Key Techniques Research On Depression Assessment For Multi-sensory Wearable Graphical Representation

Depression is a prevalent psychological disorder characterized primarily by persistent and prolonged feelings of low mood.In severe cases,individuals may experience hallucinations,delusions,or even develop tendencies towards self-harm and suicide.The prevalence of depression has remained consistently high.The researches show that depression has become the second-largest contributor to the global disease burden,ranking just below cardiovascular diseases.However,the main challenges in the screening,diagnosis,and treatment of depression are as follows: Firstly,there is a low diagnosis rate.The clinical etiology of depression remains unclear and,unlike other diseases,it lacks objective biochemical or medical imaging indicators to assess depression and its severity.Currently,the screening and diagnosis for depression primarily rely on professional doctors’ interpretation of patient questionnaire scores and symptom self-reports,which is susceptible to subjective factors,and lead to potential misdiagnosis or underdiagnosis.Secondly,there is a low healthcare-seeking rate.Due to insufficient awareness of depression,some patients erroneously equate depression with other mental disorders,leading to feelings of shame and reluctance to disclose their condition to others.These misunderstandings and biases contribute to patients’ unwillingness to seek medical help.Additionally,there is a shortage of well-trained professionals.Lastly,there is a high recurrence rate,patients who experience relief from depression after treatment are highly prone to recurrence.The research reports indicate that the recurrence rate for patients after their first episode is 50%,rising to 75% for those with two episodes,and reaching as high as 90% for patients after three episodes.To address these issues,this study has developed a wearable wristband integrating multiple perceptual sensors such as voice,acitivity behavioral,and physiological sensors to objectively and dynamically monitor and assess changes in depressive symptoms over the long term.This system is applied in collaboration with specialized hospitals to conduct long-term tracking and monitoring experiments on patients with depression,establishing a non-inductive,long-term,localized,and standardized multimodal sensory synchronous depression monitoring dataset.The study has investigated wearable-based digital biomarkers for depression screening and diagnosis.Furthermore,it has constructed mental-sensing-graphs and emotion-sensing-graphs guided by the psychological knowledge and emotion models,which achieve the fusion of multimodal perceptual,and the representation of the development and evolution patterns of depression from both micro and macro perspectives.Based on this,the interpretable graph learning models are established to effectively identify depression and detect changes in depressive symptoms.Finally,an Io T-based wearable depression assessment and management system is developed to longitudinally monitor and evaluate the psychological health status of wearers.The main research contents and innovations of this study are as follows:1.In response to the lack of wearable devices for long-term tracking and monitoring of depression and localized multimodal synchronous depression datasets,the study has investigated the impact relationships between the "bio-psycho-social" model of depression etiology and voice,activity behavioral,and physiological signals.Considering user privacy and limitations of device resource,a wearable wristband integrating voice,activity,and heart rate sensors has been developed.We have conducted online real-time extraction of depression-related voice features and proposed an adaptive filtering algorithm using recursive least squares to estimate motion heart rate by fusing acceleration signals to eliminate motion artifacts.Long-term tracking and monitoring experiments on patients with depression have been conducted to establish a long-term,natural,localized,and standardized multimodal sensory synchronous monitoring dataset for depression,laying the foundation for subsequent research.2.To address the lack of objective biological markers for effectively aiding in the diagnosis of depression,statistical methods such as correlation,variance analysis,and non-parametric tests have been employed to remove redundant,unrelated,and nondifferential features,thereby narrowing down the exploration scope of digital biomarkers and establishing a relevant differential multimodal feature set.The variational autoencoders have been applied to balance minority classes.focal loss has been introduced to emphasize difficult samples,and a method for identifying depression improvement levels based on the XGBoost model is proposed.Mechanisms for evaluating feature contributions,such as Shapley values and linear model coefficients,have been studied to respectively determine common features in the differential feature sets before and after a course of treatment in adolescents from various perspectives,individual interpretation and overall explanations.Consequently,based on the intersection of these three aspects,objective digital biomarkers for assisting in screening and diagnosis of depression are constructed.3.To tackle the issue of multimodal fusion and the interpretability of diagnostic evaluation models for depression,prior knowledge is utilized to map digital biomarkers representing depression development trends to polar coordinate systems,thereby predicting psychological states for time segments.A mental-sensing-graph is constructed based on the relationships between psychological state nodes established from overlapping regions,achieving multimodal fusion and the representation of depression.Furthermore,a graph structure learning module based on improved cosine similarity principles have been proposed to explore potential relationships between psychological state nodes,integrating these relationships into mental-sensing-graphs based on digital biomarker.Graph convolutional models of hierarchical pooling cascades are developed to extract shallow,intermediate,and deep depression global information from mentalsensing-graphs,and mutual information theory and graph interpretation algorithms are utilized to construct a mental-sensing-graph interpretation model,explaining the mechanisms of identifying depressive symptoms based on mental-sensing-graphs and their subgraphs.Finally,the differences between patients with depression and healthy control groups,as well as the effectiveness of mental-sensign-graph learning model for depression recognition are validated using a multimodal synchronous dataset4.To address unclear biomedical significance of graphs constructing from depression data and the evaluation of improvement in the high recurrence rate of depression,this study has conducted an emotional calibration in a healthy control group,combined with digital biomarkers to establish a segment-level emotion recognition method for patients with depression,and constructed emotion-sensing-graphs guided by the Mikels emotion wheel theory to represent the improvement and development patterns of depression and the differences between patients with depression and healthy control groups.Attention mechanisms incorporating adjacency matrices are designed to learn potential relationships between emotional nodes,and a multi-layer graph convolutional model embedded in residual networks has been proposed to focus on potential emotional relationships and obtain residual emotion-sensing-graphs,thus optimizing the original graphs.To eliminate weak connections,thresholds are applied to the residual emotionsensing-graphs to preserve effective emotional connections for depression improvement assessment.On the multimodal synchronous dataset,the proposed interpretable emotionsensing-graphs attention residual graph convolutional model has demonstrated superior performance,with a validation accuracy of 0.86.Finally,an Io T-based wearable depression identification and evaluation management system is developed to validate the research methods and conclusions.In summary,this study provides new solutions and technological pathways to address the three major challenges of low healthcare-seeking rate,low diagnosis rates,and high recurrence rates of depression.It offers objective and important reference indicators to assist in screening and diagnosing depression,provides intuitive tools for the representation of the developmental patterns of depression in graph form from micro to macro perspectives,and facilitates the further promotion and development of wearable technology-assisted screening,diagnosis,and treatment tracking techniques.It also provides a wearable-based full monitoring solution and technology for depression,from early warning,diagnosis and treatment to post-treatment tracking.