An Exploration On Auxiliary Identification Of Depressive Tendencies

Depressive symptoms in young people may persist into adulthood and develop into depression.Early identification and treatment of depression are essential in promoting the remission of this disease.However,depression is traditionally diagnosed through questionnaires and interviews,which rely on patients' and clinicians' reports.Therefore,the evaluation can be subjective and inconsistent.Additionally,early signs of depressive tendencies are difficult to detect and quantify.Studies in recent years have shown that identifying depression by behavioral indicators has the potential to be repeatable and generalizable.Automatic detection of depressive symptoms would potentially improve diagnostic accuracy and availability,leading to faster intervention(Haque et al.,2018).Given the current situation,this study is aimed to propose an automatic depressive tendencies prediction system using multiple behavioral features of Chinese university students in different scenarios by applying artificial intelligence technology(feature extraction and classification algorithm).In this study,a set of behavioral features were extracted from our home-made Chinese university students' behavioral dataset,which can be used to establish the mapping relationship between students' depressive tendencies and their observed behaviors.This study first established a new multi-modal dataset among Chinese university students,including participants' behavioral data and their scores of two depression scales(BDI-? and CES-D)under four experimental scenarios.Then the experimental study including 5 sub-experiments was conducted where behavioral characteristics were extracted from the dataset to establish the depressive tendencies prediction model.The performance of different prediction models was analyzed and compared to explore the benchmark for the depressive tendencies identification and provide a reference for the establishment of a systematic model for depressive tendencies identification.Experiment 1 investigated the performance of depressive tendencies prediction model based on gait information in natural walking scenario.Results:Natural gait information in our dataset did not perform well in predicting depressive tendencies.However,the classification accuracy of data-enhanced ST-GCN model was improved compared with that of ST-GCN without data-enhancement.The accuracy increased from 59.0%to 61.5%,and the sensitivity reached 80.0%.Experiment 2 investigated the performance of depressive tendencies prediction model based on dynamic facial features in emotional speeches.Results:Dynamic facial features in emotional speeches can effectively predict depressive tendencies.The highest accuracy is 76.1%.Experiment 3 investigated the performance of depressive tendencies prediction model based on voice information in emotional speeches using different voice presentation and feature extraction methods.Results:Speech recognition model based on medium and low audio features could predict depressive tendencies better than the prediction model based on visual audio features.Experiment 4 aimed to propose the baseline regression parameters of a systematic depressive tendencies prediction model based on the emotional speeches and to compare the performance of the depressive tendencies prediction model using deep-learned features and hand-crafted features.Results:Facial deep-learned features(RMSE=11.59,MAE=8.28)have better performance than facial hand-crafted features(RMSE=12.38,MAE=10.43)in depressive tendencies prediction.Experiment 5 investigated the effectiveness of depressive tendencies prediction model based on facial expression recognition in emotional videos watching scenario.Results:Facial expression recognition techniques used in this experiment could not effectively predict depressive tendencies.In conclusion,this study demonstrated that:(1)Depressive tendencies prediction models based on the facial information have moderate and stable performance across different scenarios and valence,suggesting differences in facial modes between people with and without depressive tendencies.(2)Depressive tendencies prediction models in natural scenarios(natural walking and videos watching)were less effective than that in the experimental scenario(emotional speeches).