduced sleep efficiency due to unconscious choke during sleep.While during daytime they suffer from problems such as drowsiness,fatigure,inattention and memory loss.Besides,sleep apnea-hypopnea syndrome is also linked to hypertension and diabetes.The gold standard for diagnosis of sleep apnea-hypopnea syndrome is polysomnography.However,many patients with sleep apnea-hypopnea syndrome are not diagnosed because polysomnography is expensive and need to be conducted in laboratory.At present,there are many studies for diagnosis of sleep apnea-hypopnea syndrome using limited of signals for replacement of polysomnography.However,there are some limitations such as low time resolution and not good precision.In this paper,firstly a sliding window was adopted to cut the nasal flow signal into data segments with time information.Then statistal features were extracted from each segment as the input for the classifier.The classifier was a random forest composed of classification and regression trees.And based on the output of classifier,each apneahypopnea event could be located in seconds and the apnea-hypopnea events per hour could be calculated then the final diagnosis result could be reached.Besides,on the basis of out-of-bag error rate in random forest,the importance of each feature was evaluated and feature selection was completed,which would effectively reduce the algorithm complexity and improve the model efficiency.Manual feature extraction is based on prior knowledge and may be sensitive to datasets,which may result in a decline in the generalization ability of model.Focused on this problem,based on convolutional neural network(CNN)and long-short term memory(LSTM)network,a LSTM-CNN network was built to automatic extract features from nasal flow signal and achieve the classification of SAHS.Cross validation was adopted to confirm the generalization ability of this network.Finally,50s’ subjects data analysis shows that the average accuracy,sensitivity,specificity and precision for diagnosis of sleep apnea-hypopnea syndrome severity reached 97.0%,98.9%,91.4%and 94.5% respectively.Besides,it achieved a Cohen’s KAPPA coefficient of 0.87 compared with the manual diagnosis results.The cascading detection model based on statistical features extracted from nasal flow signal and LSTM-CNN network are both able to predict the sleep apnea-hypopnea events on short time scales.And based on which,the apnea-hypopnea index could be calculated and the diagnosis result could be reached.Besides,the algorithms show a good timeliness and robustness,thus can be adopted as an effective method for clinical diagnosis of sleep apnea-hypopnea syndrome.