Deep Learning Based Research On Recognition Of Depression Patients Using EEG Data

There are a large number of patients with depression in the world currently.The number of people with depression is growing and is expected to be second only to that of heart patients in 2020 affected by fast-paced and high-stress lifestyles.If patients could be diagnosed and treated in time,the burden on patients,their relatives and society will be reduced.However,the diagnosis of depression mainly focus on scale testing,which is characterized by strong subjectivity and low accuracy.Therefore,there is an urgent need to find a way to identify patients with depression objectively and efficiently.There are significant differences in brain activity between depressed and normal populations,which have been shown by numerous studies.Electroencephalography(EEG),a bioelectrical signal that can reflect the state of brain objectively,is widely used in the field of depression research because of its non-invasive,easy-to-access,and high temporal resolution.Currently,there is no research that could clearly show the pathogenesis of depression,and researchers lack sufficient prior knowledge in designing and screening features influenced by the unclear pathogenesis of depression.Deep learning(DL)has been successfully applied in many fields due to its characterization learning ability with the rapid growth of computing power and the convenience of sample acquisition.Compared with traditional machine learning methods,DL could extract and combine features spontaneously,and there I no complex feature engineering.In this paper,eye-closed resting state EEG was collected from 17 depression and 17 normal subjects,and the identification method of depression was explored based on the collected EEG and DL.The main work of this paper is as follows:(1)The amount of available data is increased to thousands by clipping a single EEG record into segments.With methods of arranging records of electrodes in order according to the number of electrodes,distance based projection and non-distance based projection,EEG is organized as samples for DL models to preserve the time domain,frequency domain and spatial characteristics of the original EEG signals;(2)Different network structures are designed and different training strategies were adopted according to characteristics of three types of EEG frames;(3)The percentage of samples that are predicted to certain class by models in the samples of corresponding individual is taken as the probability that the individual belongs to the class.The methods of this paper are compared from the accuracy of segment classification and the accuracy of subject classification;(4)The effectiveness and reliability of methods are verified by deep learning visualization technology.The highest classification accuracy rate of 81.03% for cropped samples and individual classification accuracy of 88.24% are obtained with methods in this paper.The visualization results show that features extracted by deep learning models can distinguish the two groups of people effectively,and the prefrontal lobe contribute mostly to the prediction of models,which is consistent with the current research.The results of this paper verify the effectiveness and feasibility of further application of DL in depression research.