Recognition And Grading Of Mental Health Condition Based On Deep Learning

Mental health is an important aspect of people’s overall health.It not only affects an individual’s physical health and quality of life,but also relates to the harmony and happiness of the entire society.Early diagnosis and intervention for mental health problems can greatly improve treatment outcomes and prognosis.However,due to insufficient medical resources and lack of knowledge about mental health,many patients with mental health problems do not receive timely diagnosis and treatment.Therefore,it is necessary and meaningful to develop automated systems for mental health condition recognition and grading using computer technology.The use of deep learning technology for automated diagnosis can not only effectively reduce dependence on mental health professionals,improve efficiency and stability,but also enhance the accessibility of psychological diagnosis and address the issue of patients’ lack of initiative.Therefore,the research in this project is focused on improving the efficiency of mental health condition recognition and grading,considering the practical application scenarios,and exploring identification methods based on structured data and short text data,ultimately establishing an automated algorithm for mental health condition recognition and grading.The work completed in this article mainly includes the following two aspects:(1)Research on mental health condition recognition and grading based on structured data.To address the problem of limited feature representation of structured data in neural networks,while meeting the demand for interpretability in mental health condition identification tasks,a Fully Embedded Bi-Order Network(FE-BiONet)with interpretability is proposed.This model includes two innovative structures:the Fully Embedding(FE)feature embedding module and the high-low order parallel network.The FE module automatically performs feature engineering while addressing the issue of limited feature representation of structured data,enhancing the feature representation of structured data.The high-low order parallel network aims to balance model accuracy and interpretability.The proposed model in this study achieved the highest AUC and F1 score on three datasets.In addition,the effectiveness of the FE module and the high-low order parallel network was verified through ablation experiments,and the interpretability of the model was validated through paper research on risk factors related to mental health problems.Finally,the proposed model was applied to the grading of mental health conditions in college students,providing reference for early intervention of mental health problems in college students.(2)Research on emotion classification based on short text data.Effectively recognition of negative emotions can provide new insights and suggestions for research on mental health condition recognition and grading.This study addresses the problem of poor feature representation in short text data due to limited information by proposing Emoji-Fused BERT.This model contains a separate emoji embedding and a fusion attention mechanism between emoji and short text features,which enhances and supplements the representation of short text features.The proposed model was experimented on three Weibo emotion classification datasets.Compared with BERT and other baseline models,Emoji-Fused BERT achieved the highest classification accuracy,validating the effectiveness of the emoji and short text fusion attention mechanism.Lastly,we applied the proposed model to two practical tasks in Weibo.Improving the efficiency of the mental health condition recognition of Weibo users by identifying the negative emotions of them.