Predicting and Preventing Suicide at Entry to Mental Health Care: A Community-Engaged, Machine Learning Model Implementation

Suicide rates in the United States have increased steadily over the past twenty years, a trend coinciding with rising use of mental health services across the country. To help patients before a suicide attempt, health systems must be able to screen for suicide risk and take action at a large scale. Recently, powerful machine learning (ML) models have emerged that can accurately predict suicide attempts by using historical electronic health record (EHR) data, and yet there exists no standardized framework for implementing these models in care delivery. Here we present a case study describing the deployment of a suicide risk prediction model within a large virtual mental healthcare program at Kaiser Permanente Northern California that handles more than 5,000 intake visits per month. Our approach used data science to evaluate model validity for our novel use case (intake visits). We integrated patient and clinician voices to design a model-augmented suicide assessment workflow, which we tested iteratively with continuous input from clinician managers. Understanding the opportunities and pitfalls of model-augmented suicide risk assessment from the perspective of patients and clinicians provided an intuitive framework for mapping clinical actions to potential prediction scenarios. This playbook can be applied to ongoing co-development of ML uses, as part of health systems’ continuous learning initiatives integrating ML to serve today’s public health needs.