Deep Learning Based On EEG And EM Signals For Mild Depression Recognition

Depression is a mental illness that seriously endangers the physical and mental health of patients.In the past 30 years,the incidence rate of depression has increased sharply,and it is still on the rise.However,the prevention and treatment of depression in China is still in based on traditional inquiry and scale index,which leads to the problems of difficult to effectively identify depression and low recognition rate in the early stage of depression.The effectiveness of electroencephalography(EEG)in the field of depression recognition has been proved by many studies,but the research of mild depression recognition based on EEG is still in its infancy.The eye movement data(EM)can reflect the part of the subjects interested in the experiment,which plays an important role in understanding the behavior of the subjects.In view of this situation,this paper proposes two models based on deep neural network to help identify mild depression.The data used in this study come from the EEG data and eye movement data of 48 Lanzhou University Students(24 depressed subjects and 24 healthy controls)during the emotional face browsing task.The main work of this work is as below.(1)Mild depression recognition based on multi kernel convolution and spatiotemporal features.The performance of individuals in the same EEG acquisition task is different,which leads to different recognition performance on different convolution kernels.Therefore,this study adopts the method of multi-kernel to extract spatial features of different granularity,it not only enhances the diversity of features,but also reduces the impact of individual differences,so that the recognition accuracy of the model can reach 81.81%.However,EEG contains a lot of temporal information besides spatial information.Temporal information reflects the state changes of the whole brain during the task.Simple combination of spatial information and temporal information does not reflect the brain’s activity state well.In this paper,the matrix multiplication operation in LSTM unit is replaced by convolution operation to strengthen the temporal and spatial characteristics.New LSTM strategy can improve the performance of the model to 83.47%.In order to effectively detect mild depression,a new loss function is used to amplify the role of mild depression samples in the network training process.On this basis,the model achieves the highest accuracy of 86.18%,and the recognition accuracy of mild depression samples also reached 82.08%.This will be more conducive to the detection of mild depression in the early stage.In addition,this paper also discusses that batch regularization technology will reduce the recognition accuracy of EEG data with special characteristics。(2)A multi-task learning of depression recognition based on EEG and eye movement.Single EEG data is not effective enough to reflect the state of brain activity,and the eye movement data collected simultaneously with EEG can help improve the brain activity information to a certain extent.Therefore,how to effectively combine EEG and Eye Movement information in the field of depression recognition is particularly important.Multi-task learning can learn multiple related tasks at the same time,so that the knowledge of these tasks can be shared in the learning process,so as to improve the performance and generalization ability of the model in each task.The data in this paper are EEG data and Eye Movement data generated by subjects in the same task at the same time.There is a strong correlation between the two tasks,which can be used to build a multi-task model.The EEG private module adopts the multikernel convolution strategy in the first model,and the eye movement private module and sharing module adopt three-layer DNN structure.In the multi-task model,the accuracy of EEG recognition increased from 83.47% to 85.35%,and the accuracy of Eye Movement recognition increased from 71.78% to 74.93%.After the model training,the shared module is used as the feature extractor to generate new features,and multiple machine learning models are used to verify that the features extracted by the shared module contain the commonality among multiple tasks.These results show that the model can work effectively even when one of the data is difficult to obtain or damaged.