Literature Collection

BACKGROUND: Most people with opioid use disorder (OUD) do not receive evidence-based treatment. To increase treatment rates, primary care clinics may choose to implement risk prediction tools available in the electronic health record (EHR) to identify patients with a high risk of OUD or overdose. OBJECTIVE: To externally validate Epic's cognitive computing model to predict the Risk of Opioid Abuse or Overdose (referred to as the Opioid Risk Score; ORS) in three large integrated health systems. DESIGN: Prospective cohort study secondary to an ongoing clinical trial. PARTICIPANTS: Patients (N = 704,764) aged 18-75 who had a primary care encounter during the study period (April 2021-December 2022) and did not have an OUD diagnosis at index. MAIN MEASURES: Data were extracted from the EHR. The index date was defined as the first date within the study period where the patient met eligibility criteria and had an ORS calculated by the EHR. The binary outcome variable was whether the patient was diagnosed with OUD or experienced an opioid overdose within 12 months of the index date. KEY RESULTS: Most patients were classified as low risk on ORS (99.6%). Few patients experienced an OUD diagnosis or overdose in the 12-month follow-up period (0.3%). The model correctly classified 185 of 2362 patients who experienced an event (sensitivity 0.0783, 95% CI 0.0675, 0.0892) and 699,926 of 702,406 patients who did not experience an event (specificity 0.9965, 95% CI 0.9963, 0.9966). Few patients with high ORS experienced the event (PPV 0.0694, 95% CI 0.0598, 0.0791). The model had excellent discrimination (c-statistic = 0.815) but was poorly calibrated, underestimating risk for patients who experienced the outcomes. CONCLUSIONS: Epic's ORS demonstrated excellent discrimination but very low sensitivity across three large integrated health systems. Health systems should exercise caution before implementing vendor risk prediction models without validating their use in their patient populations. CLINICAL TRIAL NUMBER: Not applicable.