Mindfulness and Social Networks for Suicide Prevention

We will use two existing network data sets, collected by the Dr. Eric Rice from two different populations–1. youth experiencing homelessness (YEH) and 2. active duty army personnel–to explore these issues. Both YEH and active duty service members are “special” populations with life experiences that are markedly different than each other and most other civilian populations. From a social network science perspective, these two seemingly different populations share some critical features. In both contexts, there is a primary network consisting of other persons like themselves—other soldiers or other street youth—and there is a simultaneous set of multiplex network connections that surround the individuals in question, such as family, service providers, and friends from outside the primary network.

We will develop machine learning methods that can leverage rich attributed networked data to understand how the personal attributes and the network structures (e.g., help-seeking or mindfulness) relate to the personal outcomes (e.g., suicidal ideation). The hypothesis is that the target outcomes are dependent on the personal attributes as well as the personal attributes of the neighbors, the neighbors’ neighbors, and so on. We will use Graph Neural Networks on our datasets to perform classification of target behaviors (e.g., suicidal ideation). Then, we will use an “explainer” which will identify the subset of the network and attributes that contribute the most to the classification, thus revealing decisive interpretable patterns.

In our previous work (Fulginiti et al., 2021), we effectively used machine learning to predict suicidal ideation and suicide attempts among youth experiencing homelessness. In particular, we employed classification and regression tree (CART) models, which represent a subset of decision-tree (DT) analysis. Ten variables contributed to the construction of the DT for predicting suicidal ideation (in order of importance): (a) sum of traumatic childhood experiences, (b) lifetime hard drug use, (c) depression score, (d) proportion of street friends providing emotional support, (e) proportion of network members engaging in hard drug use, (f) proportion of network members providing tangible support, (g) current age, (h) proportion of home-based friends providing emotional support, (i) proportion of network members objecting to risky behavior, and (j) network size. Six variables contributed to the construction of the DT for predicting suicide attempts (in order of importance): (a) depression score, (b) proportion of network members objecting to risky behavior, (c) age first homeless, (d) proportion of home-based friends in the participant’s network, (e) sum of traumatic childhood experiences, and (f) frequency of fighting. Our work was the first known work to use machine learning, in the form of decision-tree analyses, in combination with personal AND social network data to understand suicidal experiences. This is a key advance given that socioecological factors have been underutilized in suicide prediction endeavors (even with advances in AI) and our work found that more than half of the variables that contributed to the construction of the decision trees were social network variables. Moreover, our decision trees produced higher AUC values than most studies using DT models.

In our prior work (Srivastava et. al., 2019, Petering et. al., 2021), we studied the spread of violence among homeless youth. We proposed the Uncertain Voter Model to model the spread of violence as a “non-progressive” diffusion process. The model takes into account the uncertainty in the knowledge of the network, and based on available network data, models interactions through an existing edge or a non-existent edge that may be created in the future.

A pilot study was conducted in partnership with Safe Place for Youth, a drop-in center for YEH. An algorithm identified 11 youth (of whom 8 participated) to become mindfulness and yoga peer ambassadors. These ambassadors participated in the weekly program named MYPATH (mindfulness and yoga peer ambassadors towards health) over the course of two months. Intervention programming included an intensive 3-hour workshop that discussed ambassadors’ perceptions of the causes and impacts of violence in their lives, as well as didactic and experiential mindfulness and yoga training. The intensive workshop was followed by 6 weekly 1-hour trainer-facilitated mindfulness and yoga classes that were open for attendance to the full network (i.e. not only ambassadors). Mindfulness exercises included attention-control meditation and empathy and compassion meditation exercises. Yoga was performed in street clothes without special props to improve proprioception and acceptance of physical experience. Results (Petering et al., 2021; Barr et al., 2022) showed that 7 weeks after the introduction of the MYPATH programming, there was a 40% reduction in the number of individuals involved in physical fights, and an increase of 85% in the number of individuals who practiced regular mindfulness and yoga. Network-level reductions in violence engagement and dispositional mindfulness were maintained at 1-month after follow up after the cessation of intervention programming. Moreover, the selected peer ambassadors showed qualitative high engagement during the program as predicted by the algorithm.