Journal of Technology in Behavioral Science Health (social science)
Suicide Risk Flags in Electronic Health Record Vignettes: Impact on Clinician Responses
Abstract
Abstract Electronic health record suicide risk flags generated by artificial intelligence are intended to support clinical decision making, but their utility may depend on whether clinicians understand and act on the information provided. This study examined how mental health clinicians responded to simulated algorithmic suicide risk flags and their explanatory details in electronic health record vignettes. Ninety-nine mental health clinicians were assigned nine fictional clinical vignettes across three information conditions: base clinical information (A), base information plus a newly delivered simulated algorithmic suicide risk notification (B), or the algorithmic notification followed by explanatory information (C). Each clinician reviewed three vignettes per condition at each assigned information level. Written responses were qualitatively coded to capture intended clinical management actions, and we examined whether and how clinicians’ intended management responses changed following the disclosure of a simulated suicide risk notification and subsequent explanatory information. Clinicians introduced fewer previously unmentioned action subcodes following explanatory information at C than following the notification only at B ( M s = 0.09 and 0.87, respectively), IRR = 0.11, 95% CI [0.07, 0.17], p < .001. Although 70.8% of C-level responses included at least one action subcode absent from the preceding B response, 93.5% of these additions reintroduced content previously stated at A. The number of distinct action subcodes within individual responses decreased across stages, whereas the cumulative number mentioned across successive responses increased descriptively. Sequential disclosures were associated with changes in the content of intended management responses, including the addition and reintroduction of action subcodes. The study relied on clinicians’ written responses to fictional vignettes rather than observed clinical behavior, and the sample was limited in size and demographic diversity. Sequential disclosure of simulated suicide risk notifications and explanatory information was associated with changes in stated management responses, largely through reintroduction of earlier actions rather than expansion of the action repertoire. Future research should test whether explanations help clinicians select and implement appropriate suicide risk management strategies and whether any such changes extend to patient outcomes.