Algorithmic Solutions to Improve Suicide Prevention Gatekeeper Training Among Youth Experiencing Homelessness: Seeking Equity without Compromising Coverage

Background/ Purpose: Youth experiencing homelessness (YEH) are vulnerable to suicide but most youth in suicidal crises do not seek help from trained professionals. Gatekeeper training, which involves teaching people to recognize and support someone in crisis, is a widely used approach for facilitating mental health service utilization. The goal is to recruit and train a small subset of people in a network (i.e., gatekeepers) with the aim of achieving broad network coverage. Two problems exist with typical recruitment strategies. First, they can lead to inequity in coverage across different racial and ethnic groups. Second, they do not account for the fact that some people selected to be gatekeepers may opt out of such training. The present study sought to test a newly developed artificial intelligence (AI) algorithm that can guide decisions about recruitment to maximize coverage in a YEH network while achieving coverage equity by race/ethnicity and accounting for uncertainty in potential gatekeeper participation.

Methods: A new AI was developed by computer science colleagues to address the issues in intervention design described by the social work team. Informed by work on two-stage robust optimization, we developed an algorithm to maximize the number of covered network members with a set number of recruited gatekeepers, minimizing bias and accounting for probabilistic non-participation. Data were collected from 165 YEH at a drop-in center for YEH in Los Angeles. Participants had to complete a survey about their individual characteristics and a social network interview about their relationships to one another. We then tested our algorithm performance against two competing algorithms that simulated typical gatekeeper recruitment strategies; these included a random approach (i.e., recruiting people without knowledge about their position in a network) and a degree centrality approach (i.e., recruiting people who are connected to the most people in a network). Performance was evaluated by percent of network coverage and bias in coverage.

Results: With only 30 gatekeepers, our algorithm was able to achieve 53% coverage in a network of 165 YEH. The coverage for the random and degree centrality coverage approaches were 34% and 41%, respectively. Our algorithm also showed a maximum bias of 12% between the most covered and least covered racial/ethnic group. The maximum bias for the random and degree centrality coverage were substantially larger with 31% and 37%, respectively.

Conclusion/Implications: Our algorithmic-informed approach to gatekeeper recruitment shows that we can plan for training that covers more people in a network while achieving greater equity across racial/ethnic groups relative to typical recruitment approaches. In other words, our algorithm demonstrates a concrete way to plan for gatekeeper training that is effective and equitable. We do not want to contribute to systemic inequity by using recruitment strategies that differentially benefit some youth more than others based on their racial/ethnic identity. Social work is underrepresented in the field of suicide prevention and our results highlight the importance of bringing a social justice lens to bear. Moreover, AI offers the possibility of evaluating potential intervention policies prior to costly deployment and testing.