SAHS Automatic Recognition Based On SaO₂ Signal Algorithm Research

Sleep Apnea-Hypopnea Syndrome（SAHS）is a clinical syndrome where various factors lead to repeated apnea and hypopea with sleep interruption in a sleep state.It is of great significance to identify SAHS based on physiological signals,because the main method is time-consuming and powerful to diagnose SAHS,and the poor symptoms of SAHS will cause problems such as low diagnosis rate and delay the optimal treatment time.Compared with other physiological signals,Pulse Oxygen measurement equipment is portable and easy to be setted up in the family environment,so the cost of obtaining arterial oxygen saturation signal is low.Consequently,the development of SAHS technology automatically identified by SaO₂signal is crucial.According to the deficiency of SAHS in traditional physiological signals,after the analysis and learning,this thesis presents a SAHS algorithm based on traditional features recognition and a SAHS algorithm for Res Net18 Network recognition based on migration learning.Experimental data is selected the SaO₂signal from the Sleep Heart Health Research database,the specific studies are as follows:（1）This thesis introduces a SAHS Recognition Methods based on Multi-scale Baseline Features of SaO₂Signals.Firstly,the nearest assignment method is proposed to solve the zero-level artifacts caused by sensors in the process of SaO₂signal acquisition in the database,and the random equilibrium data processing method is used to solve the unbalanced phenomenon of subjects in the database,so as to obtain the balanced data set.Secondly,on the basis of the obtained data set,based on the analysis of the relationship between SaO₂signal and SAHS,the baseline values of each data fragment with the length of 50 are defined,and the length,scale and 15 time-domain features with the baseline descending scale higher than 2%,3%and 4%are extracted to form the feature data set.Finally,Support Vector Machine and Random Forest classifier are used to identify SAHS,and the average recognition accuracy is 86.25%.In order to further improve the recognition rate,solve the problem of high correlation of some baseline features in the feature data set and increase the feature distance between classes,the algorithm analyzes the principle of Principal Component Analysis（PCA）,Kernel Principal Component Analysis（KPCA）and Kernel Linear Discriminant Analysis（KLDA）,and selects the KLDA method to reduce the dimension of the feature data set to obtain a new feature data set.Finally,the random forest classifier is used to identify SAHS.The average recognition accuracy of the KLDA method is 97.5%,which is 11.25%higher than that before dimension reduction.（2）This thesis presents a SAHS Recognition Method based on the Resnet18 Network.Firstly,the preprocessed SaO₂signal is converted into an image signal.Secondly,based on Res Net18Network and the Randomly Cropped Preprocessing Method,SaO₂signal is trained to obtain the A-Res Net Model.The spectrograph and SaO₂signal are trained to obtain the TL-X Res Net Model.The TL-ⅠRes Net Model is obtained by training the spectrograph.To make the detailed information of the SaO₂signal more obvious,the Random Cropping Preprocessing Method is optimized for Scaling Methods,and TL-ⅡRes Net is obtained by training the spectrograph based on the Resnet18Network.Finally,based on the thought of Transfer Learning,the SaO₂signal is identified SAHS as input to the above four models,the average identification accuracy is 53.5%、76%、85.75%and93.75%,respectively.Compared to the original Res Net18 Network,the accuracy of the TL-ⅡRes Net Model is improved by 40.25%.