Leveraging Physical and Online Digital Infrastructures to Infer College Students' Social Isolation and its Relationship to Risk of Suicide
Principal Investigators:
Munmun De Choudhury, PhD | Georgia Institute of Technology, School of Interactive Computing
Dorian Lamis, PhD, ABPP | Injury Prevention Research Center at Emory
Ursula Kelley, PhD, APRN, ANP-BC, PMHNP-BC | Emory University, Nell Hodgson Woodruff School of Nursing
Co-Investigators:
Gregory Abowd, PhD | Georgia Institute of Technology, School of Interactive Computing
Thomas Ploetz, PhD | Georgia Institute of Technology, School of Interactive Computing
Research Gap:
Despite the growing challenges of increased suicide rates on college campuses there exists a lack of real-time information on suicidal risk to guide prevention or intervention efforts. A student's social network can provide informational or emotional assistance which can buffer against negative stress.
Description:
Suicide is a serious mental health concern that affects millions of individuals worldwide. From 1999 through 2017, the age-adjusted suicide rate increased 33% from 10.5 to 14.0 per 100,000. These challenges are pronounced in college campuses, which have been disproportionately affected by the growing incidents of suicide. In spite of the urgency of this public health problem, there exists a lack of real-time information on suicidal risk to guide prevention or intervention efforts. One approach to this public health issue is to improve understanding of college students’ social context and its relationship to mental health – a person’s social network can provide instrumental, informational, or emotional assistance, which are known to boost self-esteem and self-efficacy, at the same time buffer against the negative effects. However current understanding the relationship between offline/online social isolation and risk of suicidal ideation/thoughts, has largely focused on population-level assessments. Researchers further note the limited power of existing risk factor studies due to a reliance on fixed time point sampling of cross-sectional data. Working with the Office of Information Technology and Student Health Services at Georgia Tech and Emory University, for a cohort of 100 consented students seeking psychiatric treatment on campus for mood and/or anxiety disorders and who have been at risk of suicide, this interdisciplinary team comprising machine learning, ubiquitous computing, and clinical researchers at Georgia Tech and Emory University/Grady Health System will leverage anonymized campus WiFi logs and de-identified Facebook data to automatically and passively assess students’ social context, interactions, and importantly isolation in the physical world the online world respectively. By employing state of the art and innovative machine learning and forecasting approaches, the team will examine if depleted social interaction and unexpected changes in social context, thus measured, predicts students’ exacerbated risk in the future. This preliminary one-year study will provide rich data toward seeking future extramural (NIH) funding focusing on developing timely, evidence-based clinical and service interventions and treatment paradigms that support college students affected by poor mental health, particularly suicidal ideation/behaviors. Notably, because of our topical focus and the approach adopted (passive/social media data; machine learning), the work fits squarely with and seeks to advance IPRCE priorities of the Center for Disease Control and Prevention’s National Center for Injury Prevention and Control (NCIPC) around suicide prevention in particular, and self-directed violence in general.
Aims:
- Computational extraction of digital phenotypes of offline and online social interaction from physical and digital infrastructure data volunteered by college student patients at risk of suicide.
- Machine learning modeling and prediction of associations between the digital phenotypes of offline and online social interaction and suicidal risk exacerbation in college student patients.
- (Emory only): Explore the risk and protective factors associated with suicidality among university students, focusing on social isolation, intrapersonal factors and social determinants of health.
- (Emory only): Describe students’ experience of surviving the loss of a close friend/fellow student, family member, or professional colleague to suicide.
Why is this study important?
This preliminary one-year study will provide rich data toward seeking future extramural (NIH) funding focusing on developing timely, evidence-based clinical and service interventions that will support college students affected by poor mental health, particularly suicidal behaviors.
The study team recruited 43 students to share their Facebook data over 2 years (Aug 2019 – Oct 2021) They then developed a participant portal to preserve participant privacy, provide transparency, and support compliance over 3 different data collection stages, approximately 6-8 weeks apart. They processed language on social media to describe temporal trends in psycholinguistic attributes of speech. Study researchers found that temporal trends in psycholinguistic features on social media were good at predicting mental health risk:
- They first found commendable evidence that temporal trends from private social media, such as Facebook, can be used to predict if a student is in a concerning mental state for Anxiety (F1-score = 0.92), Depression (F1-score = 0.85), Social Isolation (F1-score = 0.89), and Suicidal Behavior (F1-score = 0.93). These results were comparable to results from other studies predicting mental health status with Facebook data.
- Additionally, they demonstrated that these same features data can predict COVID--induced disruptions to both mental and physical wellbeing. The researchers built baseline models using scores in self-reported survey instruments (Anxiety, Social Isolation, Depression, Suicidal Behavior, and Resilience) to predict disruptions. In comparison to these baseline models, the Social Media-enabled models showed more accurate results in predicting students who had been disrupted. For mental wellbeing, their model achieved an F1-score of 0.76, a 28% improvement over baseline. For physical wellbeing, their model achieved an F1-score of 0.85, a 43% improvement.
- Finally, the study team further explored the opportunities for data minimization, given their study collected 19 months of data since the GT campus went in a COVID-induced lockdown. They searched through the entire time series and used variable lengths of data to train models and understand when the signals for disruption emerge. The length of data did not appear to be indicative of the prediction. However, accuracy of predicting disruption was relatively lower in the first 7 months.
If you or someone you know is considering suicide, know you’re not alone and help is out there. Contact the National Suicide Prevention Lifeline or 1-800-273-8255
