Study On Sleep EEG Characteristics For Depression Recognition

With the development of society and the increase in life pressure,the incidence of mental illnesses,mainly depression,has increased significantly.The current scale self-evaluation and doctor’s diagnosis methods are inefficient and depend on the doctor’s diagnostic experience.The subjective cognition will also affect the judgment of the result,which is easy to cause misdiagnosis.At present,with the in-depth study of EEG signals,it is found that EEG signals are closely related to brain activity and the emotional state of people,and the research of affective diseases based on EEG signals is becoming more and more common.Compared with the EEG signal during the awake period,this thesis chooses the EEG signal during sleep for two main reasons: First,the EEG signal during sleep has the characteristics of being difficult to disguise,undisturbed,and easy to collect.It is conducive to the study of depression;secondly,the depression of depressed patients does not have continuity in time,and the short-term waking period EEG signal contains less depression information,while the sleep EEG is more important because of its long time span.It is possible to collect EEG signals with depression information.Based on this,this thesis analyzes and explores the characteristics of sleep EEG.The main research innovations and contributions are as follows:1.Aiming at the problem of sample imbalance and different misclassification costs,this thesis proposes a cost-sensitive feature selection algorithm—minimum mis-classification cost-maximum recognition contribution feature selection algorithm.After performing sleep staging on the sleep EEG data of 36 depressed patients and normal subjects collected in this paper,it is found that there is an imbalance in the same staging data between the two groups of people,resulting in feature selection,the characteristics will be more inclined to the majority of samples recognition.Therefore,this thesis proposes a cost-sensitive feature selection algorithm.By setting different misclassification costs and recognition contributions for different categories,the feature ranking is more inclined to minority samples.In order to verify the effectiveness of the algorithm in this thesis,a comparative analysis of 6 different feature selection algorithms,including the algorithm in this thesis,was conducted on 4 public datasets.It was found that the algorithm in this thesis achieved a higher recall rate while ensuring the AUC value.2.Based on Study 1,this thesis analyzes and studies the sleep EEG feature matrix,and constructs an effective EEG feature set for different sleep stages.Using the feature selection algorithm in this thesis,it is found that among many sleep EEG features,power spectrum entropy,absolute power,relative or absolute center frequency are better features for depression recognition.From the perspective of different EEG bands,the characteristics of alpha band and beta band can distinguish depressed patients from normal subjects to a greater extent.Using different features selected from different stages to construct a depression recognition model,it is found that WK stage is the stage with the best classification effect among the5 sleep stages,with a classification accuracy rate of 0.794;and in NREM sleep,N3 stage is the stage with the best recognition effect,reaching 0.767;among the5 stages,the classification accuracy of REM staging is the lowest;among the 6leads,the right leads(C4,F4,O2)have a better recognition effect.Through statistical analysis methods,it is found that the WK period and the N2 period have the most significant differences in the number of features,while the N3 period sleep EEG data with better recognition effects have not many significant differences.To sum up,in view of the problem of data imbalance,this thesis proposes a feature selection algorithm with minimum misclassification cost-maximum recognition contribution,and verifies the effectiveness of the algorithm.Based on this algorithm,the thesis constructs the optimal feature set for different sleep stages.Through the analysis of the feature selection results,we find the sleep stages,sleep brainwave bands and leads that are closely related to depression recognition.Use the selected feature set to build a depression recognition model and achieve good recognition results.