Research On Early Depression Recognition Based On Deep Learning

Depression is a common chronic serious mental disease,which has a great negative impact on people’s thinking,feeling and behavior;On September 11,2020,the National Health Commission of China issued the first depression prevention and treatment action plan,the "Work Plan for Exploring Special Services for Depression Prevention and Treatment",which includes the routine screening of depression nationwide.The traditional scale-based screening has the limitations of low doctor-patient ratio and lack of multi-index diagnosis.Therefore,this paper studies the automatic diagnosis method integrating multiple behavioral indicators of depression,and proposes a multi-mode depression recognition method based on deep learning.The main research work and results of this paper are as follows:(1)Construction of the Chinese Emotional Dictionary of Depression(CEDD):Because there are too few Chinese emotion dictionaries in the field of depression,this paper builds and expands the Chinese emotion dictionary in the field of depression based on the general domain dictionary and word similarity calculation.Firstly,the comment corpus of depression patients is obtained through web crawler,and based on the characteristics of language variability and emotional polarity of the comment corpus of depression patients,it is processed by denoising and word segmentation;Secondly,to filter out the useless words in the text and retain the words that really affect the text,the seed word set is obtained through the TF-IDF(Term Frequency-Inverse Document Frequency)weighting algorithm;Finally,in order to expand the domain emotion dictionary,the cosine similarity calculation is used to expand the domain emotion dictionary of How Net and Dalian University of Technology,and CEDD is obtained.(2)Research on the Method of Identifying Depression Tendency in Online Social Text:The text classification method based on universal domain lacks the ability of domain generalization,while Chinese characters and words have their own information and characteristics.By combining the domain emotion dictionary with the Chinese word features,on the one hand,it has stronger domain generalization ability,on the other hand,it can more fully represent the text information.The method is applied to BERT model and BERT-W model is obtained.The experimental results show that the model is better in the recognition accuracy of depression text.(3)Construction of Chinese multimodal data set CMDD in the field of depression:The behavioral characteristics of patients with depression are manifested in multiple modes,while the available Chinese multimodal depression public data set is relatively small.This paper,together with Shanxi Intelligent Big Data Industry Technology Innovation Research Institute and Maiyata Bacterial Intelligent Technology Co.,LTD.,has jointly constructed the Chinese multimodal data set CMDD in the field of depression;Secondly,the content of the data set is analyzed and explained in terms of gender,degree of depression and score ratio of HAMD-17.(4)Research on depression recognition method based on multimodal feature fusion:The behavior of patients with depression is manifested in multiple modes such as language content,vision and hearing.The multimodal recognition model integrating multiple indicators can more accurately screen out potential objects of depression.On the basis of method(1)(2)(3),multi-class features are first fused by Z-core normalization method;Secondly,the relationship between features is learned through two-way LSTM layer,and the attention mechanism is used to select the information that is more critical to the current task goal from many features;Finally,the classification is completed through the Softmax normalized index function,which can more objectively and accurately achieve the task of depression screening.Based on the proposed multimodal depression recognition method,on the one hand,it can achieve auxiliary screening and diagnosis,reduce the diagnosis and treatment process,and save diagnosis and treatment time;Secondly,the reliability and accuracy of screening results can be enhanced by referring to the behavior characteristics of multiple modes;Finally,this method has low application cost and strong portability,which can solve the imbalance of medical resources to some extent.