Exploration Of Facial Feature Model In Suicide Risk Recognition Based On Stress State

BackgroundWith the development of affective computing and computer field,facial feature recognition,expression recognition,image big data analysis become possible.With the advent of the post-epidemic era,life events have more and more influence on people,and the prevalence rate of depression has increased compared with that before the epidemic.However,many people make mistakes about depression,and it is difficult to carry out large-scale mental health examination in some areas due to development problems.All these will lead to the delay of the optimal time for treatment,and even increase the risk of suicide.Secondly,suicide is a hot issue at domestic and international.High-risk individuals often have a certain degree of planning and concealment.But suicide is not untraceable,and the problem is manageable.Early detection and prevention of depression and suicide risk can improve the nation’s mental health and even save lives.However,the shortcomings of existing screening methods are obvious,and it has become a research trend to find potential objective feature identification indicators.Among them,facial features are easy to collect,low cost and non-invasive,which can be used as potential objective characteristics for depression attack identification and suicide risk identification.Thus,in order to better simulate the impact of life events on depression and suicide and make the study more ecological valid,this study recorded the facial features of the subjects in the three stages before,during and after Stress under the framework of Trier Social Stress Test(TSST),so as to explore the effects of stress.Effectiveness of facial action unit(AU)features in identifying depressive episodes and suicide risk.MethodsIn the evoked paradigm framework of TSST,the study record videos of subjects in three stages in TSST,and OpenFace software and non-parametric test were used to construct action unit(AU)based depression and suicide risk recognition feature subsets as input indicators.Whether depression and high or low risk of suicide were used as outcome indicators.Finally,SVM,RF and KNN machine learning methods were used to establish the identification model of depressive episode and suicide risk respectively.Results(1)Some AU features can be used as effective features to distinguish the risk of depression and suicide.In identifying depressive episodes,the number of differential AU increased with the progression of stress,including 55 dimension before TSST,76 dimension during TSST and 114 dimension after TSST.In the identification of suicide risk,there were 43 dimension before TSST,90 dimension during TSST and 48 dimension after TSST.The efficient characteristics were not completely consistent in different periods.(2)The consistent indicators for identifying depressive episode and suicide risk included AU4,AU10,AU 12,AU 15,AU23.There were specific indicators for identifying suicide risk and depressive episode before TSST and during TSST,and only specific indicators for identifying depressive episode after TSST.(3)The optimal depressive episode identification model was constructed by "76 AU during TSST+RF",the constructed depression recognition model has good generalization performance(ACC=77.7%,AUC=0.862,and F1=0.817).(4)In a single period,the overall performance of the model for identifying suicide risk was before TSST>during TSST>after TSST.The optimal suicide risk identification model was constructed by "top 10 AU during TSST+RF"(ACC=71.7%,AUC=0.780,F1=0.651).(5)The multi-period model is superior to the single-period model,and the optimal model is constructed by the feature subset of "before TSST+during TSST"combined with RF(ACC=72.1%,AUC=0.789,F1=0.646).(6)The combined subjective and objective modeling has the best performance among all models,and the optimal model is constructed by the AU feature subsets of"before TSST+during TSST+4 subjective scales"+RF(ACC=83.1%,AUC=0.884,F1=0.794).Conclusion(1)Some AU can be used to identify depression episode and suicide risk.Stress can affect the feature screening of the model.Simplifying the input features can improve the model performance.(2)The AU features used for depressive episode identification and suicide risk identification had concordant and specific indicators.(3)The optimal depressive episode identification model was constructed by "76 AU during TSST+RF",and the accuracy rate was 77.7%.(4)The combined subjective and objective modeling can improve the identification accuracy and performance of the suicide identification model.The optimal suicide risk identification model was constructed by "before TSST+during TSST+all scales+RF",with an accuracy of 83.1%.