Exploring adolescent healthcare provider perspectives on machine learning suicide risk classification using a mixed-methods approach (Preprint)

Abstract

Suicide is a major cause of death for youth aged 10 to 24, necessitating effective risk identification at clinical encounters. While machine learning applied to electronic health records shows promise, further research and application of human-centered design is needed to translate these advances into clinical practice.

This study explores pediatric healthcare providers' perspectives on suicide risk models, aiming to inform their design and implementation within the clinical workflow.

Using a convergent mixed methods design, we conducted a quantitative survey and individual interviews, guided by the Consolidated Framework for Implementation Research. Quantitative data were descriptively analyzed, while qualitative data were analyzed thematically. Integration of mixed methods was achieved through a joint display, and results were interpreted through a narrative review.

Thirty-eight participants completed the survey, and six were interviewed. Regarding the integration of a machine learning (ML) risk model into current workflows, 63% of survey respondents preferred alerts only for imminent risk immediately before the patient encounter. Participants preferred a prominent risk notification on the electronic health record (EHR), accompanied by decision-support tools. Approximately 46% of survey respondents expressed the need for ML approaches. Current screening tools pose challenges for patients with low cognitive function and providers. Two-thirds (66%) of survey respondents suggested that designing the EHR-based tool to improve communication between providers could facilitate implementation. Communication with patients and families requires close collaboration, navigating challenges around confidentiality, stigma, and youth-caregiver relationships. Providers expressed concerns about increased demand for mental health services, implications for patient confidentiality, coercive care, inaccurate diagnosis and response, and medical-legal liability.

There is tentative acceptability and enthusiasm among pediatric healthcare providers for AI risk models to identify suicide risk in adolescents. Successful implementation should incorporate provider perspectives in a user-led fashion to build trust and empower clinicians to respond appropriately to risk flags.