Research On OSAHS Recognition Based On Deep Learning

The average sleep quality of modern people is getting worse and worse.In addition to mental stress,physical discomfort also greatly affects the quality of breathing and thus affects the quality of sleep.According to research,snoring is caused by many hidden diseases,including Obstructive Apnea Syndrome(OSAHS),which is extremely harmful.The disease can be diagnosed by the gold standard polysomnography(PSG)to assist doctors.PSG can use multi-dimensional data such as brain waves,heart rate,and blood oxygen saturation to determine whether a person has OSAHS.However,the traditional measurement of OSAHS,which is the PSG instrument is cumbersome and inconvenient.We need a new system instrument to measure OSAHS.Through theoretical analysis of OSAHS-related snoring and normal snoring,there are not only differences in acoustic characteristics and time-domain features,but also in frequency spectrum.From this,it can be concluded that OSAHS-related snoring and normal snoring are classified to determine whether the patient has OSAHS.The key step of the snoring detection system studied in this paper is:1.Extracting and intercepting the audio data of the snoring sound with a duration of 3 s through the endpoint detection algorithm in the hardware equipment;2.Using the feature extraction method to extract audio features in the form of spectrograms;3.Inputting the feature signals into the trained neural network model for inference recognition,and obtaining the final recognition result.The experimental innovations in this paper are:1.Building a training database;2.Combining sliding window and segment similarity loss function to construct an E-LSTM model to identify snoring;3.Optimizing the loss function;4.Proposing a new method for calculating AHI value.The experimental results show that the recognition performance of the combination of the MFCC feature extraction algorithm and the E-LSTM model is the best,with a maximum accuracy of 92%.After improving the loss function on the basis of this model,the accuracy of the recognition results increased by 2%,and the final two-class accuracy reached 94%.In addition,in order to further judge the severity of OSAHS patients,the new data is input into the trained model for reasoning and recognition.Finally,the experimental result,that is,the severity of OSAHS disease,is judged in a quantitative manner by calculating the AHI value.The accuracy of the experimental system to reason about the severity of OSAHS reaches 80%.