160 Implementation of personalized statistical prediction models: Uptake as a function of potential risk/benefit

Objectives/Goals: Multivariable statistical models are commonly used as clinical decision aids. However, more research examining real-world implementation of models is needed. The goal of this mixed methods study is to explore to what degree clinician uptake of risk models is influenced by perceptions of the potential risks and benefits of taking action on model output. Methods/Study Population: Mixed qualitative and quantitative data for the ’low-risk/low-reward’ scenario arise from an implementation trial of a electronic medical records-based (EMR) model to alert primary care providers of potential cognitive problems in patients ages 65 and older. We evaluated adoption of the EMR model based on provider actions (cognitive screening/referrals/medications/diagnoses) following Epic best practice alert (BPA), and acceptability of the models through semi-structured qualitative interviews. Data for the ’high-risk/high-reward’ scenario will be gathered from clinical notes documenting the use of statistical models to support elective surgical decision-making in scenarios where such models have been regularly used at our institution. Mixed methods analyses will be holistic. Results/Anticipated Results: Early quantitative analyses following implementation indicated that primary care providers largely did not adopt the BPA, but did take action (referrals, diagnoses, medications) when automated cognitive screening was performed in preparation for the appointment. Qualitative analyses of interview data to understand provider perspectives are ongoing. In the high-risk/high-reward scenario, we expect that providers will more frequently utilize statistical models for decision-making and report positive attitudes regarding their use, when compared to the low-risk/low-reward scenario. Discussion/Significance of Impact: Though used in practice with increasing frequency, the implementation of personalized statistical prediction models has not been widely studied. Understanding in what scenarios providers feel risk models are useful and actionable can guide decision-making for the development and implementation of future models.