Background: Machine learning models using electronic health record data to identify individuals at risk for attempting or completing suicide are being developed and readied for implementation. A relatively new innovation in suicide prevention efforts, the implementation of these models warrants study; guidelines for acceptable and safe implementation do not yet exist.
Methods: Guided by the Consolidated Framework for Implementation Research, we interviewed 40 clinicians and health care administrators at 3 integrated health systems. Two sites are currently engaged in implementing suicide risk identification models: one site implemented their approach as part of a mental health visit, the other site integrated it into an existing telephone-based case management program. Clinicians and administrators at the third site, planning for future suicide risk identification model implementation, also were interviewed about their concerns and hopes for this technology.
Results: Content analysis of interviews revealed clinicians across sites broadly support the use of suicide risk identification models as another important tool in the suicide prevention toolkit. Clinician recommendations included early engagement of clinicians in decision-making and planning, establishing a clear clinic workflow, and providing appropriate, thorough, and ongoing training. Studying two different implementation approaches elicited different strengths and challenges. For example, the strength of an established patient-provider relationship in a visit-based approach could be compared to the opportunity to reach patients with an outreach approach and the accompanying barrier of not having a history with the patients. Other barriers included competing demands and time and lack of comfort/training in suicide risk assessment. Clinicians across sites called for consolidating all suicide-related risk information into a standardized and easy-to-access module in the electronic health record to enhance efficiency for risk identification follow-up. Finally, clinicians expressed concern that some patients would still be missed, and they were not sure what to do with patients with chronic suicide ideation who are continuously identified.
Conclusion: Lessons from two sites provided several areas of opportunity for improving implementation of machine learning models that identify suicide risk.