Background. Medication for opioid use disorder (MOUD) prevents overdoses and improves mortality in Veterans with OUD, but retention on MOUD is critical for achieving those clinical endpoints. Only 50% of Veterans are retained on MOUD at 6-months post-MOUD initiation. Poor engagement in additional needed care services is an important risk factor for early MOUD discontinuation. Consequently, providers’ ability to identify Veterans in need of additional care or support while on MOUD may increase the likelihood of Veterans’ continued use of MOUD. Valid predictive models can provide an accurate probability of an individual Veteran experiencing the outcome being modeled (e.g., MOUD discontinuation). Prediction of future MOUD discontinuation risk could provide an innovative and real-time method for identifying Veterans in need of additional care (e.g., peer support). Significance/Impact. Predictive models could be used to lower MOUD attrition risk and improve outcomes for this Veteran population by continuously monitoring their risk of MOUD discontinuation in real-time during active MOUD treatment and identifying those Veterans in need of additional care (e.g., if increasing risk between visits, providers might add peer support services to a treatment plan). Innovation. This CDA-2 encompasses three HSR&D research priority areas (opioid/pain, health care informatics, and access to care) while crosscutting HSR methods of “big” data and implementation science, all in an effort to improve care and outcomes for Veterans with OUD. This study will also be the first to develop and pilot test a clinical decision support tool (CDST), based on a predictive model, to improve Veterans’ MOUD retention. Specific Aims. (1) To develop and validate PREMMOUD, a PREdictive Model for MOUD discontinuation. Hypotheses: (H1) I will develop a predictive model with good discrimination (e.g., c-statistic, a measure of goodness-of-fit, ≥0.8) for identifying Veterans likely to discontinue MOUD within the initial 6 months of treatment; (H2) the model generated using neural network techniques will have better discrimination than the models generated using random forest and logistic regression techniques. (2) To adapt PREMMOUD into a CDST to continuously monitor risk of MOUD discontinuation and provide clinical guidelines for addressing the primary risk factors driving the PREMMOUD score. (3) To assess (a) the feasibility of conducting a large scale, randomized controlled trial (RCT) to test PREMMOUD CDST’s (P-CDST) effectiveness as well as (b) P- CDST’s acceptability among waivered providers. Hypotheses: (H3) The feasibility of conducting a large-scale RCT to evaluate P-CDST’s effectiveness will be supported; (H4) P-CDST will be acceptable among VHA waivered providers. Methodology. Using machine-learning methods and data from the VHA Corporate Data Warehouse (2006-2019), I will train and validate PREMMOUD in a national sample of Veterans initiating MOUD (Aim 1). For Aim 2, I will conduct two rounds of focus groups with key stakeholders (VHA providers, Veterans receiving MOUD, VHA operations partners) to inform the creation of a beta-version of P-CDST to be integrated into CPRS/Cerner. To build P-CDST, I will use VHA CDW data, PREMMOUD, SQL Server Reporting Services (SSRS) and the Business Intelligence Service Line (BISL) platform. P-CDST will contain the patient’s real-time PREMMOUD score as well as clinical guidelines to support the provider in addressing the Veteran’s specific risk factors driving the PREMMOUD score. For Aim 3, I will conduct a single-arm, two- site pilot trial to assess study feasibility (provider enrollment, frequency of P-CDST use, and follow-up rates) and P-CDST’s acceptability (clinical usability of P-CDST). Implementation/Next Steps. Aim 1 will support an HSR&D IIR submission in Year 3 to assess whether PREMMOUD can be used to identify which Veterans, receiving MOUD, can effectively be treated in specialty care versus non-specialty care and which Veterans benefit from additional supportive services. A second IIR proposal will be submitted post CDA-2 to conduct an RCT, using a hybrid design, to evaluate the effectiveness and implementation potential of P-CDST in VHA.

This VHA Career Development Award will ensure Dr. Hayes’s expertise in translating knowledge gained from electronic health record (EHR) data and predictive analytics (e.g., predicting the future probability of an outcome for individuals) into interventions that increase Veterans’ retention in medication treatment for opioid use disorder (MOUD; e.g., buprenorphine). His specific training areas span the spectrum of learning advanced methods for working with EHR data, how to develop computer-based tools, and how to test the effectiveness of new interventions. Dr. Hayes will apply this new knowledge to (a) develop a mathematical model that will continuously assess Veterans’ risk of discontinuing MOUD, (b) create a computer-delivered clinical tool that provides clinicians with information (i.e., patient-specific likelihood risk, patient-specific risk factors for discontinuing MOUD, and treatment guidelines, when indicated) on Veterans’ risk of discontinuing MOUD, and (c) test the feasibility/acceptability of the clinical tool among VHA providers who are prescribing buprenorphine.