Every year, approximately 703,000 people worldwide die by suicide, and the number of suicide attempts is approximately 20 times higher. Suicide is a global public health issue, and in 2019, it was the second leading cause of death among adolescents aged 15 to 19.1 Suicide poses significant harm to families, economies, and societies at large. In the United States, 7% of adolescents reported having made one or more suicide attempts during the past year.2 The latest data indicate that in 2021, China featured a suicide rate of 5.29 per 100,000 people (2022 China Health Statistics Yearbook), and the suicide rate was higher among Chinese adolescents than among other age groups.3 In 2014, the World Health Organization published its first World Suicide Report titled "Preventing Suicide: A Global Imperative" to raise awareness of the public health significance of suicide and suicide attempts and to promote suicide prevention as a priority on the global public health agenda. However, as of 2020, only 35 countries had developed independent suicide prevention strategies or plans,4 thus making the accurate prediction or identification of individuals who are at risk of suicide a significant global challenge.
At present, the primary screening tools for suicide involve questionnaires. Various psychological assessment tools are used to measure suicidal thoughts or behaviors, such as the Beck Scale for Suicidal Ideation (BSI), the Suicidal Behavior Questionnaire (SBQ), and the National Suicide Assessment Scale for Risk (NGASR).5,6 However, individuals at high risk of suicide often exhibit a degree of planning and concealment, thus making it difficult to identify them solely through questionnaires. In clinical practice, the assessment of suicide risk is influenced by the experience and cultural background of healthcare professionals, leading to some degree of subjective bias; it is also challenging to conduct large-scale screenings in various populations due to limitations pertaining to time and location. In this context, scholars have shown interest in identifying easily accessible objective indicators such as voice metrics, head movements, facial features, gait, serum markers, and salivary hormones,7–11 and significant progress has been made in this area. Braithwaite12 validated suicide models based on Twitter data and found that machine learning algorithms could effectively identify individuals at risk of suicide. Pestian et al.13,14 used a combination of vocabulary and vocal characteristics to achieve 85% accuracy in differentiating among individuals with suicidal tendencies (suicide attempts and suicidal thoughts), individuals with mental illness but no suicidal tendencies, and a control group.
Facial expressions are among the most challenging indicators to conceal, and the use of the standardized Facial Action Coding System (FACS) and software such as OpenFace15 for feature extraction offers the potential for further research in this area. Developed by American psychologist Ekman in 1978, FACS deconstructs different facial movements into corresponding muscle group actions, and the use of FACS for facial expression recognition involves the task of identifying action units (AUs). Researchers such as Hu et al.16 have suggested that facial emotions can be used to identify suicide risk and have reported an association between suicide and individual expressions of anger. In 2017, Laksana et al.17 used online video repositories to predict suicide risk in healthy control groups, depressed groups, and groups of individuals with suicidal tendencies. The findings showed that regions related to smiling (AU12) ,frowning, raising the eyebrows, and head movement speed exhibited good discriminatory power. An interview showed that in the mouth region, chin lift (AU17) could be used as an indicator to distinguish suicide attempters and suicide attempters .17 In 2019, scholars used social media video streams to differentiate between individuals at high and low risk of suicide. These scholars noted that individuals at higher risk of suicide exhibited more extended periods of silence, drooping shoulders, and certain limb movements.18 Given the mature classification system and ease of data collection offered by facial action units, they are well suited to the role of serving as objective features for suicide risk identification.
The differential activation theory of suicide4,19 posits that individuals who are at risk of suicide exhibit cognitive functioning that is similar to that of the general population in normal conditions. However, when significant stress events or emotional fluctuations occur, cognitive patterns associated with suicide are activated. In other words, the generation of relevant emotions repeatedly activates cognitive processes related to suicide, leading to stronger suicidal thoughts. When such emotions resurface, they are more likely to trigger suicidal ideation. Mann et al.14 proposed the "stress-diathesis" model, which suggests that suicide involves interactions among stress factors, protective factors, and individual susceptibility.20 Stress factors encompass the processes through which individuals perceive, evaluate, and adapt to environmental stressors, and they are significantly associated with suicidal thoughts.3,21 Individuals who experience suicidal thoughts or behaviors are often undergoing stress due to either acute stressors or long-term pressure, which leads to a perceived lack of effective solutions and learned helplessness and ultimately results in the choice to commit suicide.22 Increased negative life events, especially economic uncertainty, social isolation, substance abuse, and other such events, lead to more stress-related disorders and suicidal tendencies.23 Accordingly, to simulate the impact of life events on suicide more actively and ensure that the research is ecologically valid, this study recorded facial features during both stressful and non-stressful phases within the framework of the Trier Social Stress Test (TSST). The study aimed to investigate the effectiveness of facial action unit features with regard to suicide risk under stressful conditions.
In recent years, big data has developed rapidly, expanding the scope of application of machine learning.24 Ryu et al.25 used machine learning to identify suicide attempters among individuals who exhibited suicidal ideation, achieving an AUC of 0.95 and an accuracy of 0.89. Hill et al.26 conducted a longitudinal study involving home interviews with young people across the United States and assessments of suicide attempts 12 months later, finding that two classification tree solutions maximized risk prediction. Hettige et al.27 demonstrated that four machine learning models performed similarly on the task of identifying suicide attempts among people on the schizophrenia spectrum. Among the six machine learning methods compared by Barros et al.,6 random forests were found to predict suicide risk robustly in patients with psychiatric disorders. Research conducted in the past five years has consistently shown that machine learning enhances our understanding of and ability to predict suicidal behavior, with random forests being particularly effective in the task of predicting suicide risk.6
Currently, research on the use of facial action units as objective indicators to identify and predict suicide remains lacking. Moreover, stressful conditions exhibit greater fidelity to real-life situations, reducing the subjective concealment of patients. Therefore, this study utilized facial action unit feature sets collected using the Trier Social Stress Test (TSST) framework to construct a suicide risk prediction model based on machine learning. By studying the key factors involved in the process of suicide under stress conditions, a more authentic understanding of the development of suicidal behavior can be obtained. Constructing a suicide risk prediction model based on facial action units enhances the feasibility of large-scale suicide risk screening using objective indicators.