Center for Suicide Research and Prevention

• Aim I: Test the effects of providing ED clinicians with information about risk of patient suicidal behavior on clinician decision-making and the rate of subsequent suicide attempts. We will implement a recently developed risk calculator that provides ED clinicians with real-time estimates of the probability that their patient will attempt suicide in the next month. This will be tested with 4,000 patients presenting to two EDs, randomly assigning clinicians to receive (vs not) the risk score and testing the effect of providing this information on care and outcomes.
  • Hypotheses: (a) There will be fewer suicide attempts in the experimental (risk score) condition and (b) Clinicians provided with the risk score data will be more likely to hospitalize patients estimated by the model to be high-risk and less likely to hospitalize patients estimated to be at low risk than clinicians in the control condition. A significant part of this Aim will now focus on maximizing the generalizability of these results by examining clinician- and institutional-level facilitators/barriers to implementation in the Mass General Brigham healthcare system, as well as in an independent, less well-resourced, and more racially-, ethnically-, and linguistically diverse setting.
• Aim II: Improve detection of high-risk patients. Since the development of the risk scoring system noted above, new developments have emerged that would allow the score to be improved by adding more potentially important predictors of several kinds. In addition, new ensemble machine learning prediction methods have recently emerged that are likely to yield more accurate predictions than those based on the method we used in developing our current model.
  • Hypothesis: Prediction accuracy will improve when we revise the model to include the new features and modeling approaches.
• Aim III: Improve treatment optimization among high-risk patients. We will develop a precision treatment rule (PTR) to help ED clinicians make optimal clinical decisions about whether to treat patients with significantly elevated risk of suicidal behaviors by hospitalization partial hospitalization, or referral to outpatient care.
  • Hypotheses: (a) Significant estimated heterogeneity of treatment effects will be found based on a precision treatment model that uses information available during the ED visit (e.g., EHR, NLP, surveys) to predict subsequent suicide-related behaviors; and (b) The estimated reduction in suicide-related behaviors through treatment assignment based on the PTR will be superior to that based on merely providing information of patient high-risk status to ED clinicians.

• Aim I: Pilot test the implementation process of an enhanced EOI for post-ED discharge prevention of SRB, delivered through an Mass General Brigham-Samaritans partnership. This will include (1) developing operational workflows for using our risk algorithm to identify patients for the EOI, (2) conducting a small open pilot, and (3) assessing individual- and system-level barriers and facilitators, and feasibility, acceptability, and fidelity of implementing the EOI at an academic medical setting through mixed methods to inform intervention refinement. We hypothesize that the EOI will be feasible and acceptable, and delivered with fidelity by Samaritans.
• Aim II: Conduct a Hybrid Type I randomized trial of the EOI vs. care as usual (CAU) after ED discharge to test effectiveness and explore implementation-related factors. After refinement of the EOI, we will conduct a trial comparing outcomes of participants randomized to EOI or CAU. All participants will be drawn from those enrolled in the Signature Project (SIG) who were stratified to the top 50% of predicted suicide risk via our automated risk algorithm. We hypothesize that the EOI will be associated with (a) reductions in SRB and increased outpatient treatment attendance (primary and secondary outcomes) at post-treatment (12 weeks after discharge) and (b) acceptability, feasibility and fidelity (implementation factors).
• Exploratory Aim III: Explore moderators of intervention effects. Potential moderators of intervention effects will include, age, sex, race/ethnicity, and predicted risk stratum. We hypothesize that such factors will be associated with differential effects of the EOI on reducing post-ED SRB.

Suicide deaths among youth aged 10 to 19 years rose 56% between 2007 and 2016, and the proportion of pediatric inpatient or emergency department visits due to SRB has doubled since 2008. This public health emergency appears to be affecting youth of underrepresented backgrounds in particular. Trends in suicide attempts over the past decade have remained higher or increased among Black and Latinx youth compared to white peers. The need to develop scalable, practice-based tools with near-term potential for targeting suicide risk in youth is therefore particularly acute in the case of racial, ethnic, and linguistic (REL) minority youth.

• Aim 1: Use machine learning/natural language processing (NLP) with EHR data to develop algorithms to classify suicide risk for REL minority youth. Mindful of the need to reduce mental health inequity through the development of scalable assessment tools that can be implemented in under-resourced settings with REL minority youth, our first aim is to apply ML and NLP to EHR data, to develop risk algorithms to predict ED/inpatient visits for SRB within 3 months of discharge from index hospitalization. We will evaluate the relative performance of co-trained versus locally trained ML algorithms with EHR data from 2 pediatric psychiatric inpatient sites. To determine which of these algorithms is optimal, we will assess their portability with test sets of REL minority youth at both sites.
• Aim 2: Evaluate the incremental value of SDOH and CAT data in predicting suicide risk in REL minority youth. Our second aim is to generate a suicide risk algorithm for REL minority youth that would be applicable to optimal clinical care settings. That is, we will build on the optimal algorithm identified in Aim 1 by incorporating SDoH data, as well as broad psychopathology collected through CAT as part of standard care at one of the inpatient sites.
• Exploratory Aim: Evaluate the performance of the algorithms developed in Aims 1 and 2 for specific REL minority subsamples, such as Black or Latinx youth.

• Aim 1: Identify factors key to just-in-time intervention refinement in an initial pilot. Qualitative feedback on the intervention methods, content, amount of support, timing, and triggering, as well as overall acceptability and feasibility, will be collected in an initial pilot and used to refine the interventions before a full MRT. Hypothesis: The just-in-time interventions will be both acceptable and feasible.
• Aim 2: Determine the proximal effects of just-in-time interventions aimed to promote use of the safety plan and its components in an MRT. Randomly timed, brief follow-up surveys will assess post- intervention use of safety plan components and suicidal thoughts.
  • Hypothesis 1: When collapsing across all intervention methods and content, the interventions will be associated with increased use of the safety plan and its components, as well as reductions in suicidal thoughts.
  • Hypothesis 2: Intervention effects will vary as a function of the intervention method (phone call, text messaging, automated interactive tool, and automated non- interactive pop-up) and the intervention content (reminder to use the safety plan and its components) at high, medium, and low levels of suicidal thoughts.
• Exploratory Aim 3: Explore internal and external contextual factors as moderators of intervention effects. Potential moderators will include momentary internal (e.g., affect, physiology) and external (e.g., access to social support) factors, assessed with both intensive longitudinal self-report and passive sensing. Hypothesis: Internal and external contextual factors will be associated with differential intervention effects.

• Stephen Bartels, MD, MS
• Maurizio Fava, MD
• Sophia Kim, BA
• Sarah Knoll, BA
• Ashley Seiger, MSc
• Benjamin Silverman, MD
• Christine Vogeli, PhD
• John Weisz, PhD
• David Williams, PhD, MPH

• Samuel Aronson, ALM
• Tianxi Cai, ScD
• David Cheng, PhD
• Andrea Foulkes, ScD

• Matthew Nock, PhD
• Rajendra Aldis, MD
• Kate Bentley, PhD
• Suzanne Bird, MD
• Nick Carson, PhD
• Alexis Christie, BA
• Abigail Donovan, MD
• Rebecca Fortgang, PhD
• Ron Kessler, PhD
• Narise Ramlal, BA
• Mark Schechter, MD
• William Schleyer, MD

• Jordan W. Smoller, MD, ScD
• Kate Bentley, PhD
• Rebecca Fortgang, PhD
• Sarah Young, PsyD

• Nicholas Carson, MD
• Richard Liu, PhD
• Philip Wang, MD, DrPH

• Kate Bentley, PhD
• Walter Dempsey, PhD
• Evan Kleiman, PhD
• Alexander Millner, PhD
• Matthew Nock, PhD