Social Media Detection and Monitoring of Suicidal Risk Among Adolescents

Background: Social workers engaged in clinical practice with adolescents at risk for suicide, the 2nd leading cause of death among youth, must consider their unique needs for social media monitoring. Suicidal disclosures and risk communication occur frequently online, more so than they do in-person. Further, suicidal youth are more vulnerable to negative social media experiences, e.g. cyberbullying or exposure to self-harm content, than youth in the general population. Advances in natural language processing (NLP) offer the capacity to detect risk for suicide from language present on social media content to automate social media monitoring. For example, the OurDataHelps platform by QntfyTM has used a NLP algorithm to detect risk of suicide attempt from social media data with a high degree of accuracy (70-85% true positive rate). Such an approach shows promise for augmenting self-report measures of suicidal risk used within clinical care for adolescents. However, acceptability and feasibility within a clinical context remains unknown.

Objectives: First, we will report the findings of a mixed methods study which explored acceptance of automated social media monitoring through the perspectives of suicidal adolescents, parents, and clinicians. Second, we will report the preliminary findings from a pilot study in which we are testing the feasibility of collecting social media content from suicidal adolescents in clinical care.

Methods: Study 1: Adolescent patients (n= 15, ages 13-18), parents (n=12), and clinicians (n=12) from an intensive outpatient program for suicidal youth participated in surveys and either interviews or focus groups. Clinicians, who were predominantly clinical social workers but also included counselors and psychiatric nurses, were engaged in team-based care of acutely suicidal youth. Focus groups and interviews were analyzed using a thematic analysis approach and triangulated with descriptive survey data. Study 2: Data collection is underway (n=7 adolescents recruited) toward a goal of recruiting 35 suicidal adolescents and 15 healthy controls (recruitment estimated to be at last 75% complete by the time of the conference). Recruitment rates will be analyzed to evaluate feasibility and positive rates for suicidal risk from a NLP algorithm will be compared to clinical assessments of suicidal ideation and behavior.

Results: Study 1: Adolescents, parents, and clinicians reported facilitators toward automated monitoring, including a desire for protection in adolescents’ online environments. Barriers included concerns for privacy and limitation of free online expression. Study 2 Of youth and parents who were told about the research study, 72% were interested and eligible to participate, 17% were eligible but reported not using social media, and 11% were not interested. Those who were disinterested primarily discussed privacy concerns. NLP analyses are forthcoming, pending recruitment of a larger sample.

Conclusions: Automated monitoring of risk statements detected through social media content presents a promising avenue to augment self-reported assessments of suicidal risk among youth; however, for this method to be successful careful attention must be paid to privacy concerns and desires for freedom of expression. Incorporating such methods into clinical practice could aid social workers in supporting the safety of their adolescent clients who are vulnerable to suicide.