EEG-based Depression Recognition Using Machine Learning

In recent years,the prevalence of depression has gradually increased,and depression has become one of the most pressing common health issues.At present,the main diagnostic methods for depression are doctor-patient communication and scale analysis which are dependent on doctors' experience and professional background.It is time-consuming and laborious.Due to the advantages of EEG,such as non-invasive,low-cost,simple to operate,easy to record,researchers begin to apply EEG to depression recognition research.They use machine learning algorithms to analyze the EEG signals collected from depression patients and try to find an objective,timely and accurate method for diagnosis of depression.However,there is comparatively little research on EEG-based depression recognition using deep learning method at this stage,and whether using traditional machine learning method or deep learning method,most research demands manual extraction of multiple features based on EEG signals and simply combines extracted features with traditional classification algorithms or neural network models.How to find a better way to recognize depression remains a challenge to surmount.To this end,we use machine learning and EEG to explore new computeraided methods for diagnosis of depression.In this paper,we use a 128-channel HCGSN(Hydro Cel Geodesic Sensor Net)EEG acquisition system to collect EEG data based on an emotional face stimuli task.We select EEG data from 28 subjects for the research,including 14 depression subjects and 14 normal controls.The main work of this paper is as follows:(1)In this paper,firstly use the adaptive noise canceller based on LMS algorithm,the 0.5-40 Hz band-pass filter,Fast ICA algorithm to get rid of artifacts from EEG signals.Use the AR model and Hjorth algorithm to calculate the power spectral density and activity as the raw feature vectors based on 0.2s,0.5s,1s,and 2s time windows on the alpha,beta and theta frequency bands,respectively.(2)The statistical analysis is used to analyze changes in power spectral density and activity in six major brain regions(i.e.left frontal,left temporal lobe,left parietaloccipital,right frontal,right temporal lobe and right parietal-occipital)of depression patients and normal subjects on alpha,beta,and theta frequency bands,and traditional classifiers SVM,RF and k NN are used to analyze the constructed feature vector set.(3)An ensemble learning model based on deep forest and SVM is proposed.According to the proposed ensemble learning model and characteristics of the constructed feature vectors,two training strategies of ensemble learning model are proposed,namely the strategy based on fixed features and the strategy based on fixed time windows.Then,the ensemble learning model is trained separately according to the proposed strategies.(4)The spatial information of the EEG signal is added to the depression recognition by converting the EEG signal from a continuous signal into twodimensional images.Based on the generated images,two different convolutional neural network structures are proposed for image recognition,which are 7-layer CNN and 8-layer CNN.At the same time,EMD method is used to explore another possibility of EEG signal recognition method.We evaluate the performance of the proposed method on total and single frequency bands,respectively.For the traditional classifiers,the best classification accuracy of the total frequency band is 82.06%±7.47%,and the best classification accuracy of the single frequency band is 84.86%±7.88%.For the ensemble learning method,the best classification accuracy of the total frequency band is 89.02%±6.19%,and the best classification accuracy of the single frequency band is 88.28%±6.72%.For the deep learning method,the best classification accuracy of the total frequency band is 82.36% ±9.57%,and the best classification accuracy of the single frequency band is 84.75%± 10.35%.For the EMD method,the best classification accuracy of the total IMF is 79.85% ±6.55%,and the best classification accuracy of the single IMF is 85.25%±6.83%.These research results demonstrate the effectiveness of the proposed method in this paper and show that EEG signals could be used as reliable indicators for depression recognition,which make it possible for EEG-based portable system design and application in auxiliary depression recognition in the future.