Predicting Suicide Risk Using Facial Action Units : A Machine Learning Approach to Objective and Large-Scale Screening

Abstract

Background: Suicide is a global concern, and adolescents currently have the highest suicide risk because they are more prone to extreme behaviors such as suicide and self-injury due to their emotions and coping styles. Research suggests that suicide can be predicted, and therefore, early detection of suicide risk is crucial for suicide prevention and treatment. Methods: This study collected data from 98 participants using the Trier Social Stress Test (TSST) paradigm. Videos were recorded during the normal and stressful phases Results: In the normal state, the most efficient features for identification were action unit (AU) 17, AU20, and AU23; in the stress state, the features AU2 and AU5 were added. In the combined model integrating both subjective and objective measures, the accuracy increased to 83.1%, which was the best performance in this study. Conclusions: This study enriches the empirical evidence that AU features can be used to identify suicide risk, provides a way to formulate and implement policies related to suicide prevention.