Suicide is among the leading causes of death for adolescents,1 making the need for effective and scalable suicide risk identification methods urgent. Most individuals visit primary care (PC) in the year prior to death by suicide2 and most pediatricians have encountered patients who have experienced suicidality,3 making this an important setting for the implementation of novel suicide risk prediction procedures. Predictive algorithms that utilize machine learning have shown promise in identifying suicide risk.4, 5 Research has primarily focused on the accuracy of such prediction methods across different time horizons (e.g., a week, 90 days, lifetime),5, 6, 7 with only a few examples of real-world implementation.8, 9, 10 As novel risk identification methods gain traction in research, it is important to better understand how suicide risk predictive algorithms can be optimized for healthcare settings and end users to translate such prediction approaches into effective clinical tools.11 Incorporating the perspectives of end users is likely to make research more relevant to actual health decisions that end users will encounter, which can lead to improved uptake of research findings and increased likelihood that patients will achieve desired health outcomes.12

The limited research that exists on end users’ perceptions of suicide risk algorithms highlights important factors that may promote or impede implementation of these algorithms in clinical practice, including privacy and liability concerns, the utility of risk flags for prioritizing follow-up, the need for timely risk notifications and clear protocols for responding to suicide risk, the importance of making information available about why risk was flagged to build provider trust, alert fatigue, and increased demand on healthcare systems.13, 14, 15, 16 The current study represents the first known examination of perceptions of an adolescent-focused suicide risk predictive algorithm intended for use in pediatric PC. Of note, most research on suicide risk algorithms has focused on patient or provider perspectives on algorithms involving data from the EHR. The current study adds to the literature by investigating provider, parent, and adolescent perspectives on incorporating a novel adolescent suicide risk predictive algorithm into PC that is intended to utilize EHR data, as well as to examine whether smartphone data can be leveraged to improve risk prediction. In line with calls in the literature to incorporate implementation science principles earlier in the translational research spectrum to narrow the research-to-practice gap,17 this study aims to ensure key end user feedback is incorporated into the design and deployment of the algorithm. We relied on implementation science models and frameworks,18, 19, 20 including constructs from the Consolidated Framework for Implementation research (CFIR)18, 19 to interpret the current findings. CFIR outlines multiple levels of contextual factors that are likely to impact implementation, including aspects of the setting (in this case, PC), the context surrounding the implementation setting, individual characteristics such as beliefs and knowledge, and characteristics of the innovation such as its complexity (see Table for definitions relevant to the current study). Our approach is expected to yield knowledge necessary to enhance the likelihood of downstream adoption, implementation, and sustainment.